TA-TransUNet: An Improved Deep Learning Network Model for Water Body Extraction From Remote Sensing Images

Water body extraction, a critical task in environmental monitoring and resource management, has gained extensive research and application in recent years through the use of remote sensing image semantic segmentation. However, due to the complexity of spectral information of ground objects in remote sensing images, water extraction faces challenges such as inaccurate edge extraction and low extraction accuracy. This article proposed an improved model, namely TA-TransUNet, to address these issues, particularly poor edge recognition and difficulty in identifying water bodies such as lakes and ponds. First, the model uses an encoder-decoder architecture, with a Transformer-based encoder that enhances the ability to capture global features. Second, the Triplet Attention module is introduced into the model to improve the processing and understanding of input data by integrating local, global, and cross-channel attention mechanisms, thereby enhancing the performance of water extraction. At last, the designed module is applied to the up-sampling stage of the TransUNet decoder, enhancing the expressive power of different features by fusing spatial, channel, and dimensional features. By merging up-sampled features with those from the encoder, the model integrates local details and global semantic information, improving edge detection accuracy and distinguishing complex land cover types. To validate the effectiveness of the proposed model, two publicly available datasets, GID and LoveDA, were used to test the performance of the model. Experimental results showed that TA-TransUNet outperforms the state-of-the-art algorithm models in edge detection and spatial detail capture.