📄 agriculture

Systematic evaluation of hyperspectral imaging workflows for predicting pigment and nutrient traits in tomato leaves under varying nitrogen supply levels

This study systematically evaluated hyperspectral imaging workflows to identify optimal combinations of sample partitioning, spectral preprocessing, feature selection, and machine learning models for non-destructively predicting pigment and nutrient traits in tomato leaves, revealing that while nitrogen-related traits were reliably predicted using CNNs, chlorophyll prediction remained limited and that workflow optimization must be tailored to specific target traits.

Original authors: Yige Zhang, Xiaolong Ma, Chengda Niu, Junfeng Zhang, Jitao Wang, Zibin Xiao, Shiwei Liu, Hua Ma, Cui Yin, Longguo Wu, Yune Cao

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yige Zhang, Xiaolong Ma, Chengda Niu, Junfeng Zhang, Jitao Wang, Zibin Xiao, Shiwei Liu, Hua Ma, Cui Yin, Longguo Wu, Yune Cao

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quiet rhythm of a greenhouse, a tomato plant makes a silent demand. It needs nitrogen, a fundamental nutrient that acts as the engine for growth, driving the formation of green leaves and the development of fruit. When a farmer provides too little, the plant struggles, its leaves turning pale and its growth stalling. When too much is given, the plant may grow lush but weak, or worse, it may accumulate harmful levels of nitrates that can affect food safety and the environment. For decades, knowing exactly how much nitrogen a plant has absorbed has been a guessing game or a slow, destructive process. Farmers have had to pluck a leaf, crush it in a lab, and wait days for a chemical analysis, a method that is too slow to guide daily decisions in a fast-moving crop cycle.

To solve this, scientists have turned to a technology that sees the world in a thousand shades of color. Instead of the three colors our eyes perceive, hyperspectral cameras capture hundreds of narrow bands of light, creating a detailed fingerprint for every surface they scan. This light fingerprint changes depending on what is inside the leaf, revealing secrets about its water, its structure, and its chemical makeup. The challenge, however, is that this data is incredibly complex. It is not enough to simply point a camera at a plant and get an answer; researchers must carefully choose how to clean up the noisy data, which specific colors to focus on, and which mathematical tools to use to translate light into a number. The question is whether one single method can work for every nutrient, or if each trait requires its own unique path to discovery.

A team of researchers in Ningxia, China, set out to find the answer by working with tomato plants in a controlled greenhouse. They grew a popular variety called 'Provence' under four different feeding regimes: one group received no nitrogen at all, while the others received increasing amounts, ranging from a moderate 210 kilograms per hectare up to a heavy 390 kilograms. Over the course of the growing season, they visited the plants at three critical moments: when they were flowering and setting fruit, when the fruit was ripening, and finally at harvest. At each visit, they selected specific leaves and used a portable hyperspectral camera to scan them, capturing the way light bounced off the leaf surface. Immediately after scanning, they took those same leaves to the lab to measure their actual chemical contents, determining the total amount of chlorophyll, the total nitrogen, and the nitrate levels using standard, precise chemical tests. This created a perfect match between the light signature and the real-world chemistry for hundreds of samples.

The researchers then treated this data like a complex puzzle, testing dozens of different ways to solve it. They tried splitting the data into training and testing groups in four different ways to ensure their models were fair. They applied seven different methods to clean up the light signals, removing the static and noise that often cloud these measurements. They also tested two distinct strategies for picking out the most important colors from the hundreds available, asking which specific wavelengths held the key to the answer. Finally, they fed this refined information into four different types of computer learning models, ranging from established statistical tools to advanced deep learning networks, to see which one could best predict the chemical values based on the light alone.

The results revealed a clear truth: there is no single "best" way to predict every trait. The path to an accurate answer depended entirely on what the researchers were trying to measure. When the goal was to estimate the total chlorophyll, the best results came from a specific type of data splitting and a simple smoothing technique, but even the best model struggled to make precise predictions. The team found that the variation in chlorophyll across the different growth stages made it difficult to build a single, reliable rule. However, when the focus shifted to nitrogen and nitrate, the story changed. For total nitrogen, a method that selected a compact set of key wavelengths worked best, while for nitrate, a different selection strategy was superior.

The type of computer model used also mattered deeply. For the chlorophyll, a support vector machine, a robust and traditional algorithm, performed the best. But for the more complex tasks of predicting total nitrogen and nitrate, a convolutional neural network, a type of deep learning model that mimics the way the brain processes patterns, outperformed all others. This deep learning model was able to see subtle, non-linear connections in the light data that the simpler models missed, providing much more reliable predictions for the nitrogen traits. The study showed that the red-edge region of the spectrum, where light reflectance shifts sharply, and the near-infrared region were the most informative areas for distinguishing between the different nitrogen levels.

Ultimately, the work demonstrates that precision agriculture cannot rely on a one-size-fits-all approach. To accurately monitor the health of a tomato crop without damaging a single leaf, the entire workflow must be tailored to the specific nutrient in question. While the prediction of total chlorophyll remains challenging under the conditions tested, the ability to reliably predict total nitrogen and nitrate levels offers a powerful new tool. By matching the right data cleaning, the right wavelength selection, and the right computer model to the specific trait, farmers can move closer to a future where they know exactly what their plants need, allowing them to apply fertilizer with surgical precision rather than broad guesswork.

