Delineation of Surface Water Bodies from SAR Imagery Based on Improved MRF and CNN Model

The precise extraction and automatic identification of water bodies play an important role in water management and flood control. Synthetic aperture radar (SAR) is one of the vital sensors for monitoring the distribution and change of surface water, especially when encountering severe weather. However, the precision of water extraction from SAR images is still unsatisfactory due to the high false alarms rates over urban fringes and mountainous regions. In this paper, an improved method combining modified Markov random field (MRF) with convolutional neural network (CNN) model is proposed to extract surface water bodies from SAR image. Firstly, the initial water maps are obtained by training CNN-based model. Secondly, the primary water maps are subsequently modified by the MRF model with Gamma distribution, which could significantly reduce the false alarms and misclassifications in the water maps. The effectiveness of the proposed strategy has been validated based on Sentinel-1 SAR images, which are four typical application scenes with different characteristics of water resources. The experimental results show that the overall classification accuracy of the proposed method is much higher than that of the other methods. The proposed framework has the potential for water detection from SAR images, which are essential for water management and monitoring.