New deep learning method for efficient extraction of small water from remote sensing images

Abstract

Extracting water bodies from remote sensing images is important in many fields, such as in water resources information acquisition and analysis. Conventional methods of water body extraction enhance the differences between water bodies and other interfering water bodies to improve the accuracy of water body boundary extraction. Multiple methods must be used alternately to extract water body boundaries more accurately. Water body extraction methods combined with neural networks struggle to improve the extraction accuracy of fine water bodies while ensuring an overall extraction effect. In this study, false color processing and a generative adversarial network (GAN) were added to reconstruct remote sensing images and enhance the features of tiny water bodies. In addition, a multi-scale input strategy was designed to reduce the training cost. We input the processed data into a new water body extraction method based on strip pooling for remote sensing images, which is an improvement of DeepLabv3+. Strip pooling was introduced in the DeepLabv3+ network to better extract water bodies with a discrete distribution at long distances using different strip kernels. The experiments and tests show that the proposed method can improve the accuracy of water body extraction and is effective in fine water body extraction. Compared with seven other traditional remote sensing water body extraction methods and deep learning semantic segmentation methods, the prediction accuracy of the proposed method reaches 94.72%. In summary, the proposed method performs water body extraction better than existing methods.