Precise estimation of rice leaf macro and micro nutrients from multi-spectral images using neural architecture search with polynomial approximation functions Authors: Diksha Arora, Jhilik Bhattacharya, Chinmaya Panigrahy Published: 2026 ABSTRACT Estimating the nutritional status of rice leaves is crucial for efficient nutrient management and yield enhancement. Traditional wet lab analyses are time-consuming and labor-intensive. This study presents a novel deep learning-based approach utilizing multispectral images captured by unmanned aerial vehicles (UAVs) to estimate macro/micro nutrients in rice leaves. The proposed framework integrates a differentiable neural search technique using polynomial function approximators and an adaptive activation mechanism, which not only provides improved predictive performance but also deals efficiently with limited training data. The model performance is evaluated across different treatments and crop growth stages using mean absolute error (MAE) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values. Experiments were conducted at the Punjab Agricultural University. The results demonstrate that the proposed model achieves MAE values in the range of 0.06–0.11 for SAS-I and 0.06–0.16 for SAS-II across eleven leaf macro/micro nutrients. To further evaluate the reliability of the predicted nutrients beyond the prediction error analysis, uncertainty estimation of nutrients is also performed. Comparative analysis shows that the proposed framework outperforms conventional deep learning baselines and machine learning methods in terms of accuracy and robustness. Furthermore, the t-SNE visualization of learned feature representations effectively clusters similar nutrient values while separating dissimilar ones. The robustness of the proposed framework is further validated through ablation studies, treatment-wise and plot-wise cross-validation, highlighting the contribution of individual components and their performance under varying field conditions. These findings highlight the proposed NAS-based framework for precise and reliable nutrient assessment in precision agriculture. Introduction Rice serves as the staple food crop for the majority of the population and is the primary crop grown in the trans-Indo-Gangetic Plains, including Punjab1. Smart farming with artificial intelligence (AI) serves as an efficient mode for enabling data-driven crop monitoring and decision-making by integrating data from sensors, remote sensing platforms, and predictive models to continuously monitor crop conditions and optimize management decisions2. Among various crop management practices, nutrient management is one of the important factors, as macro/micro nutrients are the essential elements that facilitate the optimal growth and development of crops3. There are sixteen fundamental nutrients for rice crops where nitrogen (N), potassium (K), and phosphorus (P) are primary macro-nutrients; sulfur (S), calcium (Ca), and magnesium (Mg) are secondary macro nutrients; copper (Cu), manganese (Mn), iron (Fe), zinc (Zn), chlorine (Cl), and molybdenum (Mo) are micro nutrients4. Deficiency or toxicity of any of these nutrients can restrict typical plant growth5,6. For example, boron deficiency reduces rice yield and degrades grain quality7. Consequently, diagnosis of crop nutritional status enables early detection and prevention of diseases, ensuring the precise application of fertilizers or nutritional supplements8–12. Fertilization plays a crucial role in plant growth and yield by supplying essential mineral nutrients. However, nutrient absorption varies significantly across spatial and temporal scales. Despite these variations, farmers often apply fertilizers uniformly based on intuitive and visual assessments, leading to their overuse13. Excessive fertilization contributes to environmental degradation through increased greenhouse gas emissions, water and soil pollution, ecosystem disruptions, and high input costs14. Therefore, frequent analysis of leaf nutrient concentration is crucial for accurately assessing crop nutritional status and addressing potential deficiencies. Traditional agronomic methods rely on regular leaf sampling for chemical analysis, which is a time-consuming, labor-intensive, and environmentally harmful process15,16. These limitations have motivated the development of non-destructive, UAV–based nutrient estimation techniques using AI and multispectral remote sensing data. Unmanned aerial vehicles (UAVs) or drones represent a promising approach for monitoring and assessing vegetation status in crop management17. The use of drones mounted with high-tech cameras and sensors (such as RGB, thermal, multispectral, and hyperspectral) offers adequate control over several features, such as flight altitude, geographical coverage, etc., which lead to high spatial resolution data, resulting in precise crop monitoring and crop health analysis18–20. Vegetation indices have been widely explored for estimating plant nutrient concentration21,22. These were derived from spectral bands and provide interpretable features by capturing relationships between reflectance values across different wavelengths. These reflectance values are associated with plant biochemical and structural properties. They capture diverse crop characteristics, including chlorophyll content, canopy structure, and stress conditions, which are indirectly related to nutrient concentrations23,24. NDVI (Normalized Difference Vegetation Index) is the most commonly used VI, effective for assessing vegetation health and crop growth. Similarly, RVI (Ratio Vegetation Index), GNDVI (Green Normalized Difference Vegetation Index), and NDRE (Normalized Difference Red Edge) have shown strong correlations with leaf nitrogen status25. Previous studies have also demonstrated the effectiveness of UAV-based multispectral imagery combined with vegetation indices and feature-selection techniques for estimating both macro/micro nutrient concentrations in crop leaves26,27. Furthermore, integrating multiple vegetation indices, including NDVI, NDRE, GNDVI, and RVI, can provide a more comprehensive assessment of crop nutrient status by capturing complementary spectral information, thereby improving the robustness and reliability of nutrient estimation. In this study, we design a neural search space using polynomial approximation functions to estimate nutrients using UAV-based multispectral images of crops at various growth stages. It is theoretically evident that using deep spectral features from optimal neural architectures will provide better estimates of crop nutrients compared to traditional machine learning techniques. Earlier reliance on traditional machine learning techniques like linear regression, random forests, and support vector machine for canopy nitrogen weight28; gradient boosting regression tree for estimating N, P, K, Mg, Ca and S29; partial least squares regression, artificial neural network (ANN), support vector regression and Gaussian process regression for N, P, and K leaf content13; multiple regression analysis to estimate N, P, K, Ca, Mg, Cu, Fe, Mn, and Zn26 was replaced with deep learning–based approaches employing convolutional neural networks (CNNs). However, selecting an optimal architecture that balances the feature dimensionality as well as the data availability is important. A study27particularly reported cases where multilayer perceptron (MLP) models performed better compared to CNN architectures in citrus leaves nutrient estimation tasks. This motivates the fact that a careful search and design of network architecture is crucial for convergence. It was further observed that many existing deep learning–based studies primarily focus on nutrient deficiency classification of N, P, and K using RGB leaf images rather than quantitative estimation. For example30, proposed an S-RiceNet-D-LightGBM model for early diagnosis of N, P, and K deficiencies, where it outperformed several pretrained CNN architectures like VGG16, ResNet101, DenseNet169, InceptionNetV3, and MobileNetV2, achieving an accuracy of 95.01%. Similarly31, addressed the identification of Ca, Fe, K, Mg, and N deficiencies using CNN. In another study32, evaluated CNN, VGG-16, and Inception-V3 architectures for classifying N, P, and K deficiencies with highest accuracy of approximately 93%, while33 proposed a customised CNN with transfer learning for multi-nutrient deficiency detection, primarily focusing on N, P, and K in paddy leaves, achieving high classification performance, with an accuracy of 98.2%. In contrast, only limited work has explored regression-based non-destructive nutrient estimation of multiple macro/micro nutrients using multispectral imagery34. The proposed Neural Architecture Search (NAS) framework incorporates learnable polynomial approximators with adaptive activations to transform multispectral inputs into a nonlinear feature space, facilitating faster convergence and improved adaptability across varying data distributions. The hypothesis driving this study is that multispectral images possess unique spectral features that correspond to a particular nutrient amount of the crop. These features, when appropriately learnt via sequential filter transforms, can provide some mapping values for nutrient assessment. However, it is emphasized that nutrient estimation models are crop and region-specific due to variations in soil quality, climate, and cultivation practices, and it is often difficult to obtain a large dataset for local regions to train the estimation model. We hence introduce learnable polynomial approximators with adaptive activations in the network architecture to improve predictive performance with smaller dataset sizes. This work emphasizes the importance of designing and validating models on a crop-wise and region-wise basis. Although the proposed model is validated only for paddy crops in the northern region of India, the proposed search-space design may be explored for adaptation to other crops and geographical regions in future studies. The main contributions of this study are summarized as follows:A novel dataset comprising UAV-acquired multispectral images and corresponding laboratory-measured leaf macro/micro nutrient concentrations is developed for nutrient estimation.An adaptive neural model is proposed that integrates learnable polynomial approximation functions within an optimized Neural Architecture Search (NAS) framework for precise nutrient concentration estimation.An optimized and reusable neural architecture search space is designed to enhance model adaptability across varying field conditions, growth stages, and data distributions within the target crop, with potential extension to other crops.Unlike most existing studies that focus primarily on nitrogen, phosphorus, and potassium, the proposed approach enables simultaneous estimation of eleven macro/micro nutrients, providing a more comprehensive assessment of crop nutritional status to support precision nutrient management. A novel dataset comprising UAV-acquired multispectral images and corresponding laboratory-measured leaf macro/micro nutrient concentrations is developed for nutrient estimation. A novel dataset comprising UAV-acquired multispectral images and corresponding laboratory-measured leaf macro/micro nutrient concentrations is developed for nutrient estimation. An adaptive neural model is proposed that integrates learnable polynomial approximation functions within an optimized Neural Architecture Search (NAS) framework for precise nutrient concentration estimation. An adaptive neural model is proposed that integrates learnable polynomial approximation functions within an optimized Neural Architecture Search (NAS) framework for precise nutrient concentration estimation. An optimized and reusable neural architecture search space is designed to enhance model adaptability across varying field conditions, growth stages, and data distributions within the target crop, with potential extension to other crops. An optimized and reusable neural architecture search space is designed to enhance model adaptability across varying field conditions, growth stages, and data distributions within the target crop, with potential extension to other crops. Unlike most existing studies that focus primarily on nitrogen, phosphorus, and potassium, the proposed approach enables simultaneous estimation of eleven macro/micro nutrients, providing a more comprehensive assessment of crop nutritional status to support precision nutrient management. Unlike most existing studies that focus primarily on nitrogen, phosphorus, and potassium, the proposed approach enables simultaneous estimation of eleven macro/micro nutrients, providing a more comprehensive assessment of crop nutritional status to support precision nutrient management. The paper is organized as follows. Section "Materials And Methods" provides an overview of the study areas along with the corresponding treatment details and methodology, including data collection, data preparation, and a detailed description of the proposed nutrient estimation model architecture. Section “Results” presents the results of the various