Abstract: Mastering the types and distribution of nearshore mariculture is of great significance for the ecological protection of coastal zones. When conventional image semantic segmentation methods are used to extract pond culture and factory culture areas in high-resolution remote sensing images, results are prone to edge blurring and loss of detailed information. In this context, this study proposes a classification and extraction method for nearshore mariculture based on improved U-Net. First, to improve the U-Net model, an atrous channel spatio attention module (ACSAM) that extracts more detailed information is constructed. Second, to address the issue of misidentified patches caused by “same spectrum with different objects” in model detection resluts, a noise filtering module is introduced for postprocessing to further improve classification accuracy. Finally, the feasibility and effectiveness of the method are verified through experiments in a nearshore mariculture area in Xingcheng City, Liaoning province. Experimental results show that the PA, MPA, and mIoU of the proposed method are 93.40%, 92.71%, and 85.74% respectively, outperforming classic network model methods U-Net, U-Net++, and PSPNet. This method can provide a technical support for the monitoring of nearshore pond cultures and factory cultures via remote sensing.