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

• DOI: 10.1371/journal.pone.0272317
• Publication Date: 2022-08-05T17:32:29Z
• Authors: Yuanjiang Luo; Ao Feng; Hongxiang Li; Danyang Li; Xuan Wu; Jie Liao; Chengwu Zhang; Xingqiang Zheng; Haibo Pu
• Keywords: Water body; Water extraction; Pooling

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