experimental analyses. Section “Discussion” discusses these results in depth and elaborates on a series of experiments conducted to gain deeper insights into the model’s learning behavior. Finally, Section “Conclusion” concludes the paper. Materials and methods Study area description The research is conducted in paddy crop fields of Punjab Agricultural University (PAU), located in Ludhiana, Punjab, India. Data is collected over a span of two years, covering two consecutive seasons of paddy, precisely during the years 2022 and 2023. The subsequent details provide a description of two distinct study area seasons. Study area Season I (SAS-I) The study period extended from July 2022 to October 2022. Direct-seeded rice (DSR) of type PR126 was planted in the field on 9th June 2022. The study area extends between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$30^\circ 54' 38.16''\ \textrm{N},\ 75^\circ 48' 41.57''\ \textrm{E}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$30^\circ 54' 39.53''\ \textrm{N},\ 75^\circ 48' 44.01''\ \textrm{E}$$\end{document}, as illustrated in Figure 1(a). The experimental field covered a total area of 1917 sq. m. It consists of ten plots where P1 represents the control field that did not receive treatment. The remaining fields underwent three distinct treatments, in which P2, P5, and P8 received the same treatment, P3, P6, and P9 received another treatment, and P4, P7, and P10 received the third treatment. The implementation of different treatments aimed to examine the variations in model performance while predicting leaf nutrients across different treatments. The replication of treatments is carried out to augment the quantity of multispectral image data for that particular treatment. This approach has been devised for a precise estimation of leaf macro/micro nutrients. The experimental SAS-I (2022) with its field-wise measurements in meters is shown in Fig. 1(a).Fig. 1Experimental fields of PAU with its Field-wise measurements. Experimental fields of PAU with its Field-wise measurements. Study area Season II (SAS-II) The duration of the study spanned from July 2023 to September 2023. The field was planted with DSR of variety PR126. The study area extends between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$30^\circ 54' 32.89''\ \textrm{N},\ 75^\circ 49' 6.49''\ \textrm{E}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$30^\circ 54' 34.69''\ \textrm{N},\ 75^\circ 49' 4.63''\ \textrm{E}$$\end{document}, as illustrated in Fig. 1(b). The experimental field covered a total area of 2700 sq. m. The field is then further divided into four plots (P1, P2, P3, P4), where the first plot is the control with no treatment, the second has treatment in 4 splits, the third has treatment in 3 splits, and the fourth has treatment in 2 splits. Each plot is then replicated thrice (R1, R2, and R3). Hence, a total of 12 subplots are in the field. In Fig. 1(b) represents the satellite view of the field with its field-specific measurements of the SAS-II (2023) in meters. Treatment details Initially, 15 kg of urea and 20 kg of DAP (Di-ammonium Phosphate) were applied to DSR fields. Several measures were taken to ensure consistency across fields. To maintain soil type consistency, pH tests are conducted in all fields before sowing the crops. Irrigation is standardized by regulating the amount and frequency of water provided to the crops. The same variety of rice (PR126) is sown in all fields, maintaining uniform density. Leaves are always sampled on the same day from all the fields (i.e., across different treatments) for each growth stage, hence ensuring uniformity in data collection. Following this, three distinct treatments were implemented (2 splits, 3 splits, and 4 splits), with a control field that received no further treatment. These treatments were applied identically in both SAS-I (Fig. 1(a)) and SAS-II (Fig. 1(b)). The split treatments were as follows:2 Split Treatment: 10.968 kg of urea was applied twice in P2, P5, P8 in SAS-I, and all R1, R2, R3 replications in P1 in SAS-II.3 Split Treatment: 7.312 kg of urea was applied thrice in P3, P6, P9 in SAS-I and all R1, R2, R3 replications in P2 in SAS-II.4 Split Treatment: 5.484 kg of urea was applied four times in fields P4, P7, P10 in SAS-I and all R1, R2, R3 replications in P3 in SAS-II. 2 Split Treatment: 10.968 kg of urea was applied twice in P2, P5, P8 in SAS-I, and all R1, R2, R3 replications in P1 in SAS-II. 2 Split Treatment: 10.968 kg of urea was applied twice in P2, P5, P8 in SAS-I, and all R1, R2, R3 replications in P1 in SAS-II. 3 Split Treatment: 7.312 kg of urea was applied thrice in P3, P6, P9 in SAS-I and all R1, R2, R3 replications in P2 in SAS-II. 3 Split Treatment: 7.312 kg of urea was applied thrice in P3, P6, P9 in SAS-I and all R1, R2, R3 replications in P2 in SAS-II. 4 Split Treatment: 5.484 kg of urea was applied four times in fields P4, P7, P10 in SAS-I and all R1, R2, R3 replications in P3 in SAS-II. 4 Split Treatment: 5.484 kg of urea was applied four times in fields P4, P7, P10 in SAS-I and all R1, R2, R3 replications in P3 in SAS-II. Methodology The methodology of the proposed study is shown in Fig. 2. The initial step involves data acquisition “Data Collection”, which encompasses the collection of multispectral images and leaf samples for ground truth data. Subsequently, the data preparation “Data Preparation” phase includes the orthomosaicing of multispectral images and leaf analysis reports. The prepared data is further used to develop a crop nutrient estimation model "Nutrient Estimation Model", which takes multispectral image patches as input and predicts the different leaf macro/micro nutrients.Fig. 2Methodology of the study for the estimation of leaf macro/micro nutrients. Methodology of the study for the estimation of leaf macro/micro nutrients. Data collection In this research, an Agribot mounted with a Micasense RedEdge-MX camera is employed to capture multispectral images of the experimental field at PAU. It is equipped with five narrow-band multispectral sensors with a default spatial resolution of 8 cm/pixel at an altitude of 400 ft (120 m). The choice of system is suitable in this case, as the geographic region allows the Agribot to fly very close to the rice fields. For the current experiment, the drone height was maintained at 15 m, giving a resolution of approximately 1.042 cm/pixel. The captured images encompass five different bands, namely, green, red, blue, near-infrared, and red-edge, with a resolution of 1280× 960, while the resulting image after orthomosaicing achieves an approximate resolution of 4500× 6500. The main goal of the study is to estimate leaf macro/micro nutrients through multispectral images to verify whether a deep learning algorithm is well-suited for the task. To achieve this objective, diverse fertilizer treatments were applied in different stages of crop development. The soil assimilated these fertilizers and subsequently reflected them in the crops. Data collection, like capturing multispectral images and leaf sample collection, was scheduled after 2–3 days of fertilizer treatment. This provides ample time for fertilizer absorption in the leaves before the commencement of actual data collection. Leaf samples of paddy crops have been collected from the fields and utilized as ground truth data for the prediction of various macro/micro nutrients. These samples are collected on the same day as the drone flying activity. Sample collection from each field is done from 8–10 distinct points. The collected leaf samples are analyzed for eleven different macro/micro nutrients in the soil science testing lab in PAU, Ludhiana. These nutrients are measured in mg/kg and include Zinc (Zn), sulfur (S), Phosphorus (P), Sodium (Na), Manganese (Mg), Magnesium (Mn), Potassium (K), Iron (Fe), Copper (Cu), Calcium (Ca), and Boron (B). Data preparation After capturing multispectral images, orthomosaicing was performed for each day of data collection. Orthomosaic images of the field were generated for all five spectral bands, namely Red, Green, Blue, RedEdge, and Near-Infrared (NIR). From each orthomosaic image, regions of interest (ROIs) were extracted per field, each containing all five spectral bands. Subsequently, each ROI was subdivided into non-overlapping 25 × 25 pixel patches, corresponding to an approximate ground coverage of 26 cm2 per patch. In total, 20,988 and 63,737 image patches were obtained for SAS-I and SAS-II, respectively. A geographical overview of the study areas is provided in Fig. 3.Fig. 3Geographical overview of SAS-I and SAS-II: Punjab emphasizing Ludhiana and the selected study areas of Punjab Agricultural University. The map is generated using QGIS version 3.22.11 (https://version.qgis.org). Geographical overview of SAS-I and SAS-II: Punjab emphasizing Ludhiana and the selected study areas of Punjab Agricultural University. The map is generated using QGIS version 3.22.11 (https://version.qgis.org). It was further observed that the average ground truth values of different macro/micro nutrients exhibit a wide numerical range, as summarized in Table 6. For example, the average concentration values vary from 9.316 for Boron, 4478.625 for Calcium, and 174.58 for Iron. Such large variations in magnitude can lead to unstable gradient updates and slow convergence during model training. To mitigate this issue, normalization was applied to all ground truth nutrient values before training, facilitating faster convergence and reduced training loss. Using these normalized target values, all models were trained and evaluated using a spatially consistent splitting strategy, wherein 75% of the data from each field (upper region) was used for training and the remaining 25% (lower region) for testing, for both Study Area–Season I and Season II. This ensures the minimal spatial autocorrelation between training and testing samples, which enables a fair and consistent comparison across all learning approaches. Nutrient estimation model Given that robust predictive performance across varying conditions often requires both a large-scale dataset and a larger network, the proposed neural architecture is designed to be flexible and capable of converging efficiently with minimal parameters. To enable these, we introduce a neural architecture search space with Fourier and polynomial approximators, which will learn a transformed nonlinear feature space that will not only help the model converge faster but also adapt itself according to the type of data. We use four different types of search spaces, namely function approximation search (FAS), adaptive activation search (AAS), layer connection search (LCS), and feature weight search (FWS). To adaptively select the most suitable transformations and architectural components, we employ a differentiable neural architecture search strategy inspired by DARTS using a continuous search space. Specifically, each search space, i.e., FAS (function approximation search), AAS (adaptive activation search), LCS (layer connection search), and FWS (feature weight search), consists of multiple operations. Learnable architecture parameters are associated with each operation, and a softmax relaxation is applied to enable continuous optimization during training. As weights are shared across all candidate operations using a joint optimization of network weights as well as architecture pathways, this provides a cost-effective solution compared to other evolution-based strategies and reinforcement strategies. The final architecture is obtained by selecting the operations with the highest learned importance. Together, these search space components enable the network to adaptively modify its nonlinear feature space, activation behavior, information flow, and feature selection strategy, which enable the network to effectively capture complex spectral nutrient relationships. We first present the overall architecture before elaborating on the different search cells. As shown in Fig. 4, the network begins with an FAS layer. Following the FAS, the network consists of three sequential search cells (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Cell_i$$\end{document}). Each search cell comprises two identical sub-cells (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Cell_{i1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Cell_{i2}$$\end{document}). Within each sub-cell, there are two sets of convolutional operations followed by an FWS, where the first set of convolutions is followed by AAS. Additionally, each sub-cell is followed by an AAS to enhance nonlinear feature representation. After the three search cells, the resulting feature vector is then passed through an average pooling layer and then a fully connected layer with dropout to estimate the 11 nutrient parameters.Fig. 4Architecture of the proposed NAS-based framework used to estimate eleven different leaf nutrients. Architecture of the proposed NAS-based framework used to estimate eleven different leaf nutrients. FAS introduces polynomial basis expansions used to capture complex nonlinear spectral relationships that standard convolution layers may fail to capture efficiently. It