UAV-based temporal synergistic estimation of multiple alfalfa qualities integrating physics-informed network and 3D allometric operator Authors: Tao Cheng, Zhangru Gao, Weibo Ren, Ziyu Ding, Feng Yuan, Yaling Liu, Jiyuan Xie, Ruopu Zan, Dong Wang, Dongyan Zhang Published: 2026 ABSTRACT Accurately monitoring alfalfa nutritional quality is essential for optimal pasture management. Yet, current UAV remote sensing methods rely on single-temporal imagery and empirical indices, limiting their ability to handle multi-stage growth dynamics, canopy spectral saturation, and canopy-to-whole-plant scale differences. Furthermore, small sample sizes often cause purely data-driven models to overfit correlations, yielding biologically unrealistic results. Overcoming these challenges, we designed a comprehensive quality estimation framework using 127 alfalfa core germplasms, combining high-dimensional spectral mining, a physics-informed network, and a 3D allometric transfer operator. After screening 14,960 spectral operators across original and log-transformed spaces, we applied a dual dimensionality reduction strategy to isolate optimal features. Four-band dual-difference structures proved highly sensitive to fiber components (ADF/NDF, |r| = 0.896), while logarithmic decoupling operators accurately isolated protein and nitrogen signals (CP/N, |r| = 0.868). We then engineered a Physics-Informed Sparse Shallow Network (PI-SSN). By leveraging temporal attention decoupling, it adaptively assigns growth-stage weights to different components and uses carbon-nitrogen metabolic constraints to maintain biological accuracy during multi-task retrieval. Multi-stage temporal data significantly boosted accuracy over single-period spectra. PI-SSN delivered exceptional test set coefficients of determination (R2) of 0.812–0.848 and RPDs >2.0 for N, CP, ADF, and NDF, easily outperforming standard baselines. To bridge the canopy-only observation gap, we introduced a 3D allometric transfer operator that incorporates canopy coverage and plant height. This effectively corrected vertical stem-leaf observation biases, enhancing Relative Feed Value (RFV) predictions. Ultimately, this approach offers a powerful new framework for high-throughput forage phenotyping. Introduction Alfalfa (Medicago sativa L.) is valued for its high protein, digestibility, and yield, making it a key forage for global livestock production [1,2]. Alfalfa nutritional quality is commonly defined by critical metrics, including crude protein (CP), neutral detergent fiber (NDF), acid detergent fiber (ADF), and relative feed value (RFV). These parameters directly dictate the forage's economic value and are instrumental in determining its RFV and the optimal timing for harvest [3]. Consequently, the ability to rapidly, accurately, and extensively estimate alfalfa quality traits is critically important for both field-scale production and forage breeding efforts. Conventional assessment of forage quality predominantly depends on destructive field sampling coupled with laboratory chemical analyses, exemplified by the Kjeldahl method and near-infrared reflectance spectroscopy [4,5]. While these approaches are accurate, they are laborious, expensive, and inefficient. Critically, they cannot satisfy the high-throughput phenotyping demands necessary for large-scale germplasm evaluation and precision grassland management. Unmanned Aerial Vehicle (UAV) remote sensing with multispectral/hyperspectral sensors has become indispensable for crop monitoring, owing to its high spatial resolution and multi-temporal observation capacity. While UAV-based remote sensing is routinely employed in grain crops for assessing biomass, phenology, and quality traits [[6], [7], [8], [9], [10], [11], [12]], its application for monitoring the dynamic nutritional changes in perennial forages remains notably limited. Compared to grain crops, remote sensing research focusing on the dynamic changes in perennial forage quality remains relatively limited. Recently, preliminary efforts have been made to employ UAV remote sensing for forage phenotyping, including biomass estimation [13,14] and the prediction of quality attributes [15,16]. For instance, Fu et al. [17] successfully estimated a suite of alfalfa nutritional parameters (CP, ADF, NDF, RFV) by integrating UAV hyperspectral data with LASSO regression and conventional machine learning algorithms. Although these studies demonstrate the promising potential of UAV remote sensing for forage quality estimation, the expressive power of traditional spectral features in high-density canopies remains limited. The majority of remote sensing investigations into forage quality continue to depend on vegetation indices formulated from red and near-infrared wavelengths, such as Normalized Difference Vegetation Index (NDVI) and Soil Adjusted Vegetation Index (SAVI). However, Gitelson et al. [18] found that during the mid-to-late growth stages, near-infrared (NIR) reflectance tends to plateau or even decrease with increasing vegetation fraction (VF). This causes the traditional Normalized Difference Vegetation Index (NDVI) to enter a spectral saturation zone once VF exceeds 60%, significantly diminishing its ability to monitor changes in vegetation structure and biochemical parameters. To address this issue, Haboudane et al. [19] proposed using multi-band combinations, such as Modified Chlorophyll Absorption Ratio Index 2 (MCARI2), to mitigate the interference from soil background and variations in leaf area. Moreover, research by Fan et al. [20] and Marques et al. [21] has shown that employing machine learning models augmented with geometric features, in conjunction with logarithmic transformation of raw spectral data, can markedly enhance both the precision and robustness of predictions for key biophysical traits and the nitrogen balance index (NBI). Nevertheless, the prevailing approach in forage quality inversion remains centered on formulating empirical two-band indices within the original reflectance domain, with a notable absence of systematic investigation into log-transformed spaces and higher-order spectral feature combinations. Aiming to advance remote sensing retrieval capabilities, recent research has started to leverage machine learning, deep learning, and multi-task learning (MTL) frameworks for predicting crop quality traits [22]. Nonetheless, the prohibitive cost of field sampling and chemical analysis often means that remote sensing phenotyping datasets are characterized by inherently small sample sizes. Under such data-scarce conditions, complex deep neural networks, such as Convolutional Neural Networks (CNNs) or Transformers, frequently fail to exploit their full representational power and exhibit a high susceptibility to overfitting [23]. Conversely, conventional machine learning algorithms like Random Forest (RF) and Support Vector Machine (SVM), despite their robustness in single-time-point predictions, typically lack the capacity to adequately model the dynamic crop growth trajectories unfolding over multiple phenological phases [24,25]. For example, when Meng et al. [26] employed algorithms like Random Forest Regression (RFR) and Extreme Gradient Boosting (XGBoost) with multi-source UAV features to estimate maize biomass over the full growing season, they observed pronounced underestimation of the rapidly accumulating ear biomass during later stages (e.g., the dough stage). This limitation, attributed to spectral saturation, underscores the difficulty these models face in accurately representing the nonlinear dynamics of biomass accumulation across the entire phenological timeline. Critically, alfalfa's biochemical traits (such as rapidly cycling nitrogen) and structural traits (such as progressively accumulated cellulose) follow distinctly different temporal trajectories throughout plant development [27]. Conventional Single-Task Learning (STL) models address these indicators in isolation, consequently neglecting the inherent physiological trade-offs and carbon-nitrogen allocation strategies operating within the crop. Within the context of recent MTL frameworks, Feng et al. [28] developed a Long Short-Term Memory (LSTM)-based Multi-Task Learning model specifically designed to predict key alfalfa nutritional indices, namely crude protein, acid detergent fiber, and neutral detergent fiber. While this model achieved predictive accuracy surpassing that of other conventional machine learning approaches, the absence of integrated physiological constraints left it susceptible to generating predictions that could contravene established biological principles. Predominantly, current studies construct forage quality models utilizing single-temporal imagery. This approach systematically overlooks the continuous physiological dynamics inherent to crop growth, making it inadequate to fully capture the temporal evolution of quality formation. For instance, advancing plant maturity is typically accompanied by increasing fiber concentration and a concurrent decline in protein content [29]. Consequently, relying exclusively on single-temporal remote sensing data is often insufficient for accurately capturing the genuine temporal dynamics of quality variation. Moreover, constrained by the inherent penetration limits of optical remote sensing, the spectral signal captured by UAVs is primarily governed by a combination of factors, including leaf architecture, leaf area index, stem-to-leaf ratio, and the complex multiple scattering of light within the canopy [8,30]. In contrast, forage quality assessed for practical purposes is conventionally determined from whole-plant samples, encompassing stems and lower, often senescent, leaves. Allometric growth theory posits that increasing plant height drives a greater allocation of structural carbohydrates to basal tissues, a strategy essential for maintaining mechanical stability [31,32]. Consequently, alfalfa development is characteristically marked by a physiological dilution of protein and a concurrent structural accumulation of fiber [29]. Directly equating spectral information derived from the canopy surface with whole-plant quality would therefore result in a systematic overestimation of protein concentration and a commensurate underestimation of fiber concentration. To address these challenges, this study leverages 127 alfalfa core germplasms, synergizing multi-temporal UAV multispectral remote sensing with extensive ground-based measurements. Our objective is to develop a comprehensive remote sensing framework for quality assessment that integrates advanced spectral feature mining, physics-informed modeling, and canopy-scale correction. This framework is specifically designed to resolve the critical issues of multi-quality collaborative inversion with limited samples and the scale discrepancy between canopy-level observations and whole-plant characteristics. The principal contributions of this work are threefold.(1)A library of low-redundancy, highly sensitive spectral operators tailored for multiple alfalfa quality metrics was systematically developed and optimized by integrating both original and log-transformed reflectance spaces. This library effectively mitigates the spectral saturation commonly encountered by traditional indices in dense canopies.(2)A novel Physics-Informed Sparse Shallow Network (PI-SSN) is introduced. By incorporating temporal attention decoupling and biologically grounded trade-off constraints, the PI-SSN circumvents the correlation overfitting typical of purely data-driven models trained on small datasets, thereby achieving synergistic and robust predictions across multiple quality traits.(3)Guided by allometric theory, a “canopy-to-whole-plant” scale transfer model was developed by integrating UAV-derived canopy coverage and plant height data. This model effectively corrects the systematic bias associated with surface-level spectral observations, enabling, for the first time, a reliable estimation of whole-plant Relative Feed Value (RFV) from canopy-level quality assessments. (1)A library of low-redundancy, highly sensitive spectral operators tailored for multiple alfalfa quality metrics was systematically developed and optimized by integrating both original and log-transformed reflectance spaces. This library effectively mitigates the spectral saturation commonly encountered by traditional indices in dense canopies. A library of low-redundancy, highly sensitive spectral operators tailored for multiple alfalfa quality metrics was systematically developed and optimized by integrating both original and log-transformed reflectance spaces. This library effectively mitigates the spectral saturation commonly encountered by traditional indices in dense canopies. (2)A novel Physics-Informed Sparse Shallow Network (PI-SSN) is introduced. By incorporating temporal attention decoupling and biologically grounded trade-off constraints, the PI-SSN circumvents the correlation overfitting typical of purely data-driven models trained on small datasets, thereby achieving synergistic and robust predictions across multiple quality traits. A novel Physics-Informed Sparse Shallow Network (PI-SSN) is introduced. By incorporating temporal attention decoupling and biologically grounded trade-off constraints, the PI-SSN circumvents the correlation overfitting typical of purely data-driven models trained on small datasets, thereby achieving synergistic and robust predictions across multiple quality traits. (3)Guided by allometric theory, a “canopy-to-whole-plant” scale transfer model was developed by integrating UAV-derived canopy coverage and plant height data. This model effectively corrects the systematic bias associated with surface-level spectral observations, enabling, for the first time, a reliable estimation of whole-plant Relative Feed Value (RFV) from canopy-level quality assessments. Guided by allometric theory, a “canopy-to-whole-plant” scale transfer model was developed by integrating UAV-derived canopy coverage and plant height data. This model effectively corrects the systematic bias associated with surface-level spectral observations, enabling, for the first time, a reliable estimation of whole-plant Relative Feed Value (RFV) from canopy-level quality assessments. This research is intended to offer a novel perspective for high-throughput remote sensing of perennial forage quality, while simultaneously providing critical technological underpinnings for the identification of superior germplasm resources and the advancement of precision management practices