Comparative analysis of explainable machine learning integrated with hyperspectral imaging for early prediction of...

Wheat is one of the most widely cultivated crops worldwide and provides essential nutrition for millions of people. Accurate and timely wheat yield mapping is critical for strategic planning and decision making to ensure global food security, particularly through early prediction at the field scale to support precision agriculture. Recently, hyperspectral imaging (HSI) integrated with machine learning (ML) techniques has emerged as a robust and reliable approach for assessing crop characteristics and predicting yield. This study combines Visible Near Infrared (VNIR) HSI data (400-1000 nm) with explainable ML models to enhance early and accurate wheat yield prediction and enable comprehensive image based agricultural analysis. The performance of Partial Least Squares Regression (PLSR), Random Forest (RF), and Convolutional Neural Networks (CNN) was compared. Hyperparameters of the CNN model were optimized using Bayesian optimization, resulting in superior performance with R² of 0.76, RMSE of 1022.24 gm/plot, and RPD of 2.03 compared with optimized PLSR and RF models. Three distinct explainable artificial intelligence (XAI) methods (Kernel-SHAP, Tree-SHAP and DeepExplainer) were further employed to analyze the predictions of the PLSR, RF, and CNN models and to determine the relative importance of key wavelengths. The CNN model developed using important wavelengths was further employed to visualize the spatial distribution of leaf regions most influential for yield prediction. These findings demonstrate the effectiveness of integrating HSI with explainable ML for advanced agricultural analysis and reliable early yield prediction.