includes learnable polynomial functions (e.g., Chebyshev and Fourier bases), which provide an orthogonal representation of spectral signals that is particularly suitable for multispectral data whose reflectance responses exhibit smooth yet nonlinear variations across wavelengths. Unlike standard convolutional filters, these polynomial functions explicitly model higher-order interactions and frequency components, enabling compact yet expressive feature representations under limited data. The input data is represented over an interval [−1,1] to form an orthogonal basis with Legendre or Chebyshev polynomials, ensuring non-redundant and decorrelated non-linearities and compact information with fewer coefficients, such that spectral patterns can be efficiently captured, giving multiscale spatial sensitivity and promoting faster convergence during training. The input of the model is cropped multispectral images with five channels, height H, and width W, represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X} \in \mathbb {R}^{5 \times H \times W}$$\end{document}. The output vector of the model contains the estimated values for 11 nutrients, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Y} \in \mathbb {R}^{11}$$\end{document}. The forward pass of the polynomial feature extraction module is expressed as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {L}_{pq}^{(k)}=\int \int P_p (i)P_q (j)f_k (i,j) d_i d_j \, , \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_k(i,j)$$\end{document} represents the k-th spectral band. To prevent noise and restrict computational complexity, the first four polynomials are used, while local feature computation and band fusion are ensured using learnable coefficients, computed over local windows. The polynomial feature mapping is expressed as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \phi (\textbf{X})=\big [P_0^{\theta } (\textbf{X}),P_1^{\theta } (\textbf{X}),P_2^{\theta } (\textbf{X}),P_3^{\theta } (\textbf{X})] \, , \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_n^{\theta }(\textbf{X})$$\end{document} are Legendre/Chebyshev/Fourier polynomials applied element-wise to each input channel, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta \in \{a, b, c\}$$\end{document}, where a, b, and c correspond to Legendre, Chebyshev, and Fourier polynomials, respectively. The 3rd-order Legendre feature mapping is represented in Eq. (3).3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} P_0^{a} (\textbf{X})&= 1 ,&P_1^{a} (\textbf{X})&= \textbf{X} ,&P_2^{a} (\textbf{X})&= \frac{1}{2}(3\textbf{X}^2 - 1) ,&P_3^{a} (\textbf{X})&= \frac{1}{2}(5\textbf{X}^3 - 3\textbf{X}) \end{aligned} \end{aligned}$$\end{document}Chebyshev polynomials feature mapping is represented in Eq. (4)4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} P_0^{b} (\textbf{X})&= 1 ,&P_1^{b} (\textbf{X})&= \textbf{X},&P_2^{b} (\textbf{X})&= 2\textbf{X}^2 - 1 ,&P_3^{b} (\textbf{X})&= 4\textbf{X}^3 - 3\textbf{X},&P_4^{b} (\textbf{X})&= 8\textbf{X}^4 - 8\textbf{X}^2+1 \end{aligned} \end{aligned}$$\end{document}Fourier feature mapping is represented in Eq. (5)5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} P_0^{c} (\textbf{X})&= \cos (2\pi b_1 x) ,&P_1^{c} (\textbf{X})&= \sin (2\pi b_1 x) ,&P_2^{c} (\textbf{X})&= \cos (2\pi b_2 x) ,&P_3^{c} (\textbf{X})&= \sin (2\pi b_2 x) \end{aligned} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_i$$\end{document} are frequency coefficients. These have good stability near the edges and offer an approximation that enables faster convergence with a lower amount of data. Together, they provide multi-frequency, bounded, decorrelated nonlinear features which improve spatial and spectral generalization. The AAS dynamically combines multiple activation functions using learnable weights, enabling flexible nonlinear relationships in multispectral data. This is particularly beneficial for smaller datasets, which often lead to noisier gradients and irregular losses, causing the model to memorize rather than generalize. AAS mitigates this by acting as a soft constraint, enabling the network to learn which non-linearity fits the data manifold rather than relying on a fixed activation function. It uses a ReLU, GeLU, and learnable sine cosine activation. In a conventional CNN approach, the output of layer k using filter weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f^k(h)$$\end{document} and bias \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b^k$$\end{document} from layer input \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l^{k-1}$$\end{document} is given by Eq. 6:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} l^k= \phi (f^k(h)\circledast l^{k-1} + b^k) \end{aligned}$$\end{document}In contrast, adaptive activation functions are shown in Eq. 7,7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} l^k= \sum _{i=1}^n \alpha _i \phi _i(f^k(h)\circledast l^{k-1} + b^k) \end{aligned}$$\end{document}where the single activation ϕ of Eq. 6 is replaced with a weighted sum of different activation functions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi _i$$\end{document} using a scale factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _i$$\end{document}. Some common choices of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi _i$$\end{document} are relu, sine, etc. This enables the model to choose activations dynamically, improving robustness across diverse data. LCS enables adaptive selection between skip and dense connectivity patterns, improving gradient flow and feature reuse across layers to learn optimal information flow instead of fixing skip or dense connections. For smaller datasets, skip connections prevent overfitting, while dense connections enhance feature reuse and learning efficiency. The NAS framework adaptively discovers the most appropriate connection to a layer from previous layers, as shown in Fig. 4. FWS assigns adaptive importance to feature maps, enabling the model to focus on informative spectral features while suppressing noise. It is used to emphasize important spectral channels and reduce irrelevant ones. FWS searches the optimal weights of a feature map. These can include using equal weights for features, using partial feature weights, or using entire feature weights. In partial feature weights, half of the features in a layer are used as they are, while feature attention is computed over the other half. Together, these search space components enable the network to adaptively modify its nonlinear feature space, activation behavior, information flow, and feature selection strategy, which enable the network to effectively capture complex spectral nutrient relationships. Results This section details the various experiments performed with different models, data preprocessing, and splits. All experiments are performed with a learning rate = 0.01 and epochs = 100. The experiments include: (i) estimation of leaf macro/micro nutrients using conventional machine learning models (Section "Estimating Leaf Macro/Micro Nutrients Using Conventional Machine Learning Techniques"); (ii) estimation using deep learning models for both study areas in Season I and Season II (Section "Estimating Leaf Macro/Micro Nutrients using Deep Learning Baseline Models for SAS-I and SAS-II"); (iii) estimation using neural architecture search with polynomial approximation functions for Season I and Season II (Section "Estimating Leaf Macro/Micro Nutrients using Neural Architecture Search with Polynomial Approximation Functions for SAS-I and SAS-II"); (iv) growth stage–wise and treatment-wise estimation of leaf macro/micro nutrients to analyze observed variations (Section "Growth Stage-wise and Treatment-wise Estimation of Eleven Leaf Nutrients"); (v) cross validation and generalization analysis of the proposed NAS-based architecture (Section "Cross Validation and Generalization Analysis of the proposed NAS-based Architecture"); (vi) estimating leaf macro/micro nutrients through vegetation indices (Section "Estimating Leaf Macro/Micro Nutrients using Vegetation Indices"); and (vii) ablation analysis of the proposed model (Section "Ablation Analysis of the Proposed Model"). In addition, an experiment was conducted to analyze input channel pixel similarity and variance with respect to growth stages and treatment types. Four sample sets are selected from the entire dataset such that Set1 and Set2 belong to the same growth stage and treatment [Growth Stage 1 Treatment 1]-(GS1 T1). Set3 belongs to a different growth stage, but the same treatment [Growth Stage 2 Treatment 1]-(GS2 T1), and Set4 represents data samples from a different growth stage and different treatment [Growth Stage 2 Treatment 2]-(GS2 T2). This approach ensures that the sample sets include variations, which is essential for assessing model performance across diverse conditions. The experiment hypothesizes that the data from similar growth stages and treatments (e.g., Set1 vs Set2) will exhibit similar statistics. Conversely, noticeable variations are expected between samples from different growth stages (e.g., Set1 vs Set3) or treatments (e.g., Set3 vs Set4).Fig. 5Gaussian distribution using pixel mean and variances across all channels for four sample sets [GS1 T1, GS1 T1, GS2 T1, GS2 T2].Table 1Mean, Standard deviation, and Coefficient of variation across different channels for four sample sets [GS1 T1, GS1 T1, GS2 T1, GS2 T2].BandsGS1 T1 (Set1)GS1 T1 (Set2)GS2 T1 (Set3)GS2 T2 (Set4)MeanSDCVMeanSDCVMeanSDCVMeanSDCVRed42.711.7227.4541.3310.9626.5229.435.6219.138.49.4924.71Green102.611.711.41100.6311.7411.6657.6710.0117.3684.1521.6825.76Blue80.967.419.1578.667.018.9137.625.7615.345.578.5918.55NIR147.191510.19145.9715.7210.77137.9721.7215.74175.8532.0518.23RedEdge133.8418.5413.85131.8917.1112.9884.212.5514.91140.0734.7824.82 Gaussian distribution using pixel mean and variances across all channels for four sample sets [GS1 T1, GS1 T1, GS2 T1, GS2 T2]. Table 1Mean, Standard deviation, and Coefficient of variation across different channels for four sample sets [GS1 T1, GS1 T1, GS2 T1, GS2 T2].BandsGS1 T1 (Set1)GS1 T1 (Set2)GS2 T1 (Set3)GS2 T2 (Set4)MeanSDCVMeanSDCVMeanSDCVMeanSDCVRed42.711.7227.4541.3310.9626.5229.435.6219.138.49.4924.71Green102.611.711.41100.6311.7411.6657.6710.0117.3684.1521.6825.76Blue80.967.419.1578.667.018.9137.625.7615.345.578.5918.55NIR147.191510.19145.9715.7210.77137.9721.7215.74175.8532.0518.23RedEdge133.8418.5413.85131.8917.1112.9884.212.5514.91140.0734.7824.82 Mean, Standard deviation, and Coefficient of variation across different channels for four sample sets [GS1 T1, GS1 T1, GS2 T1, GS2 T2]. Initially, the Gaussian distribution using the sample mean and variance is visualized in Fig. 5 to analyze the pixel variations in each channel for the four sample sets. The results show that pixel variations are more closely aligned within sample sets that belong to the same growth stage and treatment (e.g., Set1 and Set2). In contrast, significant variation is observed between sample sets from different growth stages (e.g., Set1 vs Set3/Set4). A smaller degree of variation is noted between sample sets from the same growth stage but with different treatments (e.g., Set3 vs Set4). The observation is further supported by the statistical metrics shown in Table 1, which provides the mean, standard deviation (SD), and coefficient of variance(CV) for each sample set. The table indicates that samples from the same growth stage and treatment exhibit similar statistical characteristics, whereas notable differences are observed across other cases. Estimating leaf macro/micro nutrients using conventional machine learning techniques In this study, several conventional machine learning techniques were applied to the SAS-I dataset to establish a strong comparative baseline for leaf macro/micro nutrient estimation. The objective of this analysis is to evaluate the effectiveness of the proposed deep learning framework by comparing its performance against widely used traditional regression models with mean absolute error (MAE) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} as the evaluation metric. The conventional machine learning (CML) methods considered in this work include Linear Regression (LR), Random Forest (RF), Support Vector Regression (SVR), Extreme Gradient Boosting (XGB), Lasso Regression (LaR), Ridge Regression (RdR), Multi-Layer Perceptron (MLP), Partial Least Square Regression (PLSR), and Decision Tree (DT). For all models, the input multispectral image patches of size 25×25 are converted into feature vectors. As multispectral data consist of five spectral bands, which are then flattened and concatenated to get a feature vector of size 25×25×5. This 3125-dimensional vector serves as an input to the CML techniques. The output vector of the model is 11-dimensional, which represents the eleven leaf macro/micro nutrients. Table 2 presents the obtained MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values for all eleven macro/micro leaf nutrients using the CML techniques. The ground-truth values are normalized before training. The corresponding hyperparameters for each technique are reported in 2. The results demonstrate that XGB, MLP, and RF comparatively perform better by consistently achieving lower MAE values ranging from 0.07 to 0.14 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values greater than 0.30 across most of the nutrients. In contrast, LaR and LR perform poorly and exhibit consistently high MAE and negative \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values.Table 2MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} obtained for eleven leaf macro/micro nutrients using different machine learning techniques for SAS-I after normalizing ground-truth values, along with their corresponding hyperparameters.ModelHyper-parametersMetricBCaCuFeKMgMnNaPSZnSVRC=1.0,MAE0.160.120.120.120.100.100.140.150.120.120.11epsilon=0.