in smart grassland agriculture. Materials and methods Study area and experimental design The field experiment was conducted in Siziwang Banner, Ulanqab City, Inner Mongolia Autonomous Region, situated at coordinates 111°53′-113°58′ E and 41°10′-43°22′ N. The experimental area features typical chestnut soil, known for its stable physicochemical properties, moderate organic matter reserves, and superior capacity for water and nutrient conservation. These characteristics create an optimal growing environment for deep-rooted perennial forages like alfalfa. A visual overview of the study area is presented in Fig. S1. To comprehensively represent the genetic diversity and phenotypic variation inherent in alfalfa, a population consisting of 127 global core germplasm accessions (Medicago sativa L.) was established. The germplasm nursery was arranged using a randomized complete block design. Field transplantation of the alfalfa materials was uniformly carried out in July 2024. Each germplasm accession was assigned to an individual plot measuring 9.2 m2 (4 m × 2.3 m). Within each plot, plants were established at a uniform spacing of 0.6 m × 0.6 m in a grid pattern, ensuring exactly 15 plants per plot. Uniform agronomic management, including irrigation and manual weeding, was maintained across the entire experimental area throughout the observation period. After experiencing low-temperature freezing injury during the winter of 2024-2025 and cutting stress during the 2025 growing season, some less tolerant varieties were naturally eliminated (resulting in individual plant mortality). UAV data processing and ground sample measurements UAV remote sensing data acquisition and processing To simultaneously capture multispectral reflectance and three-dimensional structural information throughout alfalfa development, a dual-UAV configuration was employed. A DJI Matrice 350 RTK (M350) equipped with an AMS-10 multispectral imager was deployed to acquire high-fidelity spectral data, serving as the foundation for subsequent quality trait inversion modeling. Concurrently, a DJI Mavic 3 Multispectral (M3M) captured high-resolution true-color and base multispectral imagery, which were utilized for 3D reconstruction and the derivation of key canopy structural parameters, specifically plant height (PH) and canopy coverage (CC). The flight campaigns were strategically scheduled to encompass the critical growth stages within two complete alfalfa cutting cycles. The specific acquisition dates were: June 8, June 27, and July 7, 2025, for the first cutting; and July 21, July 30, and August 11, 2025, for the second cutting. A uniform stubble height of 5 cm was retained following each cutting event. All flight operations were conducted between 10:00 a.m. and 2:00 p.m. under clear, cloudless skies and calm wind conditions to ensure optimal illumination. The relative flight altitude was maintained at 15 m, with a forward overlap and side overlap of 75% each. To enable precise geometric and radiometric correction of the acquired imagery, ten ground control points (50 cm × 50 cm) and a white radiometric calibration panel (1 m × 1 m) were strategically placed within the experimental area (Fig. S1).(1)Spectral data acquisition and processing. High-resolution multispectral data were captured using the AMS-10, a 10-channel spectral imager (Azure Aviation Technology Ltd., China) mounted on the M350 platform. Key specifications of this sensor are summarized in Table S1. The acquired raw images were processed using Agisoft Metashape (Agisoft LLC, St. Petersburg, Russia). A workflow involving image alignment, dense point cloud reconstruction, and geometric distortion correction was applied to generate Digital Orthophoto Maps (DOM) and Digital Surface Models (DSM). Following this, the temporal spectral reflectance for the 127 alfalfa accessions was extracted within ENVI 5.3 (Harris Geospatial Solutions, Broomfield, CO, USA) using defined Regions of Interest (ROI). These extracted reflectance values constituted the primary input for the subsequent quality estimation models. Notably, to eliminate soil and shadow interference while accounting for dynamic canopy structural variations across growth stages, no fixed NDVI threshold was applied. Instead, an interactive histogram-based dynamic thresholding method, combined with visual inspection, was employed to determine the optimal segmentation threshold for pure canopy pixels for each flight. Subsequently, during the manual delineation of Regions of Interest (ROIs) in ENVI, plot boundaries and any residual mixed or non-pure alfalfa pixels were carefully excluded to ensure that the extracted reflectance strictly represents the core canopy.(2)Structural parameter extraction. The 20-megapixel high-resolution true-color and multispectral imagery from the M3M were processed in Pix4Dmapper (Pix4D SA, Lausanne, Switzerland) to perform 3D reconstruction, yielding DSM and DOM products. For each acquisition date, a Canopy Height Model (CHM) was derived by differencing the DSM with a corresponding Digital Elevation Model (DEM). Plant height (PH) was then extracted from these CHM datasets. The reliability of the extracted heights was effectively confirmed using ground-measured data (Fig. S2). Simultaneously, effective canopy coverage (CC) for each plot was quantified through a synergistic approach that integrated high-resolution true-color imagery with the Normalized Difference Vegetation Index (NDVI). (1)Spectral data acquisition and processing. High-resolution multispectral data were captured using the AMS-10, a 10-channel spectral imager (Azure Aviation Technology Ltd., China) mounted on the M350 platform. Key specifications of this sensor are summarized in Table S1. The acquired raw images were processed using Agisoft Metashape (Agisoft LLC, St. Petersburg, Russia). A workflow involving image alignment, dense point cloud reconstruction, and geometric distortion correction was applied to generate Digital Orthophoto Maps (DOM) and Digital Surface Models (DSM). Following this, the temporal spectral reflectance for the 127 alfalfa accessions was extracted within ENVI 5.3 (Harris Geospatial Solutions, Broomfield, CO, USA) using defined Regions of Interest (ROI). These extracted reflectance values constituted the primary input for the subsequent quality estimation models. Notably, to eliminate soil and shadow interference while accounting for dynamic canopy structural variations across growth stages, no fixed NDVI threshold was applied. Instead, an interactive histogram-based dynamic thresholding method, combined with visual inspection, was employed to determine the optimal segmentation threshold for pure canopy pixels for each flight. Subsequently, during the manual delineation of Regions of Interest (ROIs) in ENVI, plot boundaries and any residual mixed or non-pure alfalfa pixels were carefully excluded to ensure that the extracted reflectance strictly represents the core canopy. Spectral data acquisition and processing. High-resolution multispectral data were captured using the AMS-10, a 10-channel spectral imager (Azure Aviation Technology Ltd., China) mounted on the M350 platform. Key specifications of this sensor are summarized in Table S1. The acquired raw images were processed using Agisoft Metashape (Agisoft LLC, St. Petersburg, Russia). A workflow involving image alignment, dense point cloud reconstruction, and geometric distortion correction was applied to generate Digital Orthophoto Maps (DOM) and Digital Surface Models (DSM). Following this, the temporal spectral reflectance for the 127 alfalfa accessions was extracted within ENVI 5.3 (Harris Geospatial Solutions, Broomfield, CO, USA) using defined Regions of Interest (ROI). These extracted reflectance values constituted the primary input for the subsequent quality estimation models. Notably, to eliminate soil and shadow interference while accounting for dynamic canopy structural variations across growth stages, no fixed NDVI threshold was applied. Instead, an interactive histogram-based dynamic thresholding method, combined with visual inspection, was employed to determine the optimal segmentation threshold for pure canopy pixels for each flight. Subsequently, during the manual delineation of Regions of Interest (ROIs) in ENVI, plot boundaries and any residual mixed or non-pure alfalfa pixels were carefully excluded to ensure that the extracted reflectance strictly represents the core canopy. (2)Structural parameter extraction. The 20-megapixel high-resolution true-color and multispectral imagery from the M3M were processed in Pix4Dmapper (Pix4D SA, Lausanne, Switzerland) to perform 3D reconstruction, yielding DSM and DOM products. For each acquisition date, a Canopy Height Model (CHM) was derived by differencing the DSM with a corresponding Digital Elevation Model (DEM). Plant height (PH) was then extracted from these CHM datasets. The reliability of the extracted heights was effectively confirmed using ground-measured data (Fig. S2). Simultaneously, effective canopy coverage (CC) for each plot was quantified through a synergistic approach that integrated high-resolution true-color imagery with the Normalized Difference Vegetation Index (NDVI). Structural parameter extraction. The 20-megapixel high-resolution true-color and multispectral imagery from the M3M were processed in Pix4Dmapper (Pix4D SA, Lausanne, Switzerland) to perform 3D reconstruction, yielding DSM and DOM products. For each acquisition date, a Canopy Height Model (CHM) was derived by differencing the DSM with a corresponding Digital Elevation Model (DEM). Plant height (PH) was then extracted from these CHM datasets. The reliability of the extracted heights was effectively confirmed using ground-measured data (Fig. S2). Simultaneously, effective canopy coverage (CC) for each plot was quantified through a synergistic approach that integrated high-resolution true-color imagery with the Normalized Difference Vegetation Index (NDVI). Ground sample measurements Destructive sampling was performed in coordination with UAV flight campaigns, specifically at the first cutting (July 7) and second cutting (August 11) stages. For each cutting cycle, canopy samples were obtained from 92 representative germplasm accessions, with three individual plants sampled per accession, yielding a total of 184 samples. Notably, conventional alfalfa breeding and forage production assessments are fundamentally based on whole-plant sampling protocols. Under standard production conditions, whole-plant alfalfa typically exhibits CP concentrations between 15% and 22%, and NDF concentrations between 35% and 50%. In contrast, the spectral signal captured by nadir-viewing UAVs predominantly reflects the alfalfa canopy, characterized by a high proportion of young leaves and newly emerged apical stems. The probability density distributions of measured data (Fig. S3) clearly reveal that canopy-scale CP values are predominantly concentrated in a higher range (22%-30%), substantially exceeding typical whole-plant levels, whereas fiber concentrations exhibit a correspondingly lower distribution. To further elucidate the vertical quality evolution from canopy to whole-plant, a critical “canopy-to-whole-plant” homologous paired dataset was established. Specifically, 50 representative core accessions were randomly selected from the 92 materials sampled during the second cutting. Whole-plant samples for these accessions were concurrently collected using standard field cutting practices with a uniform 5 cm stubble height. Upon collection, all fresh forage samples were oven-dried at 105 °C for 30 min for enzyme deactivation, followed by drying at 65 °C until a constant weight was achieved. Following grinding and sieving, total nitrogen (N) content was quantified using the Kjeldahl method and subsequently converted to crude protein (CP) using a conversion factor of 6.25. Neutral detergent fiber (NDF) and acid detergent fiber (ADF) were analyzed following the Van Soest procedure. The distribution patterns depicted in Fig. S3 reveal that, driven by variations in accumulated temperature and precipitation across the growing season, the four alfalfa nutritional parameters exhibit not only significant temporal heterogeneity between the first and second cutting canopies, but also pronounced vertical heterogeneity when comparing canopy-level versus whole-plant measurements within the second cutting. Construction of spectral feature space and optimization of high-dimensional vegetation indices UAV multispectral imagery is characterized by pronounced structural heterogeneity at the canopy level. The confounding effects of reflectance bias from exposed soil backgrounds and spectral saturation in dense alfalfa canopies significantly limit the retrieval capability of conventional, empirically derived vegetation indices. To overcome these limitations, this study moves beyond reliance on individual empirical indices. Guided by classical vegetation optical transfer theory as a prior constraint, a comprehensive feature candidate library was constructed, encompassing spectral log-transformations and 1- to 4-dimensional band combinations. A brute-force search algorithm was subsequently implemented to identify phenotypic fingerprints characterized by robust stability across different growth stages. Mathematical structure of spectral transformation and high-dimensional features According to the Beer-Lambert law, the absorption characteristics of a plant canopy are strongly linearly correlated with the logarithmic reciprocal of reflectance [33]. Accordingly, a logarithmic transformation log(1/R) was initially applied to the original reflectance (OR) data. This transformation serves to mitigate nonlinear scattering effects while accentuating the net absorption cross-section of key biochemical constituents. Leveraging the 10 OR bands and their 10 log(1/R) counterparts, a comprehensive candidate feature library was subsequently constructed across four physical dimensions, yielding a total of 14,960 features (7480 from each spectral domain). The mathematical formulations for each feature type are detailed below:(1)Single-band features (n = 20): For each spectral band ranging from 405 nm to 850 nm, both the original reflectance Ri and its log-transformed value log(1/Ri) were directly employed. These features characterize the canopy's intrinsic radiative and absorptive responses at specific wavelengths.