WATER SEGMENTATION METHOD BASED ON IMPROVED CONVOLUTIONAL NEURAL NETWORK

Due to the complex multi-scale characteristics of water in remote sensing images, traditional methods are prone to misjudgment and omission during water extraction. To address this issue, a new network structure that integrates local and global information is proposed. The network designed a residual module with an attention mechanism at the encoder end to capture both global and local information for each positional feature, and employed multipath dilated convolution to achieve multi-scale water feature extraction. To improve segmentation accuracy at water boundaries, a refined attention fusion module was designed at the decoder end of the network. Experimental results show that the network achieves recall, precision, and F1-scores of 95.78%, 94.24% and 93.75%, respectively. Compared with traditional convolutional neural networks, these evaluation metrics are improved by 1.56, 1.72, and 1.62 percentage points, respectively.