Technical Summary: Systematic Evaluation of Hyperspectral Imaging Workflows for Predicting Pigment and Nutrient Traits in Tomato Leaves

Problem Statement
Rapid, non-destructive detection of pigment and nutrient traits in tomato (Solanum lycopersicum L.) leaves is critical for precision fertilization and greenhouse management. While conventional biochemical assays (e.g., spectrophotometry, Kjeldahl digestion) are accurate, they are destructive, labor-intensive, and unsuitable for high-throughput monitoring. Although hyperspectral imaging (HSI) has emerged as a promising alternative, existing studies often focus on individual traits using isolated models. This approach limits the establishment of robust, unified analytical workflows capable of handling varying growth stages and cultivation conditions. Furthermore, there is a lack of systematic evaluation regarding how sample partitioning, spectral preprocessing, feature wavelength selection, and model architecture interact to influence prediction accuracy for multiple traits (chlorophyll, total nitrogen, and nitrate) simultaneously.

Methodology
The study employed a systematic workflow to evaluate HSI for predicting total chlorophyll, total nitrogen (N), and nitrate content in tomato leaves under four nitrogen supply levels (0, 210, 300, and 390 kg N ha⁻¹) across three growth stages: flowering-fruiting, ripening, and harvest.

Data Acquisition: Hyperspectral images (400–1100 nm) were captured using a push-broom camera. Reflectance was calibrated using white and dark references. Regions of interest (ROIs) were manually selected, and average spectral reflectance was extracted.
Ground Truth: Corresponding leaf samples were analyzed for total chlorophyll (ethanol extraction), total N (Kjeldahl digestion), and nitrate (salicylic acid-sulfuric acid method).
Data Processing Workflow:
- Sample Partitioning: Four strategies were compared: Random Sampling (RS), Kennard-Stone (KS), Sample Set Partitioning based on joint X-Y distances (SPXY), and Random Systematic Sampling.
- Preprocessing: Seven methods were evaluated: Moving Average (MA), Savitzky-Golay (SG), Gaussian Filtering (GF), Median Filtering (MF), Normalization, Baseline Correction, and Standard Normal Variate (SNV).
- Feature Selection: Competitive Adaptive Reweighted Sampling (CARS) and Successive Projections Algorithm (SPA) were used to select characteristic wavelengths.
- Modeling: Four machine learning algorithms were developed and compared: Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), and Convolutional Neural Network (CNN).
Evaluation: Models were assessed using the coefficient of determination ( $R^2$ ), Root Mean Square Error (RMSE), and Residual Prediction Deviation (RPD).

Key Results
The study found that optimal workflow components were highly trait-dependent rather than universal:

Sample Partitioning: Systematic sampling yielded the best results for total chlorophyll; Random Sampling (RS) was optimal for total N; and SPXY performed best for nitrate.
Spectral Preprocessing: Normalization was most effective for chlorophyll; Moving Average (MA) provided the best balance for total N; and SNV yielded the highest performance for nitrate.
Wavelength Selection: CARS outperformed SPA for total chlorophyll and nitrate prediction. Conversely, SPA was superior for total N prediction, achieving an $R_p^2$ of 0.895 and RPD of 2.387, indicating good quantitative prediction ability.
Model Performance:
- Total Chlorophyll: SVM performed best among the tested models, though validation accuracy remained limited ( $R_p^2 = 0.431$ ).
- Total Nitrogen: CNN achieved the highest validation accuracy ( $R_p^2 = 0.780$ ), outperforming traditional machine learning models which suffered from overfitting (e.g., RF and XGBoost showed high $R_c^2$ but low $R_p^2$ ).
- Nitrate: CNN also achieved the best performance ( $R_p^2 = 0.753$ ), demonstrating strong generalization.
Spectral Responses: Significant spectral differentiation among nitrogen treatments was observed primarily in the red-edge (700–750 nm) and near-infrared (750–1000 nm) regions, particularly at the ripening and harvest stages.

Key Contributions

Trait-Specific Workflow Optimization: The paper demonstrates that a "one-size-fits-all" approach to HSI modeling is suboptimal. It provides empirical evidence that sample partitioning, preprocessing, wavelength selection, and model architecture must be tailored to specific target traits (pigment vs. nutrient vs. inorganic nitrogen).
Comparative Model Analysis: It offers a direct comparison of ensemble learning, traditional machine learning, and deep learning (CNN) for tomato leaf traits, highlighting that CNN is particularly effective for complex nitrogen-related traits where spectral responses are nonlinear and multifactorial.
Nitrate Prediction: The study successfully establishes a workflow for non-destructive nitrate prediction, a trait often more challenging to estimate than total nitrogen due to its indirect spectral signatures.

Significance and Claims
The authors claim that these findings provide a methodological basis for non-destructive assessment of tomato leaf nitrogen status and precision fertilization management. The study emphasizes that while leaf nitrogen-related traits (total N and nitrate) can be predicted with reasonable reliability using optimized HSI workflows, total chlorophyll prediction remains challenging under pooled-stage modeling strategies.

The paper concludes that the development of robust HSI workflows requires the optimization of each step (partitioning, preprocessing, selection, modeling) for the specific trait of interest. While the results are promising for greenhouse applications, the authors modestly note that broader validation across different cultivars, seasons, and observation scales (canopy level) is necessary before widespread deployment. The study does not claim to have solved the chlorophyll prediction challenge definitively but identifies the limitations of current pooled-stage approaches.

Technical Summary: Systematic Evaluation of Hyperspectral Imaging Workflows for Predicting Pigment and Nutrient Traits in Tomato Leaves

More like this