(2)Two-band combinations (n = 540): Formulated to capture the contrast between spectral bands, this category encompasses the Normalized Difference Spectral Index (NDSI), Ratio Spectral Index (RSI), and Difference Spectral Index (DSI). Notably, when applied to log-transformed spectra, the NDSI formulation becomes mathematically equivalent to the Normalized Difference Lignin Index (NDLI). Although DSI directly quantifies the absolute energy difference between bands, it is prone to multiplicative noise induced by variations in solar elevation angle and topography in UAV-based remote sensing. Conversely, RSI and NDSI/NDLI achieve robust radiometric normalization through their ratio-based formulations.(1)NDSI(i,j)=Ri−RjRi+Rj(2)RSI(i,j)=RiRj(3)DSI(i,j)=Ri−Rj(3)Three-band decoupling and baseline correction indices (n = 4320): Given the extremely dense canopy during alfalfa's mid-to-late growth stages and the strong soil background reflectance characteristic of arid regions, conventional two-band indices often suffer from saturation and background contamination. To mitigate these issues, a third reference band Rk was incorporated, leading to the formulation of three physically derived indices:1)Visible-band anti-noise structure (3 B-VARI):(4)3B-VARI(i,j,k)=Ri−RjRi+Rj−Rk (1)Single-band features (n = 20): For each spectral band ranging from 405 nm to 850 nm, both the original reflectance Ri and its log-transformed value log(1/Ri) were directly employed. These features characterize the canopy's intrinsic radiative and absorptive responses at specific wavelengths. Single-band features (n = 20): For each spectral band ranging from 405 nm to 850 nm, both the original reflectance Ri and its log-transformed value log(1/Ri) were directly employed. These features characterize the canopy's intrinsic radiative and absorptive responses at specific wavelengths. (2)Two-band combinations (n = 540): Formulated to capture the contrast between spectral bands, this category encompasses the Normalized Difference Spectral Index (NDSI), Ratio Spectral Index (RSI), and Difference Spectral Index (DSI). Notably, when applied to log-transformed spectra, the NDSI formulation becomes mathematically equivalent to the Normalized Difference Lignin Index (NDLI). Although DSI directly quantifies the absolute energy difference between bands, it is prone to multiplicative noise induced by variations in solar elevation angle and topography in UAV-based remote sensing. Conversely, RSI and NDSI/NDLI achieve robust radiometric normalization through their ratio-based formulations. Two-band combinations (n = 540): Formulated to capture the contrast between spectral bands, this category encompasses the Normalized Difference Spectral Index (NDSI), Ratio Spectral Index (RSI), and Difference Spectral Index (DSI). Notably, when applied to log-transformed spectra, the NDSI formulation becomes mathematically equivalent to the Normalized Difference Lignin Index (NDLI). Although DSI directly quantifies the absolute energy difference between bands, it is prone to multiplicative noise induced by variations in solar elevation angle and topography in UAV-based remote sensing. Conversely, RSI and NDSI/NDLI achieve robust radiometric normalization through their ratio-based formulations. (3)Three-band decoupling and baseline correction indices (n = 4320): Given the extremely dense canopy during alfalfa's mid-to-late growth stages and the strong soil background reflectance characteristic of arid regions, conventional two-band indices often suffer from saturation and background contamination. To mitigate these issues, a third reference band Rk was incorporated, leading to the formulation of three physically derived indices:1)Visible-band anti-noise structure (3 B-VARI): Three-band decoupling and baseline correction indices (n = 4320): Given the extremely dense canopy during alfalfa's mid-to-late growth stages and the strong soil background reflectance characteristic of arid regions, conventional two-band indices often suffer from saturation and background contamination. To mitigate these issues, a third reference band Rk was incorporated, leading to the formulation of three physically derived indices:1)Visible-band anti-noise structure (3 B-VARI): 1)Visible-band anti-noise structure (3 B-VARI): Visible-band anti-noise structure (3 B-VARI): Derived from the Visible Atmospherically Resistant Index (VARI) introduced by Gitelson et al. [18], this formulation subtracts a background-sensitive band Rk (e.g., the blue-violet band at 405 nm) in the denominator. This operation effectively strips away additive noise originating from soil background and atmospheric scattering.2)Difference ratio structure (3 B-MTCI):(5)3B-MTCI(i,j,k)=Ri−RjRj−Rk 2)Difference ratio structure (3 B-MTCI): Difference ratio structure (3 B-MTCI): By computing the relative reflectance slope between adjacent spectral bands, this index sensitively detects red-edge shifts, either blue shifts or red shifts, induced by factors such as increased lignification or nitrogen depletion. Furthermore, it effectively mitigates index saturation in canopies with high fractional cover [34].3)Absorption-scattering decoupling structure (3 B-Pigment):(6)3B-Pigment(i,j,k)=(1Ri−1Rj)×Rk 3)Absorption-scattering decoupling structure (3 B-Pigment): Absorption-scattering decoupling structure (3 B-Pigment): Grounded in spectral absorption depth theory, this formulation employs (1/Ri−1/Rj) to isolate the characteristic absorption features of biochemical constituents. Concurrently, a strongly scattering band Rk is incorporated to correct for errors arising from variations in canopy geometric structure, lending strong physiological validity to the retrieval of nitrogen and protein concentrations [35,36].(4)Four-band dual-difference normalized index (n = 10,080): Designed to achieve dual-decoupling of biochemical concentration signals from canopy structural signals within complex canopies, a higher-order dual-difference structure was formulated:(7)4B-DD(i,j,k,l)=Ri−RjRk−Rl (4)Four-band dual-difference normalized index (n = 10,080): Designed to achieve dual-decoupling of biochemical concentration signals from canopy structural signals within complex canopies, a higher-order dual-difference structure was formulated: Four-band dual-difference normalized index (n = 10,080): Designed to achieve dual-decoupling of biochemical concentration signals from canopy structural signals within complex canopies, a higher-order dual-difference structure was formulated: By leveraging a denominator difference pair, such as the structural-sensitive baseline formed by near-infrared and red bands, this operator adaptively normalizes the subtle quality degradation signals captured by the numerator difference pair. These signals frequently encompass features related to photosynthetic decline or fiber accumulation in the visible spectrum [37]. This formulation physically decouples the obscuring and confounding effects of canopy density on internal biochemical reflectance signals, thereby enabling an independent estimation of true nutrient concentrations in alfalfa. Full-space feature optimization and dual-threshold redundancy reduction The algorithm systematically traversed the 14,960 candidate formulations, implementing a dual-pronged dimensionality reduction strategy that integrated target correlation screening with cross-collinearity elimination. This rigorous approach was intended to fully mitigate multicollinearity and yield an orthogonal set of input features. Initially, feature validity screening was conducted using the temporally aggregated dataset encompassing both the first (Cut 1) and second (Cut 2) cuttings. For each candidate index, Pearson correlation coefficients (r) were computed against the target traits: total nitrogen (N), crude protein (CP), ADF, and NDF. A stringent inclusion threshold was established at |r|≥0.85, equating to an explained variance R2>0.72. Subsequently, feature independence was assessed to eliminate redundancy. To circumvent the curse of dimensionality that can arise when feeding high-dimensional spectral features into machine learning models, a sequential elimination strategy prioritizing correlation strength was employed. Specifically, features retained from the initial screening were ranked in descending order based on the absolute value of their predictive correlation and subsequently processed in that sequence. In multivariate statistical theory, a pairwise correlation coefficient exceeding 0.90 among predictor variables is generally indicative of severe multicollinearity, corresponding to a variance inflation factor VIF≥5 [38]. Consequently, if any feature in the sorted list was found to have a cross-correlation coefficient exceeding 0.90 (|rcross|>0.90) with any feature already present in the retained pool, it was flagged as highly collinear and redundant, and thus excluded. Following this two-stage process of correlation screening and collinearity elimination, each quality trait was associated with an optimal set of spectral features exhibiting minimal redundancy. These derived two-dimensional canopy spectral features will be integrated with multi-temporal data to form the core input variable set. This set will be utilized for the subsequent development of Single-Task Learning (STL) and Multi-Task Learning (MTL) baseline models, and for rigorous performance evaluation of the Physics-Informed Sparse Shallow Network (PI-SSN) proposed herein. Architecture of the physics-informed sparse shallow network (PI-SSN) In the fields of agricultural remote sensing and phenomics, one of the foremost challenges for machine learning applications is to achieve accurate inversion of crop growth status across multiple time points and traits using inherently limited sample datasets. Conventional deep networks are particularly susceptible to overfitting training set noise when sample sizes are limited [39]. Moreover, purely data-driven models, which operate without incorporating agronomic knowledge, can yield predictions that contravene established biological principles [40]. To overcome these limitations, we developed a Physics-Informed Sparse Shallow Network (PI-SSN), a framework designed for broad applicability. It effectively integrates spectral feature extraction with agronomic prior knowledge by synergistically employing three key mechanisms: temporal attention decoupling, a sparse shallow network topology, and biology-constrained loss functions (Fig. 1). This framework is demonstrated herein through its application to the multi-temporal inversion of four key nutritional quality parameters in alfalfa.Fig. 1Schematic diagram of the PI-SSN architecture. Schematic diagram of the PI-SSN architecture. Trait-driven decoupled temporal attention Diverse phenotypic traits in crops follow distinctly heterogeneous temporal trajectories. As a general rule, biochemical metabolites such as chlorophyll and total nitrogen exhibit rapid fluctuations and are more strongly correlated with recent observations [41]. Conversely, structural components including cellulose and biomass reflect a cumulative process over time, thereby displaying greater dependence on historical growth states [42]. Employing a shared temporal feature extraction mechanism in a multi-task inversion framework can obscure the unique temporal lag effects inherent to specific traits. To overcome this limitation, PI-SSN incorporates a parallel multi-stream temporal attention module, specifically designed to capture the distinct growth dynamics associated with different trait categories. Formally, the input multi-temporal feature matrix is defined as Xtime=[X1,X2,…,Xk]∈Rk×d, where k denotes the number of observation time points and d is the spectral feature dimension. For the t-th observation time point, the attention score vector et, temporal attention weight αt, and the resulting fused context vector Cm are computed according to the following equations:(8)et=tanh(WaXt+ba)(9)αt=exp(weTet)∑j=1kexp(weTej)(10)Cm=∑t=1kαt(m)Xt Here, et denotes the feedforward intermediate representation, also termed the attention score vector, for the t-th time point, derived via a feedforward operation with tanh activation, without any inherent state-updating mechanism. Wa and ba are the learnable parameters (weight matrix and bias) of the attention mechanism, while Xt corresponds to the feature vector at the t-th time step within Xtime. we is the projection vector that maps et to a scalar attention score. J serves as a summation index, iterating over all elements within the set m. αt(m) is the normalized attention weight for the t-th time step within the trait-specific stream m. m indexes the trait stream category, thereby establishing the context for the normalization procedure. In the present study, this is partitioned into the biochemical group (m=bio) and the structural group (m=struc). Cm is derived as the time-step-weighted sum of the input Xt. By employing this decoupled attention mechanism, the model can adaptively allocate distinct retrospective weights to components exhibiting different temporal dynamics, thereby extracting context features uniquely relevant to each trait category. This architecture is highly scalable, making it readily applicable to plant phenotyping tasks that encompass arbitrary numbers of observation time points and trait groupings. Dual-stream shallow topology with L1 sparsity To mitigate the risk of compromised generalization performance due to excessive model parameters under small-sample conditions, PI-SSN departs from complex, deeply stacked architectures in favor of a lightweight feature extraction framework. Guided by the distinct spatial distribution characteristics of the target variables, the model implements two parallel inference streams.(1)Smooth Inference Stream. Designed for traits characterized by continuous probability distributions in agricultural fields, specifically N and CP, this stream utilizes a single hidden layer with a Tanh activation function. The objective is to capture the overall variation trends of biochemical concentrations with parsimonious model capacity.