Deep Learning Model Sharpens Maize Nitrogen and Chlorophyll Monitoring

A new study published in the Journal of Chemometrics presented a new deep learning framework designed to dramatically improve the accuracy and interpretability of nitrogen (N) and chlorophyll (Chl) estimation in maize canopies. This study, which was led by researcher Li Tian of Heilongjiang Bayi Agricultural University in China, introduces a hybrid neural network architecture that could advance precision agriculture by delivering rapid, nondestructive biochemical assessments at field scale (1). The study’s findings demonstrate the potential of deep learning and explainable artificial intelligence (AI) in estimating nitrogen and chlorophyll.

What role do canopy nitrogen and chlorophyll content in optimizing fertilizer application?

Using fertilizers to grow crops is a common practice. Fertilizers give farmers the necessary nutrients their crops need in order to grow and produce the best yield. When it comes to canopy nitrogen and chlorophyll content, these two variables are essential in enabling precise fertilization through monitoring (2). Analyzing the nitrogen and chlorophyll content also helps ensure that agricultural practices are more sustainable and that their environmental impacts are reduced (2).

Spectroscopy has played a role in this process as well. For example, near-infrared (NIR) spectroscopy has long been used to analyze plant biochemical components. However, researchers have noted that traditional machine learning (ML) approaches often struggle to model the nonlinear behavior of spectral data (1). These limitations pose challenges for applying NIR spectroscopy to real-time nitrogen management, where both accuracy and model interpretability are vital.

What did the researchers do in their study?

In their study, the research team developed a hybrid model that merges convolutional neural networks (CNNs) with gated recurrent units (GRUs), which were further enhanced with a convolutional block attention module (CBAM). This architecture enables the system to extract spatial and temporal patterns embedded within hyperspectral measurements while assigning dynamic weights to the most informative features (1). The addition of explainable artificial intelligence (AI) tools allows researchers and agronomists to understand how specific wavelengths contribute to the model’s predictions (1).

As part of the experimental procedure, the research team collected hyperspectral images from 200 maize canopy samples and applied a rigorous preprocessing pipeline. Sequential Savitzky–Golay smoothing (SG), standard normal variate (SNV), and SG transformations boosted the mean test set R² by 0.016 units, improving the overall clarity of spectral signals (1). Next, the researchers sought to eliminate the redundant data that were collected. To do so, they employed two dimensionality reduction methods: successive projection algorithm (SPA) and competitive adaptive reweighting sampling (CARS). These two methods helped shrink the spectral feature set from 176 bands down to 10 and 22 key wavelengths (1).

When comparing the CNN-GRU-CBAM model to the most traditional machine learning (ML) and deep learning models, the CNN-GRU-CBAM model outperformed them. On the test set, the CNN-GRU-CBAM model achieved an R² of 0.934 for nitrogen and 0.788 for chlorophyll, along with low RMSE values of 1.940 and 0.216, respectively (1). To further validate the model, SHapley Additive exPlanations (SHAP) analysis identified the spectral regions most responsible for prediction accuracy, confirming the biological relevance of the model’s outputs.

What is the main takeaway from this study?

The main takeaway from this study is that integrating deep learning with explainable AI has the potential to improve agricultural monitoring practices. As the researchers demonstrated in their study, their framework allows for biochemical content inversion across multiple crop species (1). By integrating deep learning with explainable AI, the approach provides a more transparent, accurate, and scalable tool for modern precision agriculture.

References

1. Kong, H.; Tian, L.; Yi, S.; et al. Estimating Maize Canopy Nitrogen and Chlorophyll Content Using CNN-GRU-CBAM and Hyperspectral Imagery. J. Chemom. 2025, 39 (12), e70093. DOI: 10.1002/cem.70093
2. Zhao, J.; Gao, A.; Wang, B.; et al. Modeling and Visualization of Nitrogen and Chlorophyll in Greenhouse Solanum lycopersicum L. Leaves with Hyperspectral Imaging for Nitrogen Stress Diagnosis. Plants (Basel) 2025, 14 (21), 3276. DOI: 10.3390/plants14213276