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.210.520.600.220.440.370.180.670.180.700.26XGBn estimators=500,max depth=3,MAE0.130.090.100.090.080.070.110.120.090.080.09learning rate=0.05,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.430.660.700.510.600.610.300.740.520.770.47random state=42RFn estimators=200,min samples leaf=3,MAE0.130.100.110.090.080.080.110.140.090.070.09random state=42,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.380.630.670.440.500.490.300.650.420.780.36n jobs=−1LRMAE0.180.150.140.130.120.110.150.210.130.120.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.100.280.360.010.060.08−0.040.320.020.570.06Lassoalpha=10.0,MAE0.180.190.200.140.130.120.150.300.130.210.11max iter=20000\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}00000000000MLPhidden layer sizes=(512, 256),activation=’relu’, solver=’adam’,alpha=0.0001, max iter=1000,MAE0.0.120.090.100.100.080.080.100.100.090.070.08learning rate=’adaptive’,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.420.680.710.210.620.600.310.810.460.820.54random state=1Ridgealpha = 10.0MAE0.160.130.130.120.110.100.130.190.120.110.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.100.400.460.190.220.250.150.410.150.630.21DTmin samples split=10,max depth=10,MAE0.150.110.130.110.090.090.120.160.100.090.10min samples leaf=5,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.110.430.470.060.290.250.030.360.120.640.07random state=0PLSRn components=10MAE0.160.130.120.110.100.100.130.190.120.110.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.110.420.470.210.240.260.170.420.160.640.23 Table 2MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} obtained for eleven leaf macro/micro nutrients using different machine learning techniques for SAS-I after normalizing ground-truth values, along with their corresponding hyperparameters.ModelHyper-parametersMetricBCaCuFeKMgMnNaPSZnSVRC=1.0,MAE0.160.120.120.120.100.100.140.150.120.120.11epsilon=0.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.210.520.600.220.440.370.180.670.180.700.26XGBn estimators=500,max depth=3,MAE0.130.090.100.090.080.070.110.120.090.080.09learning rate=0.05,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.430.660.700.510.600.610.300.740.520.770.47random state=42RFn estimators=200,min samples leaf=3,MAE0.130.100.110.090.080.080.110.140.090.070.09random state=42,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.380.630.670.440.500.490.300.650.420.780.36n jobs=−1LRMAE0.180.150.140.130.120.110.150.210.130.120.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.100.280.360.010.060.08−0.040.320.020.570.06Lassoalpha=10.0,MAE0.180.190.200.140.130.120.150.300.130.210.11max iter=20000\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}00000000000MLPhidden layer sizes=(512, 256),activation=’relu’, solver=’adam’,alpha=0.0001, max iter=1000,MAE0.0.120.090.100.100.080.080.100.100.090.070.08learning rate=’adaptive’,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.420.680.710.210.620.600.310.810.460.820.54random state=1Ridgealpha = 10.0MAE0.160.130.130.120.110.100.130.190.120.110.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.100.400.460.190.220.250.150.410.150.630.21DTmin samples split=10,max depth=10,MAE0.150.110.130.110.090.090.120.160.100.090.10min samples leaf=5,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.110.430.470.060.290.250.030.360.120.640.07random state=0PLSRn components=10MAE0.160.130.120.110.100.100.130.190.120.110.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.110.420.470.210.240.260.170.420.160.640.23 MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} obtained for eleven leaf macro/micro nutrients using different machine learning techniques for SAS-I after normalizing ground-truth values, along with their corresponding hyperparameters. Estimating leaf macro/micro nutrients using deep learning baseline models for SAS-I and SAS-II This study also analyzed several convolutional neural network (CNN) backbone architectures, including ResNet18, ResNet34, GoogleNet, LeNet5, AlexNet, VGG16, Swin Transformer, ConvNeXt, DenseNet121, MobileNet, Inception, and EfficientNet for estimating eleven leaf macro/micro nutrients. To ensure a fair comparison, all networks were trained using identical hyperparameters of 100 epochs, batch size of 32, learning rate of 0.01, and weight decay of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1e^{-4}$$\end{document}, ensuring consistency in optimization conditions. The hyperparameters were selected based on preliminary experiments exploring different learning rates and regularization strengths. All models were adapted to accept five spectral input channels while producing eleven continuous output values corresponding to nutrient concentrations. Model performance was assessed using MAE and the coefficient of determination (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}). The results obtained for SAS-I and SAS-II are summarized in Tables 3 and 4, respectively. For SAS-I, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} performance ranking is as follows: EfficientNet > DenseNet121 > MobileNet > ResNet34 > GoogleNet > ResNet18 > Alexnet > Inception, accompanied by relatively lower MAE values for all parameters, indicating stronger agreement between predicted and measured nutrient concentrations. A similar trend is observed for SAS-II, with the performance order: DenseNet121 > EfficientNet > ResNet34 > Inception > ResNet18 > Alexnet > MobileNet > GoogleNet. However, the overall performance of SAS-II is lower than that of SAS-I, indicating greater variability and reduced model generalization under the second seasonal condition. These results suggest that EfficientNet, DenseNet, and residual learning mechanisms are performing better in capturing the complex, nonlinear relationships between multispectral features and nutrient concentrations. In contrast, VGG16 and Swin Transformer, ConvNeXt exhibit poor performance with negative \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values for both SAS-I and SAS-II, which may be attributed to the fully-connected layer size of VGG, attention collapse for SWIN, and GeLU miscalibration for ConvNeXt, given the dataset size and configuration. Driven by the fact that multispectral datasets for nutrient estimations are generally small in size and not reusable due to their dependency on weather, crop type, soil type, a single model may not suit different data. This motivates the exploration of automated architecture optimization strategies capable of learning task-specific representations more efficiently. Accordingly, the next section investigates a neural architecture search (NAS) framework augmented with polynomial approximation functions to further enhance nutrient estimation accuracy.Table 3MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values obtained for deep learning baselines after normalizing ground-truth values for SAS-I.ModelMetricBCaCuFeKMgMnNaPSZnVGGMAE0.170.180.200.130.130.120.140.30.120.190.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.080−0.01−0.0300−0.0400−0.130AlexNetMAE0.120.090.110.100.090.090.100.080.100.070.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.260.530.540.220.270.220.130.850.160.800.27SwinTransformerMAE0.170.190.200.130.130.120.150.290.130.200.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.130−0.01−0.02000−0.12−0.05−0.120DenseNet121MAE0.130.100.110.100.090.080.110.140.090.080.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.330.570.590.280.430.420.210.610.400.750.31GoogleNetMAE0.130.110.120.100.090.090.110.110.100.080.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.230.500.600.290.310.360.170.700.210.730.21EfficientNetMAE0.140.100.100.100.090.080.110.140.090.060.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.150.600.670.330.490.470.260.600.480.850.32ResNet18MAE0.120.100.110.100.090.090.110.100.100.080.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.230.500.560.300.290.370.110.730.120.760.15ResNet34MAE0.120.110.110.090.090.080.110.090.100.070.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.250.470.630.310.350.390.110.810.270.780.20InceptionMAE0.130.100.120.100.090.090.110.120.100.080.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.140.520.570.260.310.300.170.630.210.760.19ConvNeXtMAE0.170.190.200.130.130.120.150.300.130.200.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.08−0.01−0.01−0.010−0.010−0.02−0.01−0.090MobileNetMAE0.140.150.120.110.090.080.110.160.100.090.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.250.290.560.260.510.490.250.590.400.740.36Table 4Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for DL after normalizing ground-truth values for SAS-II.ModelMetricBCaCuFeKMgMnNaPSZnVGGMAE0.140.200.200.110.210.170.190.190.150.230.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.00−0.01−0.17−0.11−0.0200−0.08−0.01−0.070AlexNetMAE0.100.170.140.080.150.110.160.110.120.130.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.360.120.270.350.320.450.010.540.200.550.09SwinTransformerMAE0.140.200.200.110.210.170.190.190.150.230.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0−0.01−0.12−0.09−0.01−0.010−0.050−0.08−0.01DenseNet121MAE0.090.170.130.080.140.100.160.110.110.110.06\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.450.220.370.420.400.480.090.570.210.710.13GoogleNetMAE0.100.190.150.090.140.120.170.120.130.140.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.3100.110.340.330.370.020.490.040.520EfficientNetMAE0.100.160.130.080.130.090.150.100.110.110.06\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.290.220.370.360.400.520.180.550.270.660.15ResNet18MAE0.100.170.140.080.140.110.160.110.120.130.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.370.190.230.290.400.480.180.540.220.590.07ResNet34MAE0.090.180.140.080.140.100.160.110.120.130.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.410.210.370.360.410.490.100.520.280.630.09InceptionMAE0.100.170.130.080.150.110.170.110.110.130.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.370.270.380.230.430.450.160.510.310.640.14ConvNeXtMAE0.140.200.200.110.210.170.190.190.150.230.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}00−0.13−0.050−0.010−0.04−0.01−0.03−0.01MobileNetMAE0.100.180.160.090.150.110.170.120.120.150.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.290.090.270.230.380.430.110.420.260.530.03 Table 3MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values obtained for deep learning baselines after normalizing ground-truth values for SAS-I.ModelMetricBCaCuFeKMgMnNaPSZnVGGMAE0.170.180.200.130.130.120.140.30.120.190.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.080−0.01−0.0300−0.0400−0.130AlexNetMAE0.120.090.110.100.090.090.100.080.100.070.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.260.530.540.220.270.220.130.850.160.800.27SwinTransformerMAE0.170.190.200.130.130.120.150.290.130.200.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.130−0.01−0.02000−0.12−0.05−0.120DenseNet121MAE0.130.100.110.100.090.080.110.140.090.080.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.330.570.590.280.430.420.210.610.400.750.31GoogleNetMAE0.130.110.120.100.090.090.110.110.100.080.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.230.500.600.290.310.360.170.700.210.730.21EfficientNetMAE0.140.100.100.100.090.080.110.140.090.060.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.150.600.670.330.490.470.260.600.480.850.32ResNet18MAE0.120.100.110.100.090.090.110.100.100.080.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.230.500.560.300.290.370.110.730.120.760.15ResNet34MAE0.120.110.110.090.090.080.110.090.100.070.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.250.470.630.310.350.390.110.810.270.780.20InceptionMAE0.130.100.120.100.090.090.110.120.100.080.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.140.520.570.260.310.300.170.630.210.760.19ConvNeXtMAE0.170.190.200.130.130.120.150.300.130.200.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.08−0.01−0.01−0.010−0.010−0.02−0.01−0.090MobileNetMAE0.140.150.120.110.090.080.110.160.100.090.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.250.290.560.260.510.490.250.590.400.740.36 MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values obtained for deep learning baselines after normalizing ground-truth values for SAS-I. Table 4Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for DL after normalizing ground-truth values for SAS-II.ModelMetricBCaCuFeKMgMnNaPSZnVGGMAE0.140.200.200.110.210.170.190.190.150.230.