(2)Sparse Feature Stream. Targeting traits highly sensitive to canopy three-dimensional structure and genotypic variation, which thereby exhibit pronounced nonlinear spatial heterogeneity as exemplified by ADF and NDF, this stream incorporates a LeakyReLU activation function. Furthermore, to equip the network with feature selection capabilities analogous to those of decision tree algorithms, an L1 sparsity regularization, widely known as the Lasso penalty, is imposed on the weight matrix Wstruc of its fully connected layer:(11)LL1=∑p⁡|Wstruc,p| (1)Smooth Inference Stream. Designed for traits characterized by continuous probability distributions in agricultural fields, specifically N and CP, this stream utilizes a single hidden layer with a Tanh activation function. The objective is to capture the overall variation trends of biochemical concentrations with parsimonious model capacity. Smooth Inference Stream. Designed for traits characterized by continuous probability distributions in agricultural fields, specifically N and CP, this stream utilizes a single hidden layer with a Tanh activation function. The objective is to capture the overall variation trends of biochemical concentrations with parsimonious model capacity. (2)Sparse Feature Stream. Targeting traits highly sensitive to canopy three-dimensional structure and genotypic variation, which thereby exhibit pronounced nonlinear spatial heterogeneity as exemplified by ADF and NDF, this stream incorporates a LeakyReLU activation function. Furthermore, to equip the network with feature selection capabilities analogous to those of decision tree algorithms, an L1 sparsity regularization, widely known as the Lasso penalty, is imposed on the weight matrix Wstruc of its fully connected layer: Sparse Feature Stream. Targeting traits highly sensitive to canopy three-dimensional structure and genotypic variation, which thereby exhibit pronounced nonlinear spatial heterogeneity as exemplified by ADF and NDF, this stream incorporates a LeakyReLU activation function. Furthermore, to equip the network with feature selection capabilities analogous to those of decision tree algorithms, an L1 sparsity regularization, widely known as the Lasso penalty, is imposed on the weight matrix Wstruc of its fully connected layer: Here, Wstruc,p denotes the p-th weight parameter within the fully connected layer of the sparse feature stream. During the optimization process, this regularization term compels the network to aggressively suppress the weights of redundant features exhibiting weak correlation with the target traits, effectively driving these weights toward zero. This mechanism enables automated feature selection and denoising at the foundational level, all while preserving a parsimonious network architecture. Biology-constrained joint loss function Resource competition and trade-offs are ubiquitous phenomena throughout crop growth and development [43]. Drawing upon the established theory of carbon-nitrogen metabolic equilibrium in plants, a fundamental trade-off governs resource allocation, wherein the synthesis of structural carbohydrates such as cellulose and nitrogenous compounds such as proteins constitute inherently competitive processes when crops allocate photosynthates. [44]. For alfalfa under standard field cultivation, the progression from vegetative to reproductive growth stages is characterized by an increasing stem-to-leaf ratio. This developmental shift is typically associated with a concomitant dilution of crude protein concentration and accumulation of fiber components [45]. PI-SSN integrates this fundamental agronomic principle as a soft constraint within the model, leading to a reformulated total loss Ltotal function designed for the joint optimization of multiple target traits:(12)Ltotal=Lreg+λphy·Ψphy+λL1·LL1Lreg denotes the empirical regression error, Ψphy represents the physical correlation constraint loss, and LL1 is the network weight sparsity loss. The terms λphy and λL1 are the respective penalty coefficients controlling the contribution of each constraint. First, the heterogeneous regression loss (Lreg) is formulated using distinct loss functions tailored to the error distribution characteristics of each data type:(13)Lreg=∑i∈Ωbio⁡MSE(yˆi,yi)+∑j∈Ωstruc⁡γ·SmoothL1(yˆj,yj) Here, Ωbio and Ωstruc denote the sets of target variables belonging to the biochemical and structural groups, respectively. For biochemical traits, the Mean Squared Error (MSE) loss is applied. For structural traits, the Huber loss (Smooth L1) is employed, augmented with an asymmetric weighting factor γ to mitigate the influence of potential outliers in the ground-truth data during model optimization. Second, the inter-group physical correlation constraint (Ψphy) is introduced. To comprehensively exploit the inherent biological negative correlations among variables, the model dynamically computes the cross-covariance for every pair of variables spanning the biochemical and structural groups within each training batch:(14)Cov(yˆi,yˆj)=1B∑n=1B⁡(yˆi(n)−yˆⅈ¯)(yˆj(n)−yˆj¯)(15)Ψphy=∑i∈Ωbio∑j∈ΩstrucSoftplus(Cov(yˆi,yˆj)) Here, B denotes the batch size, while yˆi(n) and yˆj(n) represent the predicted values for biochemical variable i (e.g., N, CP) and structural variable j (e.g., ADF, NDF) for the n-th sample, respectively. The terms yˆⅈ¯ and yˆj¯ denote the mean predictions for these variables within the batch. This formulation comprehensively traverses all variable pairs that are theoretically expected to exhibit negative correlations via a double summation (ΣΣ). The underlying rationale for this mechanism is that When the network predictions align with the negative correlation pattern of carbon-nitrogen metabolism empirically observed in this study (i.e., Cov≤0), the constraint term, after passing through the Softplus activation function, approaches zero, thereby exerting no additional influence on the network optimization. Conversely, if local noise during optimization causes the network to contradict this population-level statistical trend by generating synchronized predictions indicative of a positive correlation (Cov>0), the penalty term Ψphy will increase substantially to produce a gradient that guides the model back toward biologically plausible outputs. Notably, this constraint function is designed to be highly modular and adaptable. For crops exhibiting distinct physiological characteristics such as quality parameters with positive synergistic accumulation, the corresponding agronomic priors can be seamlessly integrated without altering the core network architecture. This is achieved simply by adjusting the covariance sign, for instance, by modifying the penalty term to Softplus(−Cov). This design yields a versatile and generalizable mechanism for embedding domain knowledge into the model. Unlike a simpler Hinge loss (max(0,Cov)) which is non-differentiable at zero and may cause gradient oscillations near the constraint boundary, the Softplus function serves as a smooth and continuously differentiable approximation. This mathematical smoothness ensures stable gradient updates during the network optimization process, which significantly enhances the training stability and convergence of our shallow network topology. Model training and accuracy evaluation strategy Baseline model construction and modeling strategy To rigorously evaluate the efficacy of the proposed PI-SSN architecture for extracting alfalfa quality-related features, three representative and widely adopted machine learning algorithms were employed as baseline models: Gaussian Process Regression (GPR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost, XGB). Specifically, GPR leverages a Bayesian framework, providing distinct statistical advantages for handling nonlinear mappings and regression tasks with limited samples [46]. RF, as a representative bagging ensemble method, mitigates the risk of overfitting associated with high-dimensional features by constructing a multitude of decision trees and averaging their outputs [47]. XGBoost, a gradient boosting algorithm within the boosting framework, incorporates regularization terms to control model complexity, thereby achieving robust nonlinear fitting performance [48]. Furthermore, to elucidate how the inherent correlations among alfalfa's multiple nutritional traits, specifically between biochemical and structural components, influence predictive performance, two distinct modeling paradigms were adopted, namely Single-Task Learning (STL) and Multi-Task Learning (MTL). The STL strategy entailed building separate models for each of the four target traits, comprising total nitrogen (N), crude protein (CP), neutral detergent fiber (NDF), and acid detergent fiber (ADF). Conversely, the MTL strategy implemented in MTL-GPR, MTL-RF, and our proposed PI-SSN requires the algorithm to share underlying feature representations within a unified framework, enabling simultaneous prediction of multiple output variables. This approach is designed to leverage the physiological synergies inherent to carbon-nitrogen metabolism in alfalfa. Dataset partitioning strategy To ensure an objective assessment of generalization capacity and robustness across different models, the multi-temporal remote sensing features and corresponding ground-truth samples for each cutting cycle were randomly partitioned into training and independent test sets, maintaining a 75% to 25% split. Throughout the model training and hyperparameter optimization phase, only the training set was utilized for iterative parameter updates. The independent test set was strictly withheld from all training procedures and was employed solely for the final, unbiased evaluation of predictive accuracy subsequent to model training completion. Model accuracy evaluation metrics To holistically assess the predictive accuracy and stability of the various algorithms, four widely adopted statistical metrics were employed: the coefficient of determination (R2), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Ratio of Performance to Deviation (RPD). Multi-dimensional allometric transfer operator for canopy-to-whole-plant scaling and relative feed value assessment UAV-based multispectral observations are inherently constrained by the limited penetration depth of optical sensors; consequently, the acquired radiative signals predominantly capture the biochemical status of the upper canopy, characterized by a high leaf-to-stem ratio. In contrast, whole-plant quality assessments routinely conducted in agricultural production and forage markets encompass the middle and lower portions of the stem, which are tissues typically characterized by substantial lignification. This pronounced vertical heterogeneity in plant development inevitably results in a substantial scaling discrepancy between canopy-level observations and true whole-plant measurements. To correct this systematic bias, the principles of allometric theory were invoked [49]. Given that carbon and nitrogen allocation in plants differs along vertical and horizontal axes, the three-dimensional canopy architecture of alfalfa is effectively captured by two key structural parameters: plant height (PH) in the vertical dimension and canopy coverage (CC) in the horizontal dimension. A multi-dimensional allometric transfer operator integrating these three-dimensional structural features was developed, formalized by the following mathematical expression:(16)Qwhole=α·Qcanopyβ·CCγ1·PHγ2 Here, Qwhole denotes the target whole-plant quality parameter, Qcanopy represents the canopy-level quality retrieval value, α is the allometric scaling intercept, and β is the transfer exponent for canopy quality. The coefficients γ1 and γ2 correspond to the elasticity coefficients quantifying the respective effects of horizontal canopy coverage and vertical plant height on the vertical profile of quality. This formulation quantitatively captures how the three-dimensional canopy architecture mediates the relative proportions of lower stems and upper leaves, which critically influences whole-plant quality. Parameters for this operator were calibrated using 50 independent alfalfa samples collected during the second cutting (August 11, 2025). A full-data fitting approach was applied to establish a robust population-level statistical baseline for the relationship between spatial structure and the transfer of quality attributes. To further evaluate the practical utility of this scale transfer operator in real-world agricultural contexts, Relative Feed Value (RFV) was adopted as the ultimate validation criterion. RFV is an internationally recognized comprehensive standard for forage quality evaluation, providing critical guidance for determining harvest timing and market pricing. Its calculation strictly depends on dry matter intake (DMI) and digestible dry matter (DDM), which are derived from fiber indicators, using the following formulas:(17)DMI=120NDF(18)DDM=88.9−0.779×ADF(19)RFV=DMI×DDM1.29In the validation phase, the canopy-level fiber estimates directly retrieved from UAV imagery were compared against the whole-plant fiber estimates transformed by the 3D allometric operator. Both fiber datasets were independently input into the RFV calculation formulas. By evaluating the relative error in whole-plant RFV predictions derived from each approach, the capability of the multi-dimensional allometric operator to reconcile observational scale mismatches and accurately recover true field-scale feeding value was rigorously assessed. Results and analysis Sensitivity analysis of alfalfa quality based on high-dimensional spectral feature mining To comprehensively extract signals pertinent to nutritional quality from alfalfa canopy spectra, a total of 14,960 candidate spectral operators encompassing one-to four-dimensional band combinations were formulated and evaluated within both the original reflectance (OR) and log-transformed (log(1/R)) spaces. A low-redundancy feature subset was subsequently identified using statistical thresholding, intended for input into multivariate machine learning models. Sensitivity distribution of basic spectral response and high-dimensional feature space The Pearson correlation coefficient (r) curves between individual bands of canopy original reflectance (OR) and the four quality indicators are shown in Fig. S4, separately for the first (Cut 1) and second (Cut 2) cuttings. The figure reveals that while the overall trajectories of the correlation curves are largely consistent across both cuttings, the absolute magnitudes of the correlation coefficients differ markedly. Within each individual cutting, single bands demonstrated moderate-to-strong linear predictive capability for the nutritional traits. For instance, the maximum absolute correlation coefficients for each quality parameter were approximately 0.50 in the first cutting. In the second cutting, however, correlations were substantially stronger, with the near-infrared band (850 nm) exhibited a positive correlation of 0.78 with total nitrogen (N) and crude protein (CP), while the red band (650 nm) showed negative correlations exceeding 