Analysis of Plant Nutrient Deficiencies Using Multi-Spectral Imaging and Optimized Segmentation Model Journal: Artificial Intelligence in Agriculture
Ji-Yan Wu, Zheng Yong Poh, Anoop C. Patil, Bongsoo Park, Giovanni Volpe, Daisuke Urano · 2025
Read Analysis of Plant Nutrient Deficiencies Using Multi-Spectral Imaging and Optimized Segmentation Model Journal: Artificial Intelligence in Agriculture by Ji-Yan Wu, Zheng Yong Poh, Anoop C. Patil, Bongsoo Park, Giovanni Volpe, Daisuke Urano (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-14013.txt.
Introduction
Ji-Yan Wu and five co-authors built a deep learning system in twenty twenty-five to spot nutrient deficiencies in plant leaves from multispectral images. Early detection lets farmers fix fertilizer, disease, and stress problems before yields suffer. Their plants were grown under controlled nutrient stress so symptoms could be labeled reliably.
The system enhances the YOLOv5 detector with a transformer attention head and feeds it nine-channel multispectral input. Self-attention helps it catch faint, scattered symptoms. Against the baseline model, it improved overlap scores by about twelve percent on average. It proved especially good at hard symptoms like chlorosis, the yellowing of starved leaves, and pigment buildup. The authors pitch the approach as a step toward sharper plant phenotyping for precision agriculture.
Catalog record
| CREATOR | Ji-Yan Wu, Zheng Yong Poh, Anoop C. Patil, Bongsoo Park, Giovanni Volpe, Daisuke Urano |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-14013
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Analysis of Plant Nutrient Deficiencies Using Multi-Spectral Imaging and Optimized Segmentation Model Journal: Artificial Intelligence in Agriculture Ji-Yan Wu, Zheng Yong Poh, Anoop C. Patil, Bongsoo Park, Giovanni Volpe, Daisuke Urano Abstract Accurate detection of nutrient deficiency in plant leaves is essential for precision agriculture, enabling early intervention in fertilization, disease, and stress management. This study presents a deep learning framework for leaf anomaly segmentation using multispectral imaging and an enhanced YOLOv5 model with a transformer-based attention head. The model is tailored for processing nine-channel multispectral input and uses self-attention mechanisms to better capture subtle, spatially-distributed symptoms. The plants in the experiments were grown under controlled nutrient stress conditions for evaluation. We carry out extensive experiments to benchmark the proposed model against the baseline YOLOv5. Extensive experiments show that the proposed model significantly outperforms the baseline YOLOv5, with an average Dice score and IoU (Intersection over Union) improvement of about 12 % 12\% . In particular, this model is effective in detecting challenging symptoms like chlorosis and pigment accumulation. These results highlight the promise of combining multi-spectral imaging with spectral-spatial feature learning for advancing plant phenotyping and precision agriculture. 1 Introduction The strategic implementation of precision agriculture technologies, particularly image-based anomaly detection, is transforming modern agricultural practices [ 1 , 2 ] . With the help of satellite, drone, and proximal imaging, farmers can now monitor crop health at scale to optimize fertilization, detect early signs of nutrient deficiency, and manage disease outbreaks [ 3 ] . However, to make these systems truly automated and actionable, robust computer vision algorithms capable of accurately detecting anomalies across thousands of field images are needed. Most current image-based plant health monitoring systems rely on conventional RGB cameras, which are limited to three spectral bands. In contrast, multi-spectral imaging captures reflectance data across a broader spectral range—including near-infrared (NIR), and even short wave-length infrared bands (SWIR)—which provide richer information on plant physiology and stress response [ 4 , 5 ] . NIR reflectance, in particular, is closely associated with leaf senescence and chlorophyll content, and has been widely used in vegetation indices such as Normalized Difference Vegetation Iindex (NDVI) and Normalized Difference Red-Edge index (NDRE) [ 6 ] . Affordable multi-spectral cameras typically use dual-sensor configurations to separately capture visible and NIR bands, which is advantageous for aerial imaging but introduces parallax and alignment issues at close range [ 7 ] . In contrast, multi-channel sensors with a shared optical path eliminate registration errors and enable precise analysis of plant features from the same viewpoint. Deep learning methods have shown strong promise in plant anomaly detection tasks [ 8 ] , outperforming traditional machine learning models in both accuracy and adaptability. However, most state-of-the-art deep learning architectures, such as YOLO and Mask R-CNN, are designed for RGB images and must be adapted to handle high-dimensional multi-spectral data. Challenges include input channel compatibility, feature fusion across spectral bands, and generalization under varying stress scenarios. Recent