EXPLORING SAR-CENTERED MULTI-SOURCE DATA FUSION FOR FLOOD MAPPING WITH TRANSFORMER AND CONVOLUTION BASED ENSEMBLE MODELS

The occurrence of severe hydrometeorological disasters, especially flood, has been escalating at an alarming rate. It is of great interest in exploiting the all-time and all-weather monitoring capability of SAR data to achieve rapidly and accurately flood mapping. However, the interferential speckle noises, the hill shadow and the lacking texture water body appearance in SAR data inevitably lead to false segmentation. To address these issues, we propose an effective deep learning framework to acquire flood map with SAR-centered multi-source data fusion. Specifically, we integrate both representative transformer based (Segformer) and convolution based (DCNv4+UperNet) semantic segmentation networks, thus combines both global self-attention and local inductive bias to improve the performance. Moreover, multi-band input data preprocessing, weighted model ensemble, and additional training and testing strategies are utilized to increase the segmentation accuracy. Comprehensive experiments demonstrate the effectiveness of our proposed flood mapping method which ranks in the third place of the 2024 IEEE GRSS Data Fusion Contest Track 1 test phase (F1:79.043%).