Abstract: Winter wheat is one of the most important staple crops in China and worldwide, and the timely and accurate acquisition of winter wheat information is crucial for agricultural monitoring, yield assessment, and food security. However, the automatic generation of high-quality training samples and the interpretability of classification models remain key challenges limiting the reliability and applicability of current crop identification and mapping approaches. Based on the NDVI time-series product derived from the FengYun-3D meteorological satellite, this study proposes a winter wheat identification method integrating automatic sample generation with an interpretable CatBoost model. A High-Quality Segment Anything Model (HQ-SAM) is introduced to automatically generate high-quality training samples, and effectively reduce labor cost and uncertainty associated with manual sample collection. To further enhance classification performance, the Whale Optimization Algorithm (WOA) is employed to optimize the CatBoost classifier. In addition, the Shapley Additive Explanations (SHAP) method is applied to provide quantitative and visual interpretations of model predictions, enabling analysis of temporal feature contributions. Experimental results demonstrate that: (1) HQ-SAM can reliably generate high-quality training samples for winter wheat identification; (2) the WOA-optimized CatBoost model achieves an overall accuracy of 96.43% and an F1 value of 96.36%; (3) SHAP-based analysis reveals that the importance ranking and contribution directions of temporal phases are consistent with winter wheat phenology; and (4) the combination of the top 17 most important temporal phases yields the highest classification accuracy.

Key words: FengYun-3D, sample generation, model interpretability, winter wheat, crop mapping