SHAP-DNN Approach Advances Remote Sensing Mapping of Forested Wetlands

作者：Hengxing Xiang, Fudong Yu, Jialing Bai, Xinying Shi, Ming Wang, Hengqi Yan, Yanbiao Xi, Zongming Wang, Dehua Mao · 发表于：IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 年份：2025 · DOI：10.1109/jstars.2025.3541718 · 被引用次数：11 · 研究领域：Remote Sensing and LiDAR Applications

Mapping the spatial distribution of forested wetlands accurately is crucial for assessing the impact of climate change on terrestrial ecosystems in the mid to high latitudes of the Northern Hemisphere. However, identifying forested wetlands is challenging due to the complex interactions among vegetation, hydrology, and the distinct environmental conditions of these regions. Here, we developed an interpretable Shapley additive explanation deep neural network (SHAP-DNN) classification method for large-scale mapping of forested wetlands using Sentinel-1, Sentinel-2, and PALSAR-2 satellite imagery from 2022. This approach integrated a comprehensive set of features, including polarization, spectral, phenological, and topographical features, and provided the interpretability of classification results. The resulting forested wetlands map in the Northeast China's permafrost zones, named NP_Forested Wetlands, achieved an overall accuracy of over 80% and an F1 score above 0.7. It identified a total forested wetland area of 22 042 km2, predominantly located in continuous and island permafrost regions. The SHAP model indicates that PALSAR HV polarization, slope, Sentinel-1 VV polarization, canopy structure index, and digital elevation model are the most influential factors for distinguishing forested wetlands. Compared to previously published datasets, our SHAP-DNN method provides improved accuracy in delineating forested wetlands, demonstrating its effectiveness for rapid and reliable m...