Crop sample prediction and early mapping based on historical data: Exploration of an explainable FKAN framework

Accurate and timely crop mapping is essential for food security assessment, and high-quality feature factors are the core foundation for accurate mapping. However, deep learning model crop classification algorithms have achieved some success, while the models themselves struggle to explain the specific contribution and impact of different features on the results. In this study, a self-adaptive Feature-attention Kolmogorov-Arnold Network (FKAN) is proposed for interpretable and scalable crop mapping. The model integrated the adaptive weighted feature attention module (AWFA) and the interpretable KAN network, which can visualize the complex associations between features and target crops and automatically capture and filter effective key spatiotemporal features, thus enhancing the interpretability of the model. Experimental results demonstrate that integrating optical, radar, and terrain features yields superior performance in both sample prediction and crop mapping, surpassing existing methods. The proposed FKAN achieves an overall accuracy and F1 score exceeding 0.90. Optical and radar features contribute the most significantly to classification accuracy, while terrain data provides complementary enhancement. By aligning with key crop phenology and leveraging the Google Earth Engine (GEE), FKAN establishes the first operational platform for global winter wheat identification, enabling accurate and scalable crop mapping. The migrated model achieves over 85% accuracy across different regions and years, demonstrating strong robustness and generalization capability. The study identifies optimal phenological periods and feature indices for different crops, providing scientific guidance for future mapping efforts. The FKAN model demonstrated robustness, scalability, and interpretability, was able to automatically extract high-confidence pixels and generate crop planting probabilities, providing an efficient and scalable solution for large-scale crop monitoring. This study generated the first global winter wheat map GlobalWinterWheat10m dataset by the FKAN algorithm.