0.80 with acid detergent fiber (ADF) and neutral detergent fiber (NDF). Regarding spectral response patterns, the near-infrared band was positively correlated with biochemical components, whereas the red band was positively correlated with fiber components. This divergence is largely attributable to multiple scattering within the canopy and the reduction in photosynthetic capacity accompanying lignification. However, when samples from the first and second cuttings were pooled to form a cross-cutting dataset, the representational power of individual bands deteriorated markedly. The 1-Band violin plot in Fig. 2 reveals that, within the pooled dataset encompassing both cuttings, the upper bound of absolute correlation coefficients for all single bands fell below 0.40. This observation underscores that using original reflectance from individual wavelengths is insufficient to compensate for systematic spectral variations occurring across different cutting cycles. Consequently, the development of multi-band spectral operators to mitigate environmental noise represents a viable strategy for attaining high-precision retrieval across different cutting cycles.Fig. 2Probability density distribution of absolute correlation coefficients within the 1- to 4-dimensional spectral index search space. Probability density distribution of absolute correlation coefficients within the 1- to 4-dimensional spectral index search space. To assess how band combinations and operator dimensionality enhance signal robustness, Fig. 2 displays probability density violin plots illustrating the distribution of absolute correlation coefficients for candidate features as operator dimensionality increases from 1-Band to 4-Band within the cross-cutting dataset. The contour of each violin plot indicates the density of features within various correlation ranges, with the external red dashed line and asterisk (∗) denoting the maximum correlation coefficient attained at each respective dimensionality. Statistical analysis revealed a pronounced stepwise increase in the maximum correlation coefficients corresponding to each increment in dimensionality. For instance, the maximum absolute correlation coefficient for NDF increased from 0.39 at the 1-dimensional level to 0.896 at the 4-dimensional level. This finding indicates that high-dimensional spectral operators utilizing ratio- and difference-based combinations effectively mitigate spectral drift across cuttings, enabling the extraction of more robust signals associated with quality traits. Moreover, as dimensionality increased from 3-Band to 4-Band, the core distribution region of the violin plots exhibited minimal change, and the incremental gain in the maximum correlation coefficient displayed a converging pattern. This suggests that spectral information approaches saturation at the 3- to 4-dimensional level. Dimensional augmentation of optimal spectral operators and band contribution analysis To elucidate the band composition and spatial characteristics underlying the observed correlation maxima, the optimal spectral operators at each dimensionality were extracted for detailed structural analysis, as presented in Fig. S5 and 3. Fig. S5 provides a comparison of correlation coefficients between the conventional baseline index (NDVI) and the optimal operators across different dimensions. The results reveal a distinct divergence in spectral space preference between original reflectance (OR) and log-transformed (Log) spaces depending on the properties of the nutritional indicator. For biochemical concentration traits (N and CP), the optimal operators consistently favored the log-transformed space (log(1/R)). This preference aligns with the Beer-Lambert law, as the logarithmic reciprocal transformation effectively linearizes the response associated with the net absorption cross-section of biochemical constituents. Accordingly, absorption-scattering decoupling indices formulated within this domain including Log-Pigment attained high correlation levels(|r|>0.86) even at the three-dimensional level. In contrast, for fiber components (ADF and NDF), the optimal operators favored the original reflectance (OR) space when expanded to four dimensions, with the four-band dual-difference (4 B-DD) structure emerging as dominant(|r|=0.896). The optical rationale for this preference lies in the inherently weak direct absorption features of fiber constituents across the visible-to-near-infrared spectrum. Instead, variation in fiber content manifests primarily through modifications in alfalfa's stem-to-leaf ratio and canopy architecture, which in turn influence the macroscopic backscattering field at the canopy surface [50]. Given that original reflectance is more responsive to structural variations, its integration with the dual-difference formulation enables adaptive normalization of photosynthetic decline signals in the visible spectrum. This is achieved by leveraging a structural-sensitive baseline formed by near-infrared and red bands, thereby effectively attenuating apparent reflectance artifacts arising from non-uniform canopy density. Notably, while the single optimal operator for each quality indicator demonstrates a specific spectral space preference at the theoretical extreme, biochemical absorption and physical scattering within vegetation are inherently coupled under actual field conditions. Consequently, relying on a single spectral space is often insufficient to capture the full spectrum of information variability. Thus, the synergistic incorporation of independent operators from both OR and Log spaces facilitates complementary exploitation of multiple optical mechanisms, enabling a more holistic characterization of canopy properties. Additionally, the transition from three-to four-dimensional features reveals a saturation phenomenon in the dimensionality of spectral information. Examination of the evolutionary trajectory of the correlation maxima in Fig. S5 indicates that the incremental gain in absolute correlation from incorporating a fourth band demonstrates pronounced convergence. This suggests that optimized three-band operators exemplified by the MTCI difference ratio structure are already capable of capturing the majority of target biochemical variation while effectively suppressing dominant environmental noise. The addition of a fourth band largely contributes to localized refinement of residual complex scattering signals. This observation underscores the necessity for a judicious balance between the representational fidelity of index operators and their mathematical complexity when constructing multivariate inversion models, thereby mitigating the risk of overfitting. The band contribution heatmap matrix in Fig. 3 quantifies the usage frequency of each central wavelength within the optimal operators across different dimensionalities. The enhanced predictive accuracy achieved by high-dimensional indices is fundamentally attributable to the synergistic complementarity of the distinct physiological and structural information encoded within different wavelengths. Overall, the short-wavelength violet band (405 nm) emerged as a universally dominant feature across the optimal operators for all quality indicators, achieving nearly complete representation across the 1- to 4-dimensional selections. Positioned within the strong absorption region for chlorophyll and carotenoids, the 405 nm band is exceptionally sensitive to both the underlying soil background and atmospheric Rayleigh scattering due to its short wavelength. Within high-dimensional dual-difference (4 B-DD) and anti-noise (VARI) formulations, the 405 nm band is frequently employed as a denominator component or as a background subtraction term. Incorporating this wavelength into spectral formulations enables adaptive elimination of additive noise inherent in complex field environments, thereby establishing a clean vegetation reflectance baseline for other spectral bands.Fig. 3Feature band composition and contribution matrix of optimal 1- to 4-dimensional spectral indices for each nutritional indicator. Feature band composition and contribution matrix of optimal 1- to 4-dimensional spectral indices for each nutritional indicator. Regarding the specific selection of characteristic bands, biochemical and structural components demonstrated markedly distinct physiological preferences. For traits reflecting biochemical metabolism such as crude protein (CP) and total nitrogen (N), the optimal band combinations were heavily concentrated in the green (560 nm, 570 nm) and red-edge (710 nm) regions, notably excluding the near-infrared band (850 nm) typically utilized for biomass retrieval. This selection aligns closely with the optical principles governing coupled nitrogen-pigment metabolism. The green band exhibits maximum sensitivity to variations in moderate-to-high chlorophyll concentrations, circumventing the absorption saturation that frequently occurs in the red spectral region. Meanwhile, the 710 nm band serves as a critical wavelength for characterizing the red-edge position and sensitively detects the red-edge shift induced by increasing nitrogen concentration. Difference or ratio configurations between green and red-edge bands, exemplified by 3 B-Pigment(560, 710, 405), enable these operators to directly isolate the intrinsic biochemical concentration gradients within the canopy. For fiber components (ADF and NDF) representing cell wall constituents, the absence of distinct absorption features for cellulose and lignin across the visible-to-near-infrared spectrum results in a highly dispersed pattern of band utilization within the visible domain, spanning 430 nm, 450 nm, 550 nm, 560 nm, and 570 nm, with nuanced differences in emphasis between ADF and NDF. Increased fiber content is typically associated, albeit indirectly, with an elevated stem-to-leaf ratio and the onset of leaf senescence within the alfalfa canopy. For ADF, the optimal four-dimensional operator 4 B-DD (430, 560, 450, 570) leverages blue-green band combinations to primarily capture changes in canopy reflectance associated with progressive lignification. Notably, the optimal three-band operator for ADF, 3 B-VARI (405, 550, 850), is formulated within the log-transformed (Log) space and incorporates the near-infrared band (850 nm), which is sensitive to cellular structure. Grounded in the Beer-Lambert law, the logarithmic transformation converts reflectance into apparent absorbance, and when coupled with the 850 nm band, this formulation effectively accentuates internal absorption contrasts arising during stem and leaf tissue senescence. For NDF, the optimal features, exemplified by 3 B-MTCI (550, 710, 405) and 4 B-DD (570, 710, 405, 550), exhibit a strong reliance on the combination of the 710 nm band with green wavelengths (550, 570 nm). This finding indicates that the subtle physiological decline signals accompanying the relative accumulation of primary cell wall constituents such as hemicellulose during the transition from vegetative to reproductive growth can be indirectly amplified and quantified by exploiting the high transmittance in the green region and the associated blue shift of the red-edge position. Feature dimensionality reduction and orthogonal subset construction via dual-threshold filtering While a full-space exhaustive search effectively identifies the optimal operators exhibiting the strongest correlation with the target quality traits, it inevitably yields a substantial number of highly autocorrelated features arising from subtle band variations. Feeding all such features directly into a multivariate regression model would induce severe multicollinearity, thereby diluting feature importance and compromising model generalization performance. To address this, a rigorous dual-threshold filtering strategy incorporating target correlation screening and cross-collinearity elimination was implemented to reduce dimensionality within the candidate feature set. Initially, predictive validity screening was conducted. Using the temporally aggregated dataset encompassing both the first and second cuttings, Pearson correlation coefficients were computed between each of the 14,960 candidate operators and the target quality traits. Given the complex nature of field observations, where features with weak responses are prone to introducing random background noise, a stringent inclusion threshold of |rx,y|≥0.85 was adopted. As evidenced by the cutoff lines in the scatter plots of Fig. 4(a), (d), and (g), the majority of low-dimensional operators, comprising single bands and the basic NDVI, were excluded. The initial screening yielded 672 and 844 features that surpassed the inclusion threshold for ADF and NDF, respectively. Moreover, owing to the strong physiological coupling between N and CP in alfalfa metabolism, their spectral sensitivity patterns were highly consistent. Consequently, both traits yielded 180 features passing the initial screening, with identical ranking orders. To prevent redundancy, subsequent feature dimensionality reduction visualizations and analyses in this section are presented using CP, ADF, and NDF as representative examples.Fig. 4Spectral feature dimensionality reduction and orthogonal subset extraction process for alfalfa quality based on dual-threshold filtering.Note: (a)-(c) depict the feature screening process for CP and N; (d)-(f) for ADF; and (g)-(i) for NDF. Panels (a), (d), and (g) illustrate the first-stage predictive validity screening based on target correlation, with the red dashed line indicating the feature inclusion threshold (|r|≥0.85). Panels (b), (e), and (h) show the cross-correlation matrices of the initially retained features, where darker regions signify severe multicollinearity risks (|rxi,xj|>0.90). Panels (c), (f), and (i) present the final orthogonal feature combinations retained, along with their low cross-correlation validation matrices. Spectral feature dimensionality reduction and orthogonal subset extraction process for alfalfa quality based on dual-threshold filtering. Note: (a)-(c) depict the feature screening process for CP and N; (d)-(f) for ADF; and (g)-(i) for NDF. Panels (a), (d), and (g) illustrate the first-stage predictive validity screening based on target correlation, with the red dashed line indicating the feature inclusion threshold (|r|≥0.85). Panels (b), (e), and (h) show the cross-correlation matrices of the initially retained features, where darker regions signify severe multicollinearity risks (|rxi,xj|>0.90). Panels (c), (f), and (i) present