SAR Image Water Extraction Based on Saliency Target Detection

Abstract

Synthetic aperture radar (SAR) image water extraction has important research significance in water resources monitoring and other applications. In SAR images, the scattering properties of some land cover types with low backscattering coefficients, such as broad roads and mountain shadows, are very close to the scattering properties of water. Most water extraction methods are easy to identify those false water targets as water. Meanwhile, the water area in the scene is very small, and the proportion of water and background in the training concentration is extremely unbalanced. The imbalance of water samples affects the water extraction performance. Traditional machine learning water extraction methods require manual extraction of effective features, which are time-consuming and inefficient. Therefore, we decided to introduce the deep learning-based salient object detection network Pool-Net into SRA image water extraction. To overcome the effect of class imbalance, we introduce the focus loss function into the Pool-Net network. Experimental results show that the proposed method achieves good water extraction results on water class imbalanced SAR images. The water extraction accuracy of the proposed method is 2% higher than Pool-Net and also surpasses some well-known image segmentation methods.