The study utilizes the encoder-decoder structure of the DeepLabV3+ network to extract water information from remote sensing images. The encoder module consists of an improved Xception module and DASPP module, extracting features from the image and aggregating global features. The decoder module recovers features and fuses shallow and deep features to extract water information. The Xception module is enhanced by increasing intermediate flow layers and deep separable convolution layers. The DASPP module is introduced to adapt to changing water bodies and extract complex information effectively. It expands the receptive field and pixel extraction. A feature fusion module is proposed to combine low-level and high-level features for better segmentation. The fusion algorithm compresses features and activates them using Fully Connected layers, enhancing important features and weakening unimportant ones. The fused features are then up-sampled to create the final segmentation image. This approach improves the semantic segmentation effect of water bodies by refining edge contours and enhancing image details. The study demonstrates the effectiveness of the proposed method in extracting water information from remote sensing images.
Source link
Source link: https://www.nature.com/articles/s41598-024-65430-5
High-resolution remote sensing imagery used for water body extraction. #WaterExtraction