the final orthogonal feature combinations retained, along with their low cross-correlation validation matrices. Subsequently, collinearity diagnosis and redundancy elimination were conducted. The feature pool retained after initial screening still contained substantial information overlap. As illustrated in the heatmap matrices of Fig. 4(b), (e), and (h), cross-correlation coefficients |rxi,xj| among a substantial number of high-dimensional operators exceeded 0.90, revealing pronounced redundancy. To ensure independence among model input variables, a correlation-prioritized sequential elimination strategy was implemented. Features were processed in descending order of their target correlation, and any candidate feature exhibiting a cross-correlation coefficient exceeding 0.90 with any feature already retained was flagged as collinear redundant and consequently excluded. Following this dual-filtering procedure, the number of features ultimately retained for each quality trait consistently converged to three, as depicted in Fig. 4(c), (f), (i) and summarized in Table S2. The pairwise cross-correlation coefficients among these three representative features were all maintained at very low levels(|r|≤0.72), indicating a high degree of orthogonality. Analysis of the influence of multi-temporal features and different models on quality inversion accuracy Comparison of single-task learning (STL) model inversion accuracy and the influence of temporal features To assess the efficacy of multi-temporal spectral features for alfalfa quality inversion, the predictive capabilities of Gaussian Process Regression (GPR), Random Forest (RF), and Extreme Gradient Boosting (XGB) models were evaluated within the Single-Task Learning (STL) framework under two input configurations: single-temporal instantaneous (Inst) and multi-temporal (TS) (Table S3). The results demonstrate that incorporating temporal spectral features markedly enhanced the prediction accuracy of most baseline models across all four quality parameters. For instance, for crude protein (CP), the test set coefficients of determination (Test R2) achieved by GPR, RF, and XGB under the TS configuration were respectively 0.125, 0.122, and 0.136 higher than those obtained under the Inst configuration. This finding underscores that, relative to spectral observations acquired at a single time point (e.g., at cutting), the integration of multi-temporal data spanning the branching, budding, and early flowering stages provides a more holistic representation of both dry matter accumulation dynamics and the temporal evolution of nutritional constituents throughout the crop growth cycle. When comparing algorithmic performance, distinct sensitivity patterns emerged across different target traits. For biochemical traits (N and CP) and certain structural traits (ADF), GPR exhibited notable stability, achieving consistently high Test R2 values with minimal discrepancies between training and test set performance. Conversely, tree-based models especially RF showed a pronounced tendency toward overfitting when trained on the limited dataset (Train R2>0.94), which consequently compromised their predictive accuracy on the independent test set. However, for neutral detergent fiber (NDF), XGB leveraged its capacity to model complex nonlinear relationships to attain superior performance within the STL framework (Test R2=0.834). This divergence in model performance provides further evidence for fundamental differences in the spectral response mechanisms governing biochemical versus structural components. Performance comparison of multi-task learning (MTL) joint inversion Single-Task Learning (STL) models conventionally treat each quality parameter as an independent prediction target, thereby neglecting the inherent linkages between dry matter accumulation and nutrient metabolism throughout crop development. Under field conditions, the allocation of photosynthetic assimilates in alfalfa is governed by an inherent trade-off between carbon metabolism, which is responsible for structural carbohydrate synthesis, and nitrogen metabolism, which drives the synthesis of nitrogenous compounds. While conventional purely data-driven Multi-Task Learning models, such as MTL-GPR and MTL-RF, can simultaneously output multiple variables within a unified architecture, their feature extraction processes are vulnerable to interference from uncorrelated noise across variables, precisely because they lack the guidance of agronomic prior knowledge. As evidenced by the evaluation results in Table S4 and Fig. 5, inversion accuracy derived from multi-temporal (TS) spectral features consistently surpassed that from single-temporal instantaneous (Inst) observations across all model architectures. This underscores the critical role of cumulative phenological information throughout the crop growth cycle for accurate alfalfa quality retrieval. However, the incorporation of high-dimensional temporal data revealed distinct limitations inherent to the conventional baseline models. While statistical models exemplified by MTL-GPR demonstrated relatively stable performance, their fundamental reliance on linear kernel functions precluded them from adequately capturing the complex nonlinear relationships embedded within high-dimensional spectral data, thereby imposing an inherent bottleneck on overall inversion accuracy. Conversely, while the tree ensemble model MTL-RF exhibits exceptionally strong nonlinear fitting capabilities, it is highly susceptible to overfitting when trained on limited sample sizes. As illustrated in Fig. 5(A) and (B), MTL-RF achieved exceptionally high correlation and minimal error on the training set, yet its performance deteriorated markedly when evaluated on the independent test set. In the MAE heatmap bubble matrix (Fig. 5C), the test set bubbles displayed pronounced volumetric shrinkage accompanied by intensified coloration, visually corroborating the substantial increase in absolute error incurred during the testing phase. Moreover, owing to its failure to appropriately balance the contributions of biochemical versus structural traits, the test set accuracy of MTL-RF for N and CP (Test R2 = 0.744 for both) fell below that achieved by certain single-task baseline models (Table S3).Fig. 5Comprehensive performance comparison of quality retrieval models based on multi-temporal features. (A) Distribution of the coefficient of determination (R2) for training and test sets. (B) Comparison of RMSE between training and test sets. (C) Evolution matrix of MAE from training to test sets (bubble size and color map strictly to absolute error values, where a lower value indicates smaller error). (D) Distribution of the Ratio of Performance to Deviation (RPD) for training (dashed hollow bars) and test sets (solid colored bars). Comprehensive performance comparison of quality retrieval models based on multi-temporal features. (A) Distribution of the coefficient of determination (R2) for training and test sets. (B) Comparison of RMSE between training and test sets. (C) Evolution matrix of MAE from training to test sets (bubble size and color map strictly to absolute error values, where a lower value indicates smaller error). (D) Distribution of the Ratio of Performance to Deviation (RPD) for training (dashed hollow bars) and test sets (solid colored bars). In contrast, the Physics-Informed Sparse Shallow Network (PI-SSN) developed in this study achieved concurrent enhancements in both predictive accuracy and stability for joint multi-trait inversion. First, by integrating a temporal attention mechanism within a shallow network architecture, PI-SSN effectively extracts spectral information from critical growth stages that is highly relevant to alfalfa quality dynamics, while simultaneously mitigating the risk of overfitting inherent to small-sample conditions through its constrained model depth. Second, the incorporation of the population-level carbon-nitrogen evolutionary trajectory throughout alfalfa development as a soft constraint within the loss function provides the model with regularization grounded in actual field conditions. This effectively curtails the network's propensity for aberrant fitting when trained on limited data. As demonstrated in Table S4, PI-SSN achieved optimal predictive performance for total nitrogen (N), crude protein (CP), and acid detergent fiber (ADF), attaining Test R2 values of 0.813, 0.812, and 0.815, respectively. Remarkably, for neutral detergent fiber (NDF), which is highly susceptible to canopy structural influences, PI-SSN achieved a test set accuracy of 0.848 (Test RPD = 2.491). As illustrated in Fig. 5D, the RPD values for PI-SSN on the training and test sets were highly consistent, providing statistical evidence that the model's superior performance stems not from overfitting but from its authentic learning of the underlying physiological processes governing crop development and biomass accumulation. Further examination of PI-SSN performance using the optimal temporal features (Fig. S6) reveals that measured and predicted values for all four quality indicators tightly cluster around the 1:1 line, with no discernible stage-specific bias or clustering across the two cutting cycles. Notably, the cross-cutting model developed by pooling data from both cycles attained marginally superior overall prediction accuracy relative to models trained on either single cutting alone. This is primarily attributable to the merged dataset encompassing a broader spectrum of quality variation. In contrast, variation within a single cutting is predominantly derived from phenotypic differences among distinct germplasms under uniform environmental conditions, yielding a comparatively constrained data domain and, consequently, a statistically justified R2 moderation. Nonetheless, the model's predictive R2 within individual cuttings consistently remained above 0.73. This demonstrates that PI-SSN effectively suppresses background noise induced by seasonal shifts and canopy structural heterogeneity, accurately captures quality dynamics across distinct growth phases, and possesses the sensitivity to discern subtle spectral differences among germplasms even under relatively uniform backgrounds. Collectively, these capabilities establish PI-SSN as a robust modeling framework for cross-temporal dynamic quality assessment in complex field environments. Mechanistic deconstruction and model interpretability of the PI-SSN architecture To determine whether the proposed PI-SSN architecture authentically internalizes the physiological development patterns of crops, the underlying parameters of the converged model were extracted and subjected to interpretability analysis, integrating these with inference results from the withheld independent test set (Fig. 6).Fig. 6Internal decision mechanisms and interpretability analysis of the PI-SSN architecture. (A) Differentiated attention weight allocation strategies between the biochemical parameter prediction branch (blue bars) and the structural parameter prediction branch (orange bars) given identical temporal spectral inputs. (B) Comparison of relative weight magnitude distributions in the first hidden layer between the unpenalized biochemical branch and the L1-regularized structural branch. (C) Cross-correlation (r) distributions between predicted variables (Crude Protein and Neutral Detergent Fiber) on the independent test set, illustrating the regularization and calibration effect of the physical soft constraint against correlation over-amplification. Internal decision mechanisms and interpretability analysis of the PI-SSN architecture. (A) Differentiated attention weight allocation strategies between the biochemical parameter prediction branch (blue bars) and the structural parameter prediction branch (orange bars) given identical temporal spectral inputs. (B) Comparison of relative weight magnitude distributions in the first hidden layer between the unpenalized biochemical branch and the L1-regularized structural branch. (C) Cross-correlation (r) distributions between predicted variables (Crude Protein and Neutral Detergent Fiber) on the independent test set, illustrating the regularization and calibration effect of the physical soft constraint against correlation over-amplification. First, Fig. 6A validates the underlying logic of the trait-driven temporal attention decoupling mechanism, visually illustrating how the dual-stream network allocates attention to spectral features acquired at different time points. This study is based on 127 global germplasm accessions, with significant genetic heterogeneity in actual phenological progress among varieties. Therefore, the multi-temporal spectral sequences essentially record the dynamic matter accumulation process of the population from early vegetative growth to just before the unified cutting (T3). Analysis revealed that when processing identical feature inputs, the biochemical prediction branch (N, CP) and the structural prediction branch (ADF, NDF) evolved markedly distinct temporal weighting patterns. The biochemical branch assigned its highest attention weight (0.459) to the early stage (T1), a pattern consistent with the agronomic reality of rapid nitrogen assimilation and pronounced protein accumulation during alfalfa's early vegetative phase. Conversely, the structural branch tasked with fiber retrieval allocated its highest perceptual weight (0.532) to the T3 stage, coinciding with the period just prior to cutting. This differentiated decoupling of underlying attention weights remarkably mirrors the objective developmental progression, wherein structural constituents, which are plant cell wall components, exhibit continuous accumulation as the growth stage advances, culminating at later stages. Second, the normalized network weight decay curves presented in Fig. 6B substantiate the practical efficacy of L1 regularization in facilitating feature selection within the dual-stream topology. In the biochemical branch that did not receive L1 penalization (blue line), the weights of hidden layer neurons exhibited a gradual, continuous distribution, thereby preserving sufficient parametric flexibility to accommodate the continuous variation in nutrient concentrations. Conversely, under the influence of L1 penalization, the structural branch (orange line) displayed a rapidly decaying weight distribution. Numerous redundant temporal features exhibiting weak relevance to fiber retrieval were spontaneously eliminated by the model, driving their corresponding weights toward