studies have also shown the effectiveness of combining multi-spectral sensing with AI-based models for early plant stress detection. For instance, Wei et al. [ 9 ] used multi- and hyper-spectral images to classify disease symptoms in forests. Similarly, deep learning models have been applied for plant phenotyping and genebank analysis across multi-spectral inputs [ 10 ] . This study addresses the following key challenges in applying deep learning to multi-spectral leaf analysis. On the one hand, there is a need for spectral feature selection and fusion: Given the increased number of spectral bands, it is critical to design models that can selectively emphasize the most informative wavelengths for detecting early-stage anomalies. On the other hand, also the model architecture needs to be adapted: Conventional deep learning models require modifications to process 9-channel inputs and perform pixel-wise segmentation across subtle, overlapping symptom classes. To address these challenges, we propose a deep learning framework based on the YOLOv5 object segmentation architecture [ 11 ] , enhanced with a transformer-based attention head [ 12 ] to improve contextual reasoning across spectral-spatial features. Our model is trained on a newly constructed dataset of Marchantia polymorpha grown under five controlled nutrient conditions: fully fertilized control as well as deficiencies in nitrate, phosphate, calcium, and iron. A nine-channel multi-spectral camera was used to capture reflectance data over 17 days. Pixel-wise annotations were made for four visually distinct symptom categories: normal, chlorosis, pigment accumulation, and tipburn. These symptom classes correspond directly to specific nutrient deficiencies observed in our controlled experiments. In particular, chlorosis (yellowing of leaf tissue) is commonly associated with iron and nitrate deficiencies, while pigment accumulation (typically reddish or purple discoloration) results from stress due to phosphate or nitrate imbalance. Tipburn, characterized by necrotic leaf margins, is a hallmark of calcium deficiency. The normal class includes uniformly healthy green tissue grown under fully fertilized conditions. By establishing this clear linkage between visual symptoms and physiological nutrient stressors, our annotation framework enables biologically meaningful segmentation for downstream phenotyping and stress classification. Experimental results demonstrate that the proposed method outperforms the RGB-only YOLOv5 baseline, especially in detecting small or scattered lesion patterns. The main contributions of this work are the following. First, we develop a novel leaf anomaly segmentation framework that extends the YOLOv5 architecture to support nine-channel multi-spectral image input. To better capture subtle spatial and spectral variations, we integrate a transformer-based attention head into the YOLOv5 output layers. This optimization enhances the model’s ability to detect small or dispersed defects such as chlorosis and marginal pigment accumulation, which are often underrepresented in RGB-based detection models. Second, we construct a high-quality, pixel-level annotated dataset of multi-spectral plant images under controlled nutrient stress conditions. The dataset spans five nutrient regimes (control, -Fe, -N, -P, -Ca) and includes visual and spectral symptoms across three plant species. Each image is annotated with fine-grained polygon masks representing four classes (normal, chlorosis, pigment accumulation, tipburn), enabling robust supervised segmentation training and evaluation. Finally, we conduct extensive experiments comparing our proposed model against the baseline RGB-based YOLOv5. The results demonstrate consistent improvements in segmentation performance, achieving up to 17% higher mean Average Precision (mAP) and 12% higher average Dice score across all classes. These findings confirm the effectiveness of integrating multi-spectral information and transformer-based attention for plant stress detection. The rest of this article is organized as follows: Section 2 reviews related work on multi-spectral imaging and object segmentation. Section 3 introduces the architecture and training strategy. Section 4 presents the experimental design, dataset, and evaluation results. Section 5 concludes the study and outlines future research directions. 2 Related Work Relevant previous works can be divided into three categories: i) the analysis of plant healthy and disease status using multi-spectral imaging; ii) deep learning models for object detection and segmentation; iii) deep learning models for multi-spectral and hyperspectral image analysis. The following subsections discuss the previous efforts in these areas. 2.1 Plant Healthy Analysis Using Multi-Spectral Images RGB imaging with three wavelength channels has been the most prevalent data for plant anomaly detection. Recently, multispectral and hyperspectral imaging are gaining traction for their higher spectral resolution and sensitivity to plant health [ 13 ] . One notable study was conducted by Charles Veys et al. [ 5 ] who demonstrated proposed usingthe