# Explainable AI-Based Hyperspectral Classification Reveals Differences in Spectral Response over Phenological Stages

## TL;DR

This study shows that hyperspectral imaging combined with explainable AI can effectively monitor wheat nitrogen levels, with better results when using binary nitrogen groups and early growth stage measurements.

## Contribution

The study introduces an explainable AI framework for hyperspectral classification of wheat nitrogen status, revealing key spectral regions and optimal phenological stages for monitoring.

## Key findings

- Binary nitrogen stratification (Low–High) achieved the highest classification accuracy at both booting and heading stages.
- Plot-level aggregation improved classification performance significantly, reaching perfect accuracy at the heading stage.
- SHAP analysis identified red, red-edge, and near-infrared wavelengths as key contributors to nitrogen classification.

## Abstract

Nitrogen fertilizer has a critical role in determining grain yield and quality in cereal crops, but overfertilization can significantly impact environmental quality. This study tested whether proximal hyperspectral reflectance measurements can be used to identify wheat nitrogen status in the field. Wheat was grown with different nitrogen fertilizer rates, and measurements were taken at two growth stages, booting and heading. Three nitrogen stratification strategies (binary Low–High, Extreme, and three-level) were evaluated using Random Forest, SVM-RBF, and XGBoost classifiers. The results showed that separating nitrogen levels into only two groups, low and high, worked best and gave the most reliable results. Trying to split nitrogen levels into three groups gave poor results because the groups were too similar. Combining several measurements from the same plot improved the final decision compared with using single measurements. Overall, the study shows that simple nitrogen grouping and early measurements can support better nitrogen monitoring in wheat fields.

Optimizing nitrogen (N) fertilization is essential for sustaining durum wheat yield and grain quality while reducing the environmental impacts associated with N over-application. Hyperspectral sensing provides a rapid and non-destructive approach for monitoring crop N status. However, high-dimensional data, phenology-dependent spectral responses, and spatial autocorrelation in field measurements limit robust nitrogen classification and interpretation. This study evaluated hyperspectral-based nitrogen status classification in durum wheat under Mediterranean field conditions and identified key spectral regions using explainable artificial intelligence. A field experiment was conducted in Southern Italy using ten N fertilization rates (0–180 kg N ha−1). Canopy reflectance was acquired at the booting and heading stages from georeferenced sampling locations. Three nitrogen stratification strategies (binary Low–High, Extreme, and three-level) were evaluated using Random Forest, SVM-RBF, and XGBoost classifiers. Model performance was assessed using spatially independent Leave-One-Plot-Out cross-validation at both the sample and plot levels, with plot-level predictions derived through majority voting. Classification robustness was strongly influenced by the stratification strategy and phenological stage. The binary Low–High stratification achieved the highest sample-level accuracy, with a maximum of 0.78 at booting (SVM-RBF) and 0.75 at heading (SVM-RBF), whereas the Extreme stratification produced intermediate performance, with maximum accuracies of 0.73 at booting (SVM-RBF) and 0.63 at heading (XGBoost). Plot-level aggregation improved performance, reaching up to 0.90 at booting and 1.00 at heading. SHAP analysis highlighted red, red-edge, and near-infrared wavelengths as the dominant contributors, with increased reliance on longer wavelengths at the heading. Overall, explainable machine learning provides a robust framework for hyperspectral nitrogen monitoring in durum wheat.