zero. This induced sparsity within the underlying network topology effectively mitigates feature redundancy and curtails the risk of overfitting, particularly when the model confronts the complexities of canopy structural heterogeneity. Finally, Fig. 6C leverages inference results from the independent test set to demonstrate the regularization and calibration effect of the biology-constrained joint loss function on multi-task collaborative inversion outcomes. Scatter plot regression analysis reveals that within the purely data-driven paradigm, baseline models (gray squares) successfully captured the negative correlation trend characteristic of carbon-nitrogen metabolism. However, when inferring unknown samples, the degree of negative correlation between predicted variables (r=−0.895) displayed a discernible offset relative to the true laboratory baseline (r=−0.835). In contrast, the PI-SSN model (purple diamonds), by incorporating the physical loss function as a fundamental constraint, achieved improved calibration of its multivariate joint output distribution. The resulting cross-correlation (r=−0.878) of the predicted outputs more closely approximated the true distribution observed in the germplasm nursery compared to the baseline model. This finding demonstrates that embedding agronomic statistical priors as constraints within the network effectively rectifies inference biases on unknown samples, thereby safeguarding the biological plausibility of multi-task collaborative inversion outcomes. Spatial distribution and interpretation of large-scale germplasm quality based on UAV imagery Fig. 7 illustrates the spatial distribution of four canopy apparent quality indicators across 127 alfalfa germplasms over two cuttings based on the PI-SSN model. Globally, the overall quality of the second cutting was significantly superior to that of the first. This inter-cutting difference profoundly highlights the impacts of meteorology and management. Continuous rainfall forced the first harvest's delay until early July. Consequently, plant aging caused severe physiological protein dilution. Conversely, the second cutting underwent a standard 35-day regeneration, reaching the optimal bud-to-early-bloom stage. The top young leaves endowed it with excellent apparent feed value.Fig. 7Spatial distribution of four apparent canopy quality indicators across 127 alfalfa germplasms over two cuttings, derived from the PI-SSN model. Spatial distribution of four apparent canopy quality indicators across 127 alfalfa germplasms over two cuttings, derived from the PI-SSN model. Within the same cutting, distribution maps exhibit rich spatial heterogeneity across different plots. This intuitively reflects the diverse genetic backgrounds of the 127 germplasms. Consequently, they were not at identical phenological stages during the same UAV flight observation. Spectral signals acutely captured the intrinsic differences between early- and late-maturing varieties regarding leaf-to-stem ratios and carbon-nitrogen metabolic rhythms. These differences were then expressed with high spatial differentiation. This proves the model can macroscopically reveal population fluctuations caused by meteorological disturbances. Furthermore, it microscopically identifies germplasm developmental diversity, providing robust mapping support for elite cultivar screening. However, it must be emphasized that these results represent only the apparent canopy quality of alfalfa. Limited by optical penetration depth, UAV signals essentially capture only the top young leaves. Consequently, they severely lack information regarding the basal lignified stems. Directly equating this to actual field harvest values inevitably leads to a systematic overestimation of crude protein and an underestimation of fiber. To break this surface observation barrier, subsequent evaluations must overcome 2D spectral limitations. This requires introducing 3D architectural parameters, like plant height and canopy coverage, as allometric compensation terms. By constructing a vertical profile scale transfer operator, a rigorous conversion from the apparent canopy state to the absolute whole-plant quality can be achieved. Vertical profile quality evolution and “canopy-to-whole-plant” scale transfer Observational bias analysis between canopy and whole plant A paired comparison of canopy and whole-plant assay data from 50 representative alfalfa core germplasms was conducted (Fig. 8A). Results revealed systematic numerical shifts across all nutritional indicators during this scale conversion. Biochemical components, specifically N and CP, decreased significantly at the whole-plant scale. This demonstrates a dilution effect driven by the basal stems. Their population mean concentrations dropped from 4.35% to 27.28% in the canopy to 3.40% and 21.26% in the whole plant, respectively. Conversely, structural components, including ADF and NDF, increased significantly at the whole-plant scale. Their mean contents surged from 21.52% to 31.08% in the canopy to 27.27% and 41.53% in the whole plant. This reflects the substantial accumulation of structural carbohydrates within the stems. If canopy predictions derived from the UAV PI-SSN model are directly applied for field evaluation (Fig. 8B, gray scatter), the predictive distribution severely deviates from the 1:1 baseline. Consequently, this generates a systematic overestimation of field protein content and a systematic underestimation of fiber content.Fig. 8Vertical profile quality evolution of alfalfa and correction analysis of the 3D architecture allometric transfer operator. (A) Data distribution shifts of four nutritional indicators (N, CP, ADF, NDF) between canopy and whole-plant scales, revealing the dilution effect of biochemical components and the enrichment characteristics of structural components. (B) Scatter plots and error convergence comparison of whole-plant quality predictions based on raw UAV canopy retrievals (gray dots) versus 3D operator corrections (colored dots). Vertical profile quality evolution of alfalfa and correction analysis of the 3D architecture allometric transfer operator. (A) Data distribution shifts of four nutritional indicators (N, CP, ADF, NDF) between canopy and whole-plant scales, revealing the dilution effect of biochemical components and the enrichment characteristics of structural components. (B) Scatter plots and error convergence comparison of whole-plant quality predictions based on raw UAV canopy retrievals (gray dots) versus 3D operator corrections (colored dots). Impact of 3D structural parameters on quality transfer and operator correction To correct these biases, we applied a 3D multi-dimensional allometric transfer operator for model fitting. The structural elasticity coefficient distribution for each quality indicator (Fig. 9A) reveals a clear pattern. Plant height (PH) and canopy coverage (CC) exert highly specific physical constraints during the cross-scale transfer of various biochemical and structural components.Fig. 9Cross-scale prediction loop of Relative Feed Value (RFV) and analysis of 3D architecture elasticity parameters. (A) Distribution of nonlinear elasticity coefficients of horizontal canopy coverage (γ1) and vertical plant height (γ2) for the cross-scale transfer of quality indicators (B) Comparison of whole-plant RFV prediction accuracy between raw canopy retrievals and 3D operator corrections. Cross-scale prediction loop of Relative Feed Value (RFV) and analysis of 3D architecture elasticity parameters. (A) Distribution of nonlinear elasticity coefficients of horizontal canopy coverage (γ1) and vertical plant height (γ2) for the cross-scale transfer of quality indicators (B) Comparison of whole-plant RFV prediction accuracy between raw canopy retrievals and 3D operator corrections. For biochemical metabolism indicators like total nitrogen (N) and crude protein (CP), both the plant height elasticity (γ2) and coverage elasticity (γ1) coefficients are positive. Agronomically, N and CP are primarily concentrated within plant leaves. A positive coverage elasticity (γ1 >0) indicates a dense canopy with a high leaf-to-stem ratio. This effectively offsets the physiological dilution caused by basal stems, thereby maintaining whole-plant biochemical concentrations. Meanwhile, a positive plant height elasticity (γ2 > 0) reflects robust vertical plant growth. Taller plants typically possess broader root systems and stronger source-sink assimilation capacities. During 3D spatial expansion, this vigorous nitrogen absorption and metabolism positively compensates for maintaining whole-plant protein concentrations. For structural components characterizing cell wall materials (ADF and NDF), their coverage elasticity coefficient γ1 is negative. High coverage inherently indicates an increased leaf proportion. Consequently, a lush canopy largely buffers the rise in overall fiber content. However, they exhibit completely opposite evolutionary trends regarding the plant height elasticity coefficient (γ2). Specifically, ADF's γ2 is positive, whereas NDF's γ2 is negative. This divergence accurately reflects the asynchronous accumulation dynamics of alfalfa cell wall components during 3D spatial expansion. Specifically, ADF primarily comprises recalcitrant cellulose and lignin, representing highly lignified secondary cell walls. As plant height increases, the crop must heavily deposit lignin at the basal stems. This enhances mechanical support to resist wind loads, counteract gravity, and prevent lodging. This bottom-up structural reinforcement causes a positive increase in whole-plant ADF concentrations compared to the canopy in taller plants (γ2 > 0). Conversely, NDF additionally encompasses hemicellulose, which is widely present within the primary cell walls of young shoots and leaves. During rapid internode elongation, basal secondary cell walls (ADF) thicken locally. However, rapid vacuolar expansion within the stems is accompanied by a massive influx of non-structural carbohydrates, such as soluble sugars. This causes a physiological dilution of the relative accumulation rate of total NDF at the whole-plant scale (γ2 < 0). By feeding the PI-SSN-derived UAV canopy retrievals into this 3D transfer operator, this study validates the critical value of physical operators in eliminating spatial misalignments inherent to remote sensing observations. Consequently, this framework accurately eliminates retrieval errors regarding true field quality caused by relying solely on canopy spectra (Fig. 9B). Comprehensive assessment of field relative feed value (RFV) Because RFV calculations depend simultaneously on NDF and ADF, minor deviations in these fiber parameters are heavily amplified within the comprehensive index. As illustrated in Fig. 9B, relying exclusively on UAV canopy spectra causes severe overestimation of whole-plant RFV. Consequently, its root mean square error (RMSE) fails to satisfy practical production standards. Following the application of the 3D allometric transfer operator to correct fiber parameters, the scatter distribution of RFV predictions improved markedly. Simultaneously, the RMSE decreased substantially. These results demonstrate that fusing multispectral information with physical structural parameters, such as plant height and canopy coverage, effectively enables accurate conversion from canopy phenotypes to whole-plant feed values. Discussion Deep exploration of spectral feature space and band combination mechanisms When retrieving vegetation physicochemical parameters, single bands or traditional two-band vegetation indices (e.g., NDVI) are frequently constrained by canopy structural heterogeneity, soil background, and spectral saturation effects under high canopy closure [18]. By constructing a full-space feature library of 1 to 4 dimensions, this study demonstrates the superiority of high-dimensional band combinations and spectral space transformations in complex field environments. These findings are highly consistent with classical optical theories while revealing unique response patterns specific to distinct quality indicators. Globally, the shortwave violet band (405 nm) distinctly dominates the optimal operators for all quality parameters as a background deduction term. This effectively suppresses interference from soil background and atmospheric additive scattering [51]. The dual-threshold redundancy reduction strategy serves as a critical filter that ensures the extracted features are grounded in biophysical mechanisms. While the initial correlation threshold isolates high-sensitivity signals, the subsequent collinearity exclusion guarantees that the retained orthogonal features represent independent physiological dimensions, such as biochemical absorption, canopy scattering, and background noise deduction, rather than mathematical artifacts. The consistent retention of biologically meaningful wavelengths, specifically the 405 nm band for background correction and the 710 nm band for tracking nitrogen-induced spectral shifts, confirms that the final feature subsets are fundamentally driven by crop physiology rather than spurious statistical correlations. For the retrieval of biochemical components (N, CP), the optimal features do not rely on the near-infrared band (850 nm) commonly used for biomass estimation. Instead, they focus heavily on the green region (560 nm, 570 nm) and the red-edge band (710 nm), synergized with the logarithmic (Log) transformation space. This aligns with the expectations of the Beer-Lambert law, indicating that logarithmic transformation mitigates nonlinear scattering. Furthermore, the combination of green and red-edge bands effectively prevents absorption saturation in the red region. Crude protein in alfalfa is highly enriched within chloroplast macromolecules (e.g., RuBisCO enzymes). Its metabolic accumulation exhibits a strong homologous coupling characteristic with nitrogen nutrition. Consequently, this band combination precisely isolates the red-edge shift induced by nitrogen concentration while simultaneously capturing the internal optical response of crude protein within the canopy. This finding is highly consistent with the conclusions of Fu et al. [52] and Fan et al. [53], confirming the extreme sensitivity of the red-edge band for predicting crop protein and nitrogen contents. Conversely, structural components like cellulose (ADF, NDF) lack direct characteristic absorption peaks in the visible to near-infrared region. Consequently, their utilized bands exhibit a more scattered distribution. This