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.00−0.01−0.17−0.11−0.0200−0.08−0.01−0.070AlexNetMAE0.100.170.140.080.150.110.160.110.120.130.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.360.120.270.350.320.450.010.540.200.550.09SwinTransformerMAE0.140.200.200.110.210.170.190.190.150.230.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0−0.01−0.12−0.09−0.01−0.010−0.050−0.08−0.01DenseNet121MAE0.090.170.130.080.140.100.160.110.110.110.06\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.450.220.370.420.400.480.090.570.210.710.13GoogleNetMAE0.100.190.150.090.140.120.170.120.130.140.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.3100.110.340.330.370.020.490.040.520EfficientNetMAE0.100.160.130.080.130.090.150.100.110.110.06\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.290.220.370.360.400.520.180.550.270.660.15ResNet18MAE0.100.170.140.080.140.110.160.110.120.130.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.370.190.230.290.400.480.180.540.220.590.07ResNet34MAE0.090.180.140.080.140.100.160.110.120.130.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.410.210.370.360.410.490.100.520.280.630.09InceptionMAE0.100.170.130.080.150.110.170.110.110.130.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.370.270.380.230.430.450.160.510.310.640.14ConvNeXtMAE0.140.200.200.110.210.170.190.190.150.230.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}00−0.13−0.050−0.010−0.04−0.01−0.03−0.01MobileNetMAE0.100.180.160.090.150.110.170.120.120.150.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.290.090.270.230.380.430.110.420.260.530.03 Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for DL after normalizing ground-truth values for SAS-II. Estimating leaf macro/micro nutrients using neural architecture search with polynomial approximation functions for SAS-I and SAS-II In this experiment, the SAS-I and SAS-II datasets were trained and evaluated independently using the proposed NAS–based framework augmented with polynomial approximation functions. The quantitative performance of the proposed model for SAS-I and SAS-II is summarized in Table 5. As observed, the proposed approach achieves consistently low prediction errors for SAS-I, with MAE values ≤ 0.11 across all eleven macro/micro nutrients, along with improved \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values, indicating a better fit between predicted and ground truth values. In contrast, relatively higher MAE values (up to 0.16) are observed for SAS-II compared to SAS-I, reflecting increased prediction difficulty across seasons. To further evaluate the reliability of the predicted nutrients beyond the prediction error analysis, uncertainty estimation of nutrients is performed, which is presented in Section “Discussion”. When compared with predefined deep learning backbone architectures (Tables 3 and 4), the NAS-based model demonstrates superior performance for both SAS-I and SAS-II data. For SAS-I, substantial improvements are observed in both MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} across all nutrients. Additionally, Table 6 presents the results after transforming the predictions back to their original physical units (mg/kg). The table reports the mean target values, mean predicted values, and MAE for each nutrient for both seasons. It can be observed that the prediction errors remain low in absolute terms; for example, for Cu, the mean target value is 8.16 mg/kg, the mean predicted value is 8.32 mg/kg, and the corresponding MAE is 1.16 mg/kg, indicating good agreement between predicted and actual values. These results highlight the advantage of automatically learning task-specific network architectures that are more relevant to capture complex spectral features in multispectral agricultural imagery. The incorporation of polynomial approximation functions further enhances the model’s ability to represent nonlinear feature transformations, leading to improved regression accuracy. Motivated by these results, additional experiments are conducted to analyze the robustness of the proposed model under varying agronomic conditions, including growth stage-wise and treatment-wise evaluations, to further investigate variations in nutrient estimation performance.Table 5MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values obtained with the proposed model for SAS-I and SAS-II.NutrientsBCaCuFeKMgMnNaPSZn SAS-IMAE0.110.090.090.080.080.070.100.090.080.060.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.680.750.380.620.600.340.810.580.860.53 SAS-IIMAE0.090.160.130.080.140.100.150.110.120.110.06\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.470.210.370.450.320.520.140.630.140.710.18Table 6Mean target, mean predicted, and their corresponding MAE across all macro/micro leaf nutrients in mg/kg.NutrientsBCaCuFeKMgMnNaPSZnMean Target SAS-I9.31654478.6258.16375174.5847516241.4251106.77533.2945186.55351590.951624.47525.17125Mean Prediction SAS-I8.3854781.8628.332173.87816374.2571082.55533.544178.3381600.9441634.77424.972MAE SAS-I2.129628.0761.16736.6941347.164151.4585.59025.870176.506179.7142.611 Table 5MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values obtained with the proposed model for SAS-I and SAS-II.NutrientsBCaCuFeKMgMnNaPSZn SAS-IMAE0.110.090.090.080.080.070.100.090.080.060.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.680.750.380.620.600.340.810.580.860.53 SAS-IIMAE0.090.160.130.080.140.100.150.110.120.110.06\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.470.210.370.450.320.520.140.630.140.710.18 MAE and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values obtained with the proposed model for SAS-I and SAS-II. Table 6Mean target, mean predicted, and their corresponding MAE across all macro/micro leaf nutrients in mg/kg.NutrientsBCaCuFeKMgMnNaPSZnMean Target SAS-I9.31654478.6258.16375174.5847516241.4251106.77533.2945186.55351590.951624.47525.17125Mean Prediction SAS-I8.3854781.8628.332173.87816374.2571082.55533.544178.3381600.9441634.77424.972MAE SAS-I2.129628.0761.16736.6941347.164151.4585.59025.870176.506179.7142.611 Mean target, mean predicted, and their corresponding MAE across all macro/micro leaf nutrients in mg/kg. Growth stage-wise and treatment-wise estimation of eleven leaf nutrients A growth stage–wise evaluation was conducted to analyze the variation in macro/micro nutrient estimation across different crop developmental phases. The multispectral data were grouped according to four distinct crop growth stages, denoted as GS1, GS2, GS3, and GS4, corresponding to the respective data collection periods. Table 7 presents the growth stage–wise MAE values obtained for eleven leaf nutrients. The results indicate that the proposed model achieves consistently lower estimation errors across all growth stages, with GS1 exhibiting comparatively lower MAE, which is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\le 8\%$$\end{document} for most nutrients. GS2 shows a slight increase in error for certain nutrients, such as Na and P, indicating moderate variability in estimation performance. In GS3 and GS4, the MAE values fluctuate across nutrients, with some parameters (e.g., Fe and S) showing lower errors, while others (e.g., B and Ca) exhibit relatively higher deviations. Nevertheless, the MAE values across all stages remain within an acceptable range, demonstrating the robustness of the proposed approach in capturing nutrient dynamics throughout the crop growth stages.Table 7Growth stage-wise MAE obtained for eleven Leaf nutrients after normalizing target values.NutrientsBCaCuFeKMgMnNaPSZnGS10.060.070.10.170.050.060.130.060.050.080.08GS20.120.120.10.080.100.070.070.130.110.090.1GS30.130.070.080.040.070.070.110.10.110.050.05GS40.120.110.10.040.090.070.070.080.070.030.11 Table 7Growth stage-wise MAE obtained for eleven Leaf nutrients after normalizing target values.NutrientsBCaCuFeKMgMnNaPSZnGS10.060.070.10.170.050.060.130.060.050.080.08GS20.120.120.10.080.100.070.070.130.110.090.1GS30.130.070.080.040.070.070.110.10.110.050.05GS40.120.110.10.040.090.070.070.080.070.030.11 Growth stage-wise MAE obtained for eleven Leaf nutrients after normalizing target values. A treatment-wise evaluation was conducted for Study Area–Season I to analyze the impact of different treatments on leaf macro/micro nutrient estimation accuracy. As described in Section “Treatment Details”, plots P2, P5, and P8 received Treatment 1 (T1); P3, P6, and P9 received Treatment 2 (T2); and P4, P7, and P10 received Treatment 3 (T3), while plot P1 served as the control (C) without any treatment. Hence, three distinct treatments were applied across different fields during the season. The treatment-wise MAE values for eleven leaf nutrients are presented in Table 8. The results indicate that the proposed model maintains stable and consistently low estimation errors across all treatments. As compared to C, other treatments, T1, T2, and T3, generally exhibit reduced MAE values for most nutrients. This highlights the stable and improved prediction under varied nutrient conditions. Minor variations are observed among treatments for certain nutrients, like B in T2. Additionally, relatively higher errors observed for the control condition may be influenced by the smaller sample size, as the control group includes only a single plot, whereas each treatment group consists of three plots. Despite these variations, the overall performance remains comparable, demonstrating the robustness of the proposed model across different treatments.Table 8Treatment-wise MAE obtained for eleven Leaf nutrients after normalizing target values.NutrientsBCaCuFeKMgMnNaPSZnC0.150.10.100.130.110.220.080.160.140.100.12T10.090.090.100.060.080.060.160.10.080.060.07T20.120.090.10.060.080.060.050.070.070.050.08T30.10.090.090.120.070.060.080.090.090.070.09 Table 8Treatment-wise MAE obtained for eleven Leaf nutrients after normalizing target values.NutrientsBCaCuFeKMgMnNaPSZnC0.150.10.100.130.110.220.080.160.140.100.12T10.090.090.100.060.080.060.160.10.080.060.07T20.120.090.10.060.080.060.050.070.070.050.08T30.10.090.090.120.070.060.080.090.090.070.09 Treatment-wise MAE obtained for eleven Leaf nutrients after normalizing target values. Cross validation and generalization analysis of the proposed NAS-based architecture To evaluate the cross-validation of the proposed NAS-based framework, a treatment-wise cross-validation was performed to estimate leaf macro/micro nutrients. Since the dataset consists of three treatments (T1, T2, T3) and one control (C), with no treatment, multiple experiments were conducted by training the model on three groups and testing on the remaining unseen group. For instance, the model was trained using samples with T1, T2, and T3 and evaluated on the control field samples. Similarly, other combinations were also explored, such as training on samples with T2, T3 treatments with Control field samples, and testing on the field samples with T1, and so on. The results obtained are presented in Table 9. It is observed that the model performs consistently for several nutrients. Specifically, Cu and Na achieve positive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values across all treatments. Nutrients such as B and Ca also show stable performance, with positive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values across most treatments, except for T3. Similarly, S shows positive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} across all treatments except the control (C). For Fe, K, Mg, and P, the model achieves positive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values primarily for treatments T1 and T2, while Mn shows positive performance for T2 and T3. In contrast, Zn exhibits consistently negative \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values across all treatments, indicating that its prediction remains challenging, likely due to weaker spectral sensitivity or higher variability in the dataset. Furthermore, plot-wise validation was performed to evaluate the generalization capability of the proposed model across unseen plots. In this experiment, the model was trained using 8 plots, and 2 were used for testing. This combination was taken such that the two testing plots belonged to 2 different treatments. This was repeated across different 8:2 sets. The results presented in Table 10 show relatively low MAE values (less than 14%) across all nutrients. For most of the nutrients, positive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values were obtained, indicating reasonable predictive capability under unseen plot conditions. However, Mg and Mn exhibited negative \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values, while Zn showed comparatively weak performance, suggesting that certain nutrients remain challenging to estimate due to weaker spectral sensitivity, limited variability, or complex physiological interactions. Overall, these results confirm that the proposed model exhibits reasonable generalization across unseen treatments and plots, while also highlighting the limitations and challenges associated with specific nutrients, which may require further investigation.Table 9Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model treatment-wise cross-validation for SAS-I data.NutrientsBCaCuFeKMgMnNaPSZnTrain: T1, T2, T3MAE0.140.120.110.160.200.240.120.140.180.130.13Test: C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.510.540.64−0.10−6.38−1.87−20.73−2.32−0.16−0.13Train: C, T2, T3MAE0.100.090.120.070.080.070.180.090.090.070.09Test: T1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.590.690.170.450.510.46−0.730.800.500.84−0.05Train: C, T1, T3MAE0.140.120.120.070.090.090.080.100.110.060.10Test: T2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.060.420.630.150.400.390.360.850.300.90−0.26Train: C, T1, T2MAE0.180.150.150.180.120.110.140.120.150.100.14Test: T3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.3600.29−0.14−0.04−0.470.040.55−0.830.58−0.06Table 10Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model plot-wise cross-validation for SAS-I data.NutrientsBCaCuFeKMgMnNaPSZnMAE0.120.110.100.110.110.070.080.130.040.070.