Explainable cotton mapping in northern and southern Xinjiang using Boruta feature selection and SHAP analysis

Cotton is a globally significant crop, and its accurate mapping is crucial for efficient agricultural monitoring and sustainable management. Remote sensing combined with machine learning offers a powerful solution, yet the inclusion of numerous features often leads to spectral confusion and data redundancy. This study integrated multi-source Sentinel-1/2 data, extracting 104 initial features, including spectral bands and vegetation indices. To mitigate redundancy, the Boruta feature selection algorithm was employed, optimizing the feature subsets to 31 for northern Xinjiang (S1) and 29 for southern Xinjiang (S2). Eight classification models were systematically compared, with the Convolutional Neural Network (CNN) demonstrating superior performance. The CNN achieved an overall accuracy (OA) of 96.43% and a $Kappa$ coefficient of 0.9482 in S1, and an OA of 94.61% with a $Kappa$ of 0.9168 in S2. Furthermore, Shapley Additive Explanations (SHAP) provided an interpretable analysis of the model's decision-making. SHAP results highlighted regional differences: GDVI and NIR (B9) were dominant in S1, while GI and NDVIre3 were key in S2. This Boruta-CNN-SHAP framework provides a robust, interpretable approach for large-scale cotton mapping in arid oases.