Multi-level Feature Attention Fusion Network for Water Extraction from Multi-source SAR Images

Objective At present, water extraction from single frequency-band synthetic aperture radar （SAR） images has achieved great success. However, due to the large differences in the characteristics of water bodies and other surface targets in SAR images with different frequency-bands, deep learning network designed for single frequency-band has poor extraction accuracy when applied to SAR images with different frequency bands. So, how to achieve excellent extraction performance for water automatically from multi-source SAR images is still a challenge.

Methods We propose a new network framework, namely multi-level feature attention fusion network (MFAFNet). The network consists of encoder and decoder. The encoder utilizes ResNet-101 to generate four-level features with different resolutions, which are then processed in parallel by the intermediate level feature fusion module and efficient channel atrous spatial pyramid pooling module, and the intermediate and high-level features are output after deep fusion. The attention modulation module attention mechanism is introduced into decoder to distribute the weight of low-level features, then which are further fused with the high-level features from the encoder to obtain the water extraction results.

Results Experiments are carried out on SAR images with different frequency bands and resolutions (Sentinel-1, TerraSAR and Gaofen-3), and three good networks have been compared. The results show that the MFAFNet extraction effect is obviously better than other networks, with an average pixel accuracy of 87% and an average intersection over union of 0.80.

Conclusions The water body extraction from multi-band and multi-resolution SAR images is achieved, which promotes the practical application value of deep learning network in water body detection.