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.530.600.380.230.36−0.81−0.130.780.700.490.05 Table 9Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model treatment-wise cross-validation for SAS-I data.NutrientsBCaCuFeKMgMnNaPSZnTrain: T1, T2, T3MAE0.140.120.110.160.200.240.120.140.180.130.13Test: C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.510.540.64−0.10−6.38−1.87−20.73−2.32−0.16−0.13Train: C, T2, T3MAE0.100.090.120.070.080.070.180.090.090.070.09Test: T1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.590.690.170.450.510.46−0.730.800.500.84−0.05Train: C, T1, T3MAE0.140.120.120.070.090.090.080.100.110.060.10Test: T2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.060.420.630.150.400.390.360.850.300.90−0.26Train: C, T1, T2MAE0.180.150.150.180.120.110.140.120.150.100.14Test: T3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.3600.29−0.14−0.04−0.470.040.55−0.830.58−0.06 Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model treatment-wise cross-validation for SAS-I data. Table 10Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model plot-wise cross-validation for SAS-I data.NutrientsBCaCuFeKMgMnNaPSZnMAE0.120.110.100.110.110.070.080.130.040.070.07\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.530.600.380.230.36−0.81−0.130.780.700.490.05 Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model plot-wise cross-validation for SAS-I data. Estimating leaf macro/micro nutrients using vegetation indices In this section, we evaluate the effectiveness of vegetation indices (VIs) for estimating leaf macro/micro nutrient concentrations. Specifically, three aspects are explored: (i) the impact of using VIs as standalone inputs, (ii) their effectiveness when incorporated as complementary features alongside five spectral bands, and (iii) the impact of feature selection in identifying the most informative vegetation indices and spectral bands for nutrient estimation. Initially, VIs were incorporated as standalone inputs to the best-performing deep learning baseline model, namely EfficientNet, for our dataset to investigate their impact on estimating different macro/micro nutrients. Several widely used indices such as normalized difference vegetation index (NDVI), green normalized difference vegetation index (GNDVI), Chlorophyll Index (CI), normalized difference water index (NDWI), modified chlorophyll absorption in reflectance index (MCARI), normalized difference red-edge (NDRE), optimized soil adjusted vegetation index (OSAVI), soil adjusted vegetation index (SAVI), ratio spectral index (RESI), were employed in our study for analyzing macro/micro nutrients. The corresponding results are shown in Table 11, indicating poor performance with a mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} of 0.18 across all nutrients, including negative \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values observed for certain nutrients such as B and Zn. Subsequently, VIs were incorporated as complementary features alongside the five spectral bands. This combination resulted in improved performance compared to using VIs alone, attaining a mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} of 0.39. To further capture the most relevant spectral features for nutrient estimation, feature selection was performed. Initially, channel-wise features were extracted from an image patch denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_c}$$\end{document} with height H, and width W using global average pooling, as shown in Eq. 8.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F_c = \frac{1}{HW} \sum _{i=1}^{H} \sum _{j=1}^{W} X_{c,i,j} \end{aligned}$$\end{document}This results in a feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F \in \mathbb {R}^{N \times C}$$\end{document}, where N denotes the number of samples and c represents the number of spectral channels. Subsequently, Pearson correlation analysis was performed using Eq. 9 to evaluate linear relationships between each feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_j$$\end{document} and nutrient concentration \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_k$$\end{document}, with the results presented in Fig. 6. 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r_{jk} = \frac{\sum _{i=1}^{N} (F_{i,j} - \bar{F}_j)(Y_{i,k} - \bar{Y}_k)}{\sqrt{\sum _{i=1}^{N} (F_{i,j} - \bar{F}_j)^2} \; \sqrt{\sum _{i=1}^{N} (Y_{i,k} - \bar{Y}_k)^2}} \end{aligned}$$\end{document}Additionally, non-linear dependencies were captured using mutual information (M), as defined in Eq. 10, which was estimated using a non-parametric k-nearest neighbor (KNN)-based approach.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} M(F_j; Y_k) = \int _{f_j} \int _{y_k} p(f_j, y_k)\log \left( \frac{p(f_j, y_k)}{p(f_j)p(y_k)}\right) d_{f{_j}} d_{y{_k}} \end{aligned}$$\end{document}Finally, global feature importance, shown in Fig 7, was computed by averaging mutual information scores across all nutrients to select the most informative features, as shown in Eq. 11.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} G_j = \frac{1}{K} \sum _{k=1}^{K} M(F_j; Y_k) \end{aligned}$$\end{document}Fig. 6Pearson Correlation between spectral features and leaf macro/micro nutrients.Fig. 7Global feature importance for all macro/micro nutrients. Pearson Correlation between spectral features and leaf macro/micro nutrients. Global feature importance for all macro/micro nutrients. It is observed that several vegetation indices (NDVI, GNDVI, NDRE, OSAVI, SAVI, NDWI, RESI) exhibit highly similar correlation and mutual information patterns, indicating redundancy due to shared spectral dependencies. Therefore, a subset of representative indices was selected based on feature importance to reduce dimensionality and avoid multicollinearity. Based on this analysis, a compact and informative feature set comprising blue, red, red-edge, MCARI, and GNDVI was selected for model training. The results obtained using this selected feature set (Table 11) show improved performance with a mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} value of 0.33 compared to using only VIs. To determine whether the inference is model-dependent, the same experiments were conducted using the Resnet18 baseline. A similar trend was observed, with mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values of 0.19 (VIs only), 0.32 (feature-selected inputs), and 0.40 (all bands and indices), confirming the consistency of the findings across architectures. Additionally, we have also repeated the above three experiments for our NAS-based architecture to verify if using 5 spectral bands outperforms other combinations. Results in Table 12 confirm that the proposed NAS-based model using only five spectral bands achieves a higher mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} of 0.59, outperforming configurations using only VIs with a mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} of 0.19, combined bands and VIs with a mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} of 0.55, and feature-selected inputs with a mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} of 0.34. The results indicate that the neural architectures (baseline as well as proposed) are capable of extracting suitable spectral features required for mapping the nutrient information, hence providing VIs explicitly alongside the bands do not show significant improvement in model performances.Table 11Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained SAS-I using (i) only vegetation indices (VIs), (ii) a combination of spectral bands (SB) and VIs, and (iii) their feature-selected subset (FS), evaluated using deep learning baseline models.NutrientsBCaCuFeKMgMnNaPSZn EfficientNet with VIsMAE0.160.120.140.120.100.100.120.220.110.130.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.120.400.360.140.080.230.170.210.130.47−0.08 EfficientNet with SB+VIsMAE0.130.120.120.100.090.080.120.120.100.090.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.130.400.630.320.350.450.120.740.330.760.07 EfficientNet with FSMAE0.130.100.120.110.100.090.110.130.110.080.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.170.530.510.240.220.250.160.700.110.740.03 ResNet18 with VIsMAE0.160.130.150.120.100.100.120.220.110.130.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.040.350.300.130.120.260.160.220.140.420.05 ResNet18 with SB+VIsMAE0.120.100.110.090.090.080.110.100.110.070.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.280.540.570.300.320.360.170.780.170.770.14 ResNet18 with FSMAE0.140.110.120.110.090.090.110.160.110.090.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.130.510.530.180.210.250.200.560.150.690.13Table 12Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained SAS-I using (i) only vegetation indices (VIs), (ii) a combination of spectral bands (SB) and VIs, and (iii) their feature-selected subset (FS), evaluated using the proposed NAS-based model.NutrientsBCaCuFeKMgMnNaPSZn Proposed Model with VIsMAE0.150.130.140.120.100.100.120.210.110.130.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.020.400.350.120.140.220.190.150.130.370.04 Proposed Model with SB+VIsMAE0.110.090.100.080.080.070.100.080.090.060.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.660.670.370.560.560.270.860.470.860.46 Proposed Model with FSMAE0.140.100.110.100.090.090.110.140.110.070.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.100.540.560.250.280.250.110.590.150.770.15 Table 11Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained SAS-I using (i) only vegetation indices (VIs), (ii) a combination of spectral bands (SB) and VIs, and (iii) their feature-selected subset (FS), evaluated using deep learning baseline models.NutrientsBCaCuFeKMgMnNaPSZn EfficientNet with VIsMAE0.160.120.140.120.100.100.120.220.110.130.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.120.400.360.140.080.230.170.210.130.47−0.08 EfficientNet with SB+VIsMAE0.130.120.120.100.090.080.120.120.100.090.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.130.400.630.320.350.450.120.740.330.760.07 EfficientNet with FSMAE0.130.100.120.110.100.090.110.130.110.080.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.170.530.510.240.220.250.160.700.110.740.03 ResNet18 with VIsMAE0.160.130.150.120.100.100.120.220.110.130.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.040.350.300.130.120.260.160.220.140.420.05 ResNet18 with SB+VIsMAE0.120.100.110.090.090.080.110.100.110.070.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.280.540.570.300.320.360.170.780.170.770.14 ResNet18 with FSMAE0.140.110.120.110.090.090.110.160.110.090.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.130.510.530.180.210.250.200.560.150.690.13 Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained SAS-I using (i) only vegetation indices (VIs), (ii) a combination of spectral bands (SB) and VIs, and (iii) their feature-selected subset (FS), evaluated using deep learning baseline models. Table 12Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained SAS-I using (i) only vegetation indices (VIs), (ii) a combination of spectral bands (SB) and VIs, and (iii) their feature-selected subset (FS), evaluated using the proposed NAS-based model.NutrientsBCaCuFeKMgMnNaPSZn Proposed Model with VIsMAE0.150.130.140.120.100.100.120.210.110.130.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.020.400.350.120.140.220.190.150.130.370.04 Proposed Model with SB+VIsMAE0.110.090.100.080.080.070.100.080.090.060.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.660.670.370.560.560.270.860.470.860.46 Proposed Model with FSMAE0.140.100.110.100.090.090.110.140.110.070.