Large-scale rice mapping under spatiotemporal heterogeneity using multi-temporal SAR images and explainable deep learning

Timely and accurate mapping of rice cultivation distribution is crucial for ensuring global food security and achieving SDG2. From a global perspective, rice areas display high heterogeneity in spatial pattern and SAR time-series characteristics, posing substantial challenges to deep learning (DL) models’ performance, efficiency, and transferability. Moreover, due to their “black box” nature, DL often lack interpretability and credibility. To address these challenges, this paper constructs the first SAR rice dataset with spatiotemporal heterogeneity and proposes an explainable, lightweight model for rice area extraction, the eXplainable Mamba UNet (XM-UNet). The dataset is based on the 2023 multi-temporal Sentinel-1 data, covering diverse rice samples from the United States, Kenya, and Vietnam. A Temporal Feature Importance Explainer (TFI-Explainer) based on the Selective State Space Model is designed to enhance adaptability to the temporal heterogeneity of rice and the model’s interpretability. This explainer, coupled with the DL model, provides interpretations of the importance of SAR temporal features and facilitates crucial time phase screening. To overcome the spatial heterogeneity of rice, an Attention Sandglass Layer (ASL) combining CNN and self-attention mechanisms is designed to enhance the local spatial feature extraction capabilities. Additionally, the Parallel Visual State Space Layer (PVSSL) utilizes 2D-Selective-Scan (SS2D) cross-scanning to capture the global spatial features of rice multi-directionally, significantly reducing computational complexity through parallelization. Experimental results demonstrate that the XM-UNet adapts well to the spatiotemporal heterogeneity of rice globally, with OA and F1-score of 94.26 % and 90.73 %, respectively. The model is extremely lightweight, with only 0.190 M parameters and 0.279 GFLOPs. Mamba’s selective scanning facilitates feature screening, and its integration with CNN effectively balances rice’s local and global spatial characteristics. The interpretability experiments prove that the explanations of the importance of the temporal features provided by the model are crucial for guiding rice distribution mapping and filling a gap in the related field. The code is available in https://github.com/SAR-RICE/XM-UNet.