Multi-Source Attention-Based Fusion for Segmentation of Natural Disasters

The detection of natural disasters is a critical step in supporting Sustainable Development Goal (SDG) 15. In this letter, we introduce a novel segmentation method to identify the areas affected by floods and landslides, using multiple sources such as multispectral imagery, synthetic aperture radar (SAR) data, digital elevation models (DEMs), slope, and hydrology maps. Traditional deep learning semantic segmentation algorithms encounter difficulties in handling multi-source data. In response to this challenge, we have refined the single-channel encoder-decoder architecture to better discern features derived from multiple sources. We introduce multi-source semantic segmentation with channel and spatial attention fusion (MSSCSAF-Net), a novel solution for semantic segmentation. MSSCSAF-Net incorporates a multi-source encoder (MSE) designed to extract features from diverse inputs. This technique involves the addition of a multi-source channel spatial attention fusion module (MSCSAFM) to fuse the extracted features and a decoder channel spatial attention fusion module (DCSAFM) to improve the fusion with attention of spatial and abstract features. The evaluation on the MMFlood and Landslide4Sense datasets finds that our approach exceeds the performance of current state-of-the-art (SOTA) segmentation techniques.