An AttSDNet model for multi-scale feature perception enhanced remote sensing classification of coastal salt-marsh wetlands.

Coastal salt-marsh wetlands have important ecological value, and play an important role in coastal blue carbon sink. However, under the influence of various external and natural factors, coastal wetland ecosystems worldwide have severely degraded, leading to biodiversity loss and ecological damage. Based on satellite remote sensing data and deep learning methods, it is an effective means to quickly monitor the spatial distribution of coastal wetlands, which is very important for the protection and restoration of coastal wetlands. The U-Net deep learning framework, because of its low data requirements, fast training speed, and efficient architectural design, has seen rapid development and widespread application in the field of image segmentation. However, applying the classic U-Net architecture to the classification of coastal wetland images, which have rich and complex cover types. It struggles to effectively capture the spatial dependencies and multi-scale feature information present in remote sensing images. To address this issue, this study introduces an enhanced U-Net model that integrates attention mechanisms and multi-scale feature extraction. This model employs stacked dilated convolutions to improve the U-Net's single receptive field limitation, thereby enhancing the model's ability to learn the multi-scale features of typical land covers in complex coastal wetlands. Furthermore, a combined channel-spatial attention mechanism module is incorporated to strengthen the extraction and learning of spectral and spatial features of remote sensing image land covers. This highlights the feature of small-scale land covers that are difficult to capture. Remote sensing image classification was conducted using Sentinel-2 optical imagery on the coastal wetlands of the Yellow River Estuary and Jiaozhou Bay located in Shandong Peninsula, China. An independent test dataset was used to validate the model's accuracy, and comparative experiments were conducted with several existing classification methods. The results show that the proposed model achieved the highest classification accuracy in coastal wetland remote sensing image classification compared to SVM, VGG, FCN, U-Net, ResU-Net, and SDU-Net models. The overall accuracy of the two study areas is 92.73% and 98.69%, and the MIoU is 77.68% and 83.76%, respectively. For different scales of land cover types, such as larger-scale distributions of Tamarix chinensis and ponds, the improved model's MIoU increased by 17.72% and 5.45%, respectively. For elongated structures like artificial roads and tidal channels, the MIoU improved by 9.82% and 5.41%. The proposed method effectively extracts and learns the remote sensing feature information of land cover targets at different scales, enhances the classification accuracy of large-scale land covers, and effectively addresses the issues of detail loss in small target classification and disconnection in linear land cover classification. It provides a more accurate and robust technical method for coastal wetland remote sensing classification, offering a solid data foundation for analyzing the distribution of typical land covers. Additionally, it has significant implications for efficiently monitoring biodiversity and protecting the ecological environment in coastal wetlands.