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.100.540.560.250.280.250.110.590.150.770.15 Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained SAS-I using (i) only vegetation indices (VIs), (ii) a combination of spectral bands (SB) and VIs, and (iii) their feature-selected subset (FS), evaluated using the proposed NAS-based model. Ablation analysis of the proposed model The proposed NAS architecture consists of four key search space components, i.e., FAS, FWS, AAS, and LCS. To analyze the contribution of each component in the proposed NAS architecture, an ablation study was conducted by systematically removing each component from the complete architecture. The performance of each variant was evaluated using MAE, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} as shown in Table 13. The results indicate that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} provides a clearer distinction among different variants, as MAE exhibits relatively smaller variations across most configurations. The results demonstrate that the removal of AAS leads to the most significant degradation in performance among all components, with several macro/micro leaf nutrients showing negative or zero \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values. This degradation in performance without AAS suggests that adaptive activation plays an important role in capturing non-linear relationships among spectral features, which are essential for the precise estimation of macro/micro leaf nutrients. While removing FWS results in comparatively smaller performance drops, indicating its contribution is less dominant than that of other search spaces. Furthermore, the exclusion of LCS also leads to notable performance degradation, particularly for some nutrients such as B, K, Mn, and P, emphasizing the importance of layer connection in predicting leaf macro/micro nutrients through multispectral images. Similarly, removing FAS reduces the model’s ability to effectively extract complex spectral features, leading to moderate declines in performance of the NAS-based model. Overall, the results illustrate that each search space component contributes to the effectiveness of the proposed framework, with the complete model generally achieving improved performance across most nutrients.Table 13Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model and ablation models for SAS-I data.NutrientsBCaCuFeKMgMnNaPSZnProposed ModelMAE0.110.090.090.080.080.070.100.090.080.060.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.680.750.380.620.600.340.810.580.860.53w/o FASMAE0.110.090.100.090.080.080.100.090.090.070.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.630.630.380.510.500.270.790.380.810.38w/o FWSMAE0.110.090.090.080.080.070.090.090.090.060.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.670.680.430.570.500.240.830.460.860.48w/o LCSMAE0.120.100.110.090.090.080.100.080.100.070.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.200.510.560.310.270.350.100.850.110.800.20w/o AASMAE0.170.190.200.130.130.120.150.290.130.200.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.1−0.01−0.02−0.02−0.0100−0.08−0.02−0.110 Table 13Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model and ablation models for SAS-I data.NutrientsBCaCuFeKMgMnNaPSZnProposed ModelMAE0.110.090.090.080.080.070.100.090.080.060.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.680.750.380.620.600.340.810.580.860.53w/o FASMAE0.110.090.100.090.080.080.100.090.090.070.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.630.630.380.510.500.270.790.380.810.38w/o FWSMAE0.110.090.090.080.080.070.090.090.090.060.08\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.340.670.680.430.570.500.240.830.460.860.48w/o LCSMAE0.120.100.110.090.090.080.100.080.100.070.09\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}0.200.510.560.310.270.350.100.850.110.800.20w/o AASMAE0.170.190.200.130.130.120.150.290.130.200.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}−0.1−0.01−0.02−0.02−0.0100−0.08−0.02−0.110 Mean absolute errors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} errors obtained for the proposed model and ablation models for SAS-I data. Discussion The results demonstrate the effectiveness of the proposed NAS-based framework in estimating leaf macro/micro nutrients with improved accuracy compared to conventional machine learning techniques as well as predefined deep learning baseline techniques. Both growth stage-wise and treatment-wise analyses show that the model consistently maintains low mean absolute errors across multiple nutrients, treatments, and seasons, indicating reasonable robustness within the evaluated experimental setting. Although the proposed model demonstrates robust performance in the estimation of macro/micro nutrients across different growth stages, treatments, and seasons, its practical applicability also requires an understanding of the confidence associated with its predictions. To gain a deeper insight into whether the model is able to learn unique features in each case, we perform a series of experiments. These are further elaborated below:Table 14Confidence interval range (CI range) quantified using the width of the 95% confidence interval along with its relative uncertainty percentage(RU) and coefficient of variance (CV) across all nutrients.NutrientsBCaCuFeKMgMnNaPSZnCI range0.0620.0740.0810.0560.0610.0560.0490.1040.0370.0450.033RU\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 9.04\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 6.62\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 7.32\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 7.34\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 3.99\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 6.72\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 4.62\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 35.52\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 4.51\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 5.72\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 3.75\%$$\end{document}CV0.050.040.040.040.020.040.030.260.020.030.02MC Dropout based Visualization of Uncertainty in Nutrient prediction Uncertainty estimation is performed using Monte Carlo (MC) Dropout. For every input sample, 50 dropout-enabled stochastic forward passes are carried out. A Bayesian predictive interval that quantifies epistemic uncertainty is computed with 95% confidence intervals, estimated by the 2.5th and 97.5th percentiles of the prediction distribution. This enables us to observe the distribution of model predictions, thereby capturing epistemic uncertainty arising from model parameters. For each nutrient, the predictions obtained from multiple stochastic forward passes are aggregated to compute the predictive mean and variability. The CI range (upper–lower bound) represents the spread of predictions and is used as a measure of model confidence for each nutrient. Lower CI ranges indicate higher confidence in predictions, whereas larger ranges reflect increased uncertainty. Overall, the proposed model exhibits relatively low uncertainty for most nutrients, with CI ranges generally below 0.08, demonstrating stable and reliable predictions. Nutrients such as Mn, Mg, P, S, and Zn exhibit comparatively lower uncertainty, whereas Na shows relatively higher uncertainty, likely due to greater variability in its spectral response and field-level heterogeneity. This is evidently shown in Fig. 8, which presents the distribution of 95% CI ranges for all eleven nutrients. The width of each violin reflects the density of uncertainty values, while the central markers indicate the mean uncertainty.Fig. 8Monte Carlo Dropout-based confidence interval range across nutrients. Table 14Confidence interval range (CI range) quantified using the width of the 95% confidence interval along with its relative uncertainty percentage(RU) and coefficient of variance (CV) across all nutrients.NutrientsBCaCuFeKMgMnNaPSZnCI range0.0620.0740.0810.0560.0610.0560.0490.1040.0370.0450.033RU\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 9.04\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 6.62\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 7.32\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 7.34\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 3.99\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 6.72\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 4.62\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 35.52\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 4.51\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 5.72\%$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 3.75\%$$\end{document}CV0.050.040.040.040.020.040.030.260.020.030.02 Confidence interval range (CI range) quantified using the width of the 95% confidence interval along with its relative uncertainty percentage(RU) and coefficient of variance (CV) across all nutrients. MC Dropout based Visualization of Uncertainty in Nutrient prediction Uncertainty estimation is performed using Monte Carlo (MC) Dropout. For every input sample, 50 dropout-enabled stochastic forward passes are carried out. A Bayesian predictive interval that quantifies epistemic uncertainty is computed with 95% confidence intervals, estimated by the 2.5th and 97.5th percentiles of the prediction distribution. This enables us to observe the distribution of model predictions, thereby capturing epistemic uncertainty arising from model parameters. For each nutrient, the predictions obtained from multiple stochastic forward passes are aggregated to compute the predictive mean and variability. The CI range (upper–lower bound) represents the spread of predictions and is used as a measure of model confidence for each nutrient. Lower CI ranges indicate higher confidence in predictions, whereas larger ranges reflect increased uncertainty. Overall, the proposed model exhibits relatively low uncertainty for most nutrients, with CI ranges generally below 0.08, demonstrating stable and reliable predictions. Nutrients such as Mn, Mg, P, S, and Zn exhibit comparatively lower uncertainty, whereas Na shows relatively higher uncertainty, likely due to greater variability in its spectral response and field-level heterogeneity. This is evidently shown in Fig. 8, which presents the distribution of 95% CI ranges for all eleven nutrients. The width of each violin reflects the density of uncertainty values, while the central markers indicate the mean uncertainty. MC Dropout based Visualization of Uncertainty in Nutrient prediction Uncertainty estimation is performed using Monte Carlo (MC) Dropout. For every input sample, 50 dropout-enabled stochastic forward passes are carried out. A Bayesian predictive interval that quantifies epistemic uncertainty is computed with 95% confidence intervals, estimated by the 2.5th and 97.5th percentiles of the prediction distribution. This enables us to observe the distribution of model predictions, thereby capturing epistemic uncertainty arising from model parameters. For each nutrient, the predictions obtained from multiple stochastic forward passes are aggregated to compute the predictive mean and variability. The CI range (upper–lower bound) represents the spread of predictions and is used as a measure of model confidence for each nutrient. Lower CI ranges indicate higher confidence in predictions, whereas larger ranges reflect increased uncertainty. Overall, the proposed model exhibits relatively low uncertainty for most nutrients, with CI ranges generally below 0.08, demonstrating stable and reliable predictions. Nutrients such as Mn, Mg, P, S, and Zn exhibit comparatively lower uncertainty, whereas Na shows relatively higher uncertainty, likely due to greater variability in its spectral response and field-level heterogeneity. This is evidently shown in Fig. 8, which presents the distribution of 95% CI ranges for all eleven nutrients. The width of each violin reflects the density of uncertainty values, while the central markers indicate the mean uncertainty. Monte Carlo Dropout-based confidence interval range across nutrients. Additionally, the relative uncertainty (RU=CI range/mean prediction) was computed and expressed as a percentage to assess the reliability of the uncertainty estimates. Table 14 reports the average CI range, relative uncertainty, and coefficient of variation (CV) for each nutrient. The results indicate that nutrients such as K, Mn, P, and Zn exhibit low relative uncertainty (≤ ±5%), making them suitable for automated field mapping. Nutrients such as Ca, Cu, Fe, Mg, and S show moderate uncertainty (±5% to ±8%), indicating reliable predictions for agronomic decision-making. In contrast, higher uncertainty (> ±8%) is observed for B, suggesting that these predictions should be interpreted with caution. The alarmingly high RU for Na depicts that even though Na estimation achieves competitive predictive accuracy, the uncertainty analysis indicates lower confidence and stability in Na predictions relative to several other nutrients. Furthermore, the coefficient of variation is used to assess prediction stability. Nutrients with both low CV (<5%) and low RU (≤ ±5%), such as K, P, Zn, and Mn, demonstrate consistent and reliable predictions. In contrast, Na exhibits the highest variability, indicating unstable predictions across stochastic passes. Figure 9 further illustrates the predictive mean along with the associated 95% confidence interval for representative nutrients across test samples. The shaded region corresponds to the obtained uncertainty, where narrow confidence bands across most samples indicate stable predictions, while widening of the bands reflects higher uncertainty in regions with increased variability. Hence, incorporating