study reveals that higher-order dual-difference operators (4 B-DD) successfully capture the apparent canopy absorbance contrast caused by increased stem-leaf ratios and senescence. They achieve this by combining visible bands (e.g., 430 nm, 550 nm) with specific reference bands. This further corroborates the theory proposed by Haboudane et al. [19] and Saeed et al. [54], demonstrating that multi-band combinations can effectively decouple biochemical absorption signals from physical structural signals. This breakthrough overcomes the retrieval bottlenecks of traditional indices in evaluating fibrous materials. Superiority of the Physics-Informed Sparse Shallow Network (PI-SSN) over traditional purely data-driven algorithms In small-sample scenarios of agricultural remote sensing, balancing the nonlinear fitting capability and generalization performance of deep learning models remains a critical challenge. By comparing traditional baselines under Single-Task Learning (STL) and Multi-Task Learning (MTL) architectures, this study reveals the inherent limitations of purely data-driven algorithms. Although tree-based ensemble models like RF fit the training set highly, they suffer from severe performance degradation on independent test sets [38]. Conversely, while GPR exhibits relatively stable generalization, it struggles to adequately capture the nonlinear dynamics within multi-stage phenological time series. In contrast, the proposed Physics-Informed Sparse Shallow Network (PI-SSN) demonstrates significant advantages in both accuracy and mechanistic consistency. First, the Trait-Driven Decoupled Temporal Attention (TDTA) mechanism of PI-SSN autonomously allocates extremely high observation weights at the early stage (T1) for biochemical components. Simultaneously, it shifts the perceptual focus for fiber components to the pre-harvest stage (T3). This foundational attention allocation perfectly aligns with the objective agronomic developmental pattern, namely that early growth determines the protein baseline, while late phenology dominates fiber accumulation [55,56]. Secondly, during joint multi-task retrieval, traditional models frequently suffer from correlation overfitting due to the absence of logical constraints between variables. For instance, the baseline models herein amplified the negative correlation among components in the test set to r = −0.895. PI-SSN innovatively introduces a biology-constrained joint loss function (cross-covariance penalty) based on carbon-nitrogen metabolic tradeoffs. This mechanism forces the distribution of retrieved results to converge within natural, biologically permissible boundaries (r = −0.878). This demonstrates that embedding biological statistical priors into the network as penalty terms serves as an excellent model regularizer. It eliminates local perturbations while ensuring the ecological rationality of synergistic multi-indicator outputs [57]. Empirical evaluations demonstrate that the training and inference speeds of PI-SSN are highly comparable to traditional machine learning baselines, as the physics-informed constraint imposes negligible computational overhead during the inference phase of the sparse shallow network. This near-instantaneous execution capability confirms the framework's robust feasibility for high-throughput, pixel-level quality mapping across extensive UAV orthomosaics in large-scale field applications. Evolution patterns of vertical profile quality and the value of “canopy-to-whole-plant” scale conversion Historically, limited by the penetration depth of optical sensors, UAV-based phenotyping has largely remained restricted to the canopy scale. However, alfalfa exhibits significant vertical heterogeneity. The shedding of senescent basal leaves and the presence of highly lignified stems cause physiological protein dilution and structural fiber enrichment at the whole-plant level. This study quantified this observation bias and found that if the canopy inversion results are directly applied to the field, an overestimation of protein of about 22.0% and an underestimation of fiber of more than 30% will occur. This systematic bias severely limits the practical application of UAV remote sensing in forage trading and feed formulation, such as RFV assessment. To bridge this gap, we developed a 3D allometric transfer operator using canopy coverage (CC) and plant height (PH). Increased CC partially buffers the dilution effect caused by a higher proportion of stems. Meanwhile, the PH parameter effectively captures the plant's allometric strategy. This strategy preferentially allocates structural carbohydrates (ADF) to the stem base to enhance mechanical support [58]. Applying this operator brought biased UAV predictions close to the 1:1 ground-truth line and significantly reduced RFV estimation RMSE. This confirms that coupling 3D architectural parameters is essential for high-throughput UAV phenotyping. Translating canopy surface observations into 3D whole-plant metrics is a necessary, viable step for capturing true field quality. Furthermore, vertical errors originating from baseline topography interpolation and photogrammetric smoothing could propagate into the CHM, potentially introducing uncertainties into downstream 3D allometric calculations. Research limitations and future prospects Although the 127 global germplasm accessions evaluated herein provide a highly representative foundation of genetic diversity for multi-temporal quality retrieval, validating the spatiotemporal transferability of the PI-SSN model and 3D transfer operator across broader ecological zones and multi-year scales remains essential to account for complex environmental perturbations. Furthermore, expanding beyond current statistical covariance constraints to assimilate classical crop mechanistic models (e.g., APSIM or DSSAT) into deep temporal networks would enable a critical transition from passive state retrieval to proactive quality prediction and dynamic harvest scheduling. Conclusion To address critical issues in UAV-based alfalfa quality monitoring, including canopy spectral saturation, obscure multi-stage dynamic responses, and canopy-to-whole-plant scale misalignment, this study developed a novel quality estimation framework. This framework integrates high-dimensional spectral feature mining, a Physics-Informed Sparse Shallow Network (PI-SSN), and a 3D allometric transfer operator. The main conclusions are as follows.(1)High-dimensional spectral operators effectively decouple spectral variations across different harvests. Using a 1- to 4-dimensional feature library based on original reflectance and logarithmic transformation spaces, we found that feature correlations increased stepwise with higher dimensions. The maximum absolute correlation coefficient for NDF increased from 0.39 in 1D to 0.896 in 4D. Following a dual-threshold filtration, the three retained orthogonal features for each indicator exhibited cross-correlation coefficients below 0.70. This significantly reduced the risks of feature redundancy and multicollinearity.(2)The PI-SSN demonstrates both accuracy and generalization advantages in small-sample multi-task retrievals. The Trait-Driven Decoupled Temporal Attention mechanism enables the biochemical branch to focus on early observations (T1 weight: 0.459). Meanwhile, the structural branch emphasizes late-stage accumulation (T3 weight: 0.532), aligning with the temporal heterogeneity of alfalfa's carbon-nitrogen metabolism. A biological tradeoff constraint aligns the cross-correlation of predicted variables with the actual distribution of measured populations. The test set accuracy (R2 = 0.812-0.848) outperformed traditional multi-task baselines. Furthermore, the closely matched relative errors between the train and test sets confirm the absence of overfitting.(3)The 3D allometric transfer operator effectively corrects systematic observation biases between the canopy and whole-plant scales. Constructed using canopy coverage and plant height, this operator simulates the carbon allocation compensation mechanism during 3D plant spatial expansion. Following correction by this operator, biases involving protein overestimation and fiber underestimation caused by superficial UAV retrievals were significantly suppressed. Consequently, the estimation accuracy for Relative Feed Value (RFV) improved substantially. This validates the necessity and effectiveness of coupling 3D structural parameters for scale conversion. (1)High-dimensional spectral operators effectively decouple spectral variations across different harvests. Using a 1- to 4-dimensional feature library based on original reflectance and logarithmic transformation spaces, we found that feature correlations increased stepwise with higher dimensions. The maximum absolute correlation coefficient for NDF increased from 0.39 in 1D to 0.896 in 4D. Following a dual-threshold filtration, the three retained orthogonal features for each indicator exhibited cross-correlation coefficients below 0.70. This significantly reduced the risks of feature redundancy and multicollinearity. High-dimensional spectral operators effectively decouple spectral variations across different harvests. Using a 1- to 4-dimensional feature library based on original reflectance and logarithmic transformation spaces, we found that feature correlations increased stepwise with higher dimensions. The maximum absolute correlation coefficient for NDF increased from 0.39 in 1D to 0.896 in 4D. Following a dual-threshold filtration, the three retained orthogonal features for each indicator exhibited cross-correlation coefficients below 0.70. This significantly reduced the risks of feature redundancy and multicollinearity. (2)The PI-SSN demonstrates both accuracy and generalization advantages in small-sample multi-task retrievals. The Trait-Driven Decoupled Temporal Attention mechanism enables the biochemical branch to focus on early observations (T1 weight: 0.459). Meanwhile, the structural branch emphasizes late-stage accumulation (T3 weight: 0.532), aligning with the temporal heterogeneity of alfalfa's carbon-nitrogen metabolism. A biological tradeoff constraint aligns the cross-correlation of predicted variables with the actual distribution of measured populations. The test set accuracy (R2 = 0.812-0.848) outperformed traditional multi-task baselines. Furthermore, the closely matched relative errors between the train and test sets confirm the absence of overfitting. The PI-SSN demonstrates both accuracy and generalization advantages in small-sample multi-task retrievals. The Trait-Driven Decoupled Temporal Attention mechanism enables the biochemical branch to focus on early observations (T1 weight: 0.459). Meanwhile, the structural branch emphasizes late-stage accumulation (T3 weight: 0.532), aligning with the temporal heterogeneity of alfalfa's carbon-nitrogen metabolism. A biological tradeoff constraint aligns the cross-correlation of predicted variables with the actual distribution of measured populations. The test set accuracy (R2 = 0.812-0.848) outperformed traditional multi-task baselines. Furthermore, the closely matched relative errors between the train and test sets confirm the absence of overfitting. (3)The 3D allometric transfer operator effectively corrects systematic observation biases between the canopy and whole-plant scales. Constructed using canopy coverage and plant height, this operator simulates the carbon allocation compensation mechanism during 3D plant spatial expansion. Following correction by this operator, biases involving protein overestimation and fiber underestimation caused by superficial UAV retrievals were significantly suppressed. Consequently, the estimation accuracy for Relative Feed Value (RFV) improved substantially. This validates the necessity and effectiveness of coupling 3D structural parameters for scale conversion. The 3D allometric transfer operator effectively corrects systematic observation biases between the canopy and whole-plant scales. Constructed using canopy coverage and plant height, this operator simulates the carbon allocation compensation mechanism during 3D plant spatial expansion. Following correction by this operator, biases involving protein overestimation and fiber underestimation caused by superficial UAV retrievals were significantly suppressed. Consequently, the estimation accuracy for Relative Feed Value (RFV) improved substantially. This validates the necessity and effectiveness of coupling 3D structural parameters for scale conversion. In conclusion, the methodological framework proposed herein provides novel insights for high-throughput, multi-temporal remote sensing estimation of perennial forage quality. Furthermore, it offers a reliable technological tool for precision pasture management and the screening of superior germplasm resources. Author contributions Tao Cheng: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Validation, Visualization, Writing-original draft, Writing-review & editing. Zhangru Gao: Formal analysis, Investigation, Methodology, Validation, Visualization, Writing-original draft. Weibo Ren: Methodology, Resources, Validation, Writing-review & editing. Ziyu Ding: Investigation, Methodology, Validation. Feng Yuan: Methodology, Writing-review & editing. Yaling Liu: Resources, Writing-review & editing. Jiyuan Xie: Investigation. Ruopu Zan: Investigation. Dong Wang: Conceptualization, Resources, Supervision, Writing-review & editing. Dongyan Zhang: Conceptualization, Funding acquisition, Project Administration, Supervision, Writing-original draft, Writing-review & editing. Data and code availability The pre-trained model weights, inference code, and usage instructions are publicly available at https://github.com/AeroPheno/PI-SSN.git. The raw spectral data and ground-truth quality data used in this study are not publicly available due to ongoing collaborative projects, but are available from the corresponding author on reasonable request. Funding This work was supported by the 2023 Hohhot to introduce high-level innovative and entrepreneurial talents (team) [grant number 2023RC-High Level 7], the Inner Mongolia Autonomous Region Science and Technology Plan Project [grant number 2025KYPT0170, 2025YFHH0205], National Center of Pratacultural Technology innovation Special fund for innovation platform construction [grant numbers CCPTZX2024N02], and the 2023 Inner Mongolia Autonomous Region “Unveiling and Hanging” Project [grant number 2023JBGS0008]. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.