Interpretable Multi-Year Winter Wheat Mapping with Sentinel-1/2 Time Series: SHAP-Based Feature Selection and Bayesian-Optimized Machine Learning

Abstract

Accurate winter wheat mapping is important for interannual planting-area monitoring, yet models developed independently for each year require repeated sample collection, feature selection, and parameter tuning. This study developed an interpretable cross-year transfer approach for winter wheat mapping in the Tailan River Irrigation District, Xinjiang, China, using Sentinel-1/2 time-series imagery. The model was developed entirely from 2025 source-year data, including feature selection, Bayesian hyperparameter optimization, and model training. The resulting model was then applied directly to the 2023 and 2024 target-year test sets without retraining, retuning, or feature reselection. From 189 optical, SAR, texture, and vegetation-index temporal-change features, SHAP-based selection retained 28 informative variables. The Bayesian-optimized Random Forest model achieved overall accuracies of 93.82% in the 2025 source-year validation and 92.13% and 91.01% in the 2023 and 2024 target-year tests, with corresponding Kappa coefficients of 0.92, 0.90, and 0.89. SHAP and ablation analyses showed that May-to-June vegetation index change rate features derived from NDVI, NDre1, and EVI provided the most consistent phenological information, while Sentinel-1 features supplied complementary structural and backscatter information. The extracted winter wheat area decreased from 11,270.27 ha in 2023 to 9834.97 ha in 2025, consistent with the downward trend in official statistics. These results show that a model developed from a single source year maintained consistent classification performance across the evaluated historical years within the same irrigation district.