uncertainty estimation enhances the practical feasibility of the proposed framework, as it enables evaluating the confidence associated with each estimate. Such predictions are particularly valuable for precision agriculture applications, where decision-making can benefit from both predicted nutrient levels and their associated reliability. Although certain nutrients exhibit comparatively higher uncertainty under the evaluated experimental conditions, the results demonstrate that the proposed NAS-based model not only achieves low estimation errors across several nutrients but also provides reliable uncertainty estimates.Fig. 9Monte Carlo Dropout based uncertainty obtained with 95% confidence interval for all nutrients.Feature Clusters w.r.t. Ground Truth Nutrient EstimatesData Preparation: A total of 80 samples were selected from the entire dataset, grouped into four sets of 20 each, ensuring maximum interclass variance in ground truth nutrient values between the groups.Hypothesis: It is expected that feature points from similar ground truth nutrient values will be close, whereas feature points from dissimilar nutrient ground truth values will be far from each other.Observation: The experimentation is carried out by extracting feature vectors of 10816 dimensions from the fifth layer from the end of the network. The 10 principal components are further used for generating the t-SNE plots. Fig. 10(a) and (c) clearly show the presence of four distinct clusters with respect to growth-stages and treatments, respectively. To quantitatively evaluate the separability of these clusters, centroids for each group were computed in the reduced feature space, and pairwise Euclidean distances were calculated. The corresponding heatmaps are shown in Fig. 10(b) and (d) for both growth stage-wise and treatment-wise experiments, respectively. The results from the treatment-wise distance matrix show that T1 exhibits the largest separation from the control (C), with a distance of 11.56. In contrast, T2 shows the smallest distance from C 3.42, suggesting similar characteristics, while T3 demonstrates moderate separation. Similarly, in the growth stage-wise analysis, GS1 and GS3 are relatively closer (3.40), whereas GS2 shows higher separation from GS1 (8.41), indicating variability across growth stages. These quantitative results provide stronger evidence of the discriminative capability of the extracted features.Fig. 10Visualization of 80 extracted features using t-SNE and inter-group distance matrices across different growth stages and treatment conditions. Monte Carlo Dropout based uncertainty obtained with 95% confidence interval for all nutrients. Feature Clusters w.r.t. Ground Truth Nutrient EstimatesData Preparation: A total of 80 samples were selected from the entire dataset, grouped into four sets of 20 each, ensuring maximum interclass variance in ground truth nutrient values between the groups.Hypothesis: It is expected that feature points from similar ground truth nutrient values will be close, whereas feature points from dissimilar nutrient ground truth values will be far from each other.Observation: The experimentation is carried out by extracting feature vectors of 10816 dimensions from the fifth layer from the end of the network. The 10 principal components are further used for generating the t-SNE plots. Fig. 10(a) and (c) clearly show the presence of four distinct clusters with respect to growth-stages and treatments, respectively. To quantitatively evaluate the separability of these clusters, centroids for each group were computed in the reduced feature space, and pairwise Euclidean distances were calculated. The corresponding heatmaps are shown in Fig. 10(b) and (d) for both growth stage-wise and treatment-wise experiments, respectively. The results from the treatment-wise distance matrix show that T1 exhibits the largest separation from the control (C), with a distance of 11.56. In contrast, T2 shows the smallest distance from C 3.42, suggesting similar characteristics, while T3 demonstrates moderate separation. Similarly, in the growth stage-wise analysis, GS1 and GS3 are relatively closer (3.40), whereas GS2 shows higher separation from GS1 (8.41), indicating variability across growth stages. These quantitative results provide stronger evidence of the discriminative capability of the extracted features. Feature Clusters w.r.t. Ground Truth Nutrient EstimatesData Preparation: A total of 80 samples were selected from the entire dataset, grouped into four sets of 20 each, ensuring maximum interclass variance in ground truth nutrient values between the groups.Hypothesis: It is expected that feature points from similar ground truth nutrient values will be close, whereas feature points from dissimilar nutrient ground truth values will be far from each other.Observation: The experimentation is carried out by extracting feature vectors of 10816 dimensions from the fifth layer from the end of the network. The 10 principal components are further used for generating the t-SNE plots. Fig. 10(a) and (c) clearly show the presence of four distinct clusters with respect to growth-stages and treatments, respectively. To quantitatively evaluate the separability of these clusters, centroids for each group were computed in the reduced feature space, and pairwise Euclidean distances were calculated. The corresponding heatmaps are shown in Fig. 10(b) and (d) for both growth stage-wise and treatment-wise experiments, respectively. The results from the treatment-wise distance matrix show that T1 exhibits the largest separation from the control (C), with a distance of 11.56. In contrast, T2 shows the smallest distance from C 3.42, suggesting similar characteristics, while T3 demonstrates moderate separation. Similarly, in the growth stage-wise analysis, GS1 and GS3 are relatively closer (3.40), whereas GS2 shows higher separation from GS1 (8.41), indicating variability across growth stages. These quantitative results provide stronger evidence of the discriminative capability of the extracted features. Data Preparation: A total of 80 samples were selected from the entire dataset, grouped into four sets of 20 each, ensuring maximum interclass variance in ground truth nutrient values between the groups. Data Preparation: A total of 80 samples were selected from the entire dataset, grouped into four sets of 20 each, ensuring maximum interclass variance in ground truth nutrient values between the groups. Hypothesis: It is expected that feature points from similar ground truth nutrient values will be close, whereas feature points from dissimilar nutrient ground truth values will be far from each other. Hypothesis: It is expected that feature points from similar ground truth nutrient values will be close, whereas feature points from dissimilar nutrient ground truth values will be far from each other. Observation: The experimentation is carried out by extracting feature vectors of 10816 dimensions from the fifth layer from the end of the network. The 10 principal components are further used for generating the t-SNE plots. Fig. 10(a) and (c) clearly show the presence of four distinct clusters with respect to growth-stages and treatments, respectively. To quantitatively evaluate the separability of these clusters, centroids for each group were computed in the reduced feature space, and pairwise Euclidean distances were calculated. The corresponding heatmaps are shown in Fig. 10(b) and (d) for both growth stage-wise and treatment-wise experiments, respectively. The results from the treatment-wise distance matrix show that T1 exhibits the largest separation from the control (C), with a distance of 11.56. In contrast, T2 shows the smallest distance from C 3.42, suggesting similar characteristics, while T3 demonstrates moderate separation. Similarly, in the growth stage-wise analysis, GS1 and GS3 are relatively closer (3.40), whereas GS2 shows higher separation from GS1 (8.41), indicating variability across growth stages. These quantitative results provide stronger evidence of the discriminative capability of the extracted features. Observation: The experimentation is carried out by extracting feature vectors of 10816 dimensions from the fifth layer from the end of the network. The 10 principal components are further used for generating the t-SNE plots. Fig. 10(a) and (c) clearly show the presence of four distinct clusters with respect to growth-stages and treatments, respectively. To quantitatively evaluate the separability of these clusters, centroids for each group were computed in the reduced feature space, and pairwise Euclidean distances were calculated. The corresponding heatmaps are shown in Fig. 10(b) and (d) for both growth stage-wise and treatment-wise experiments, respectively. The results from the treatment-wise distance matrix show that T1 exhibits the largest separation from the control (C), with a distance of 11.56. In contrast, T2 shows the smallest distance from C 3.42, suggesting similar characteristics, while T3 demonstrates moderate separation. Similarly, in the growth stage-wise analysis, GS1 and GS3 are relatively closer (3.40), whereas GS2 shows higher separation from GS1 (8.41), indicating variability across growth stages. These quantitative results provide stronger evidence of the discriminative capability of the extracted features. Visualization of 80 extracted features using t-SNE and inter-group distance matrices across different growth stages and treatment conditions. These observations highlight the model’s potential to learn meaningful patterns related to target variations. Future work should focus on extending the model’s validation across multiple seasons and study areas to further enhance its generalizability and accuracy. The successful performance of a deep learning model for leaf nutrient estimation from multispectral images can help farmers adopt the technique. However, translating these precise measurements into actionable insights for region-specific fertilizer application remains a critical area for further study and practical implementation. There is also a need to deploy drone-based crop monitoring facilities as an easily available service for large-scale adoption. Conclusion This study presents a comprehensive analysis of estimating eleven leaf macro/micro nutrients (B, Ca, Cu, Fe, K, Mg, Mn, Na, P, S, and Zn) using multispectral drone imagery and a Neural Architecture Search (NAS)-based framework with polynomial approximation functions. Comparative experiments were conducted using predefined deep learning baseline architectures as well as conventional machine learning techniques. Among the conventional machine learning models, XGB, RF, and MLP demonstrated competitive performance by outperforming DL baselines in some cases. Among deep learning models, DenseNet and EfficientNet consistently achieved better performance across both SAS-I and SAS-II. Although overall performance on SAS-II was slightly lower, achieving \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values ranging from 0.34 to 0.86 as compared to that of SAS-I, whose \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values are in the range of 0.14 to 0.71. Despite this, these models still outperformed other DL baselines, which exhibit lower \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values up to 0.15 for SAS-I and 0.09 for SAS-II with the best performing networks. Furthermore, the proposed NAS-based model surpasses both predefined DL architectures and traditional CML methods, achieving superior accuracy and demonstrating robust performance across different treatments and crop growth stages. A series of experiments was conducted, including growth stage-wise and treatment-wise analyses, season-specific training/testing, treatment-wise cross-validation, and an ablation study for the proposed network. In addition, a Monte Carlo dropout-based confidence interval analysis was performed to assess predictive uncertainty. The results demonstrate that the model effectively captures nutrient variations under diverse conditions. Furthermore, statistical analyses of channel-wise mean, standard deviation, coefficient of variation, and Gaussian distribution of pixel-level features confirmed that the model can differentiate learned features, exhibiting higher similarity within the same growth stage and treatment, and greater variability across different conditions. Overall, the proposed NAS-based approach provides precise and uncertainty-aware nutrient estimation, demonstrating strong potential for precision agriculture applications. We observe that intra-season testing loss on SAS-I ranging from 0.06–0.11 across different parameters previously, has now increased to 0.08–0.22. This indicates that a larger variation in the number of fields is required to further improve cross-season estimates. Such variations can alter spectral reflectance characteristics and influence nutrient estimation accuracy. Future work will focus on enhancing generalizability by incorporating data from additional study areas and seasons, further improving the model’s reliability for broader crop monitoring and nutrient management.