Research On Sea-land Segmentation Of Remote Sensing Images Based On Deep Learning

Achieving accurate sea-land segmentation is important for China’s sea-land resource survey,ecological environment monitoring,marine biological production estimation,disaster forecasting and assessment,and marine environmental protection.The article aims at achieving pixel-level sealand segmentation of near-shore scenery,and carries out research on sea-land segmentation based on high-definition remote sensing images using high-definition remote sensing satellite images.Aiming at the large amount of semantic information in remote sensing images and the intricate sea and land environment,which lead to the problems of pixel classification error,inaccurate segmentation and easy gradient disappearance or gradient explosion in forward convolution in the traditional sea and land segmentation methods,a sea and land segmentation network based on deep learning is proposed for high resolution remote sensing images.The main research contents are as follows:(1)On the basis of the original high-resolution remote sensing sea-land image dataset,preprocessing and data enhancement work are performed on the dataset from both data and image dimensions by using histogram equalization,data normalization,optical transformation and scale transformation,and then all images are labeled by using the open source software Labelme.(2)Based on the original network U-Net,the down-sampling part was replaced with Res2Net,and the coding structure of the network was reconstructed to form the Res2U-Net network.This network model combines jump links,shortcut connections and residual-like structures,makes full use of contextual information,solves the problem that forward convolution is prone to gradient disappearance or gradient explosion,and effectively improves the feature extraction ability of the model without adding additional computational load.(3)A hybrid attention mechanism BAM module and DF(Dice loss and Focal loss)combined loss function are inserted into the Res2U-Net network to form a new BAM-Res2UNet-DF network applicable to sea-land segmentation of high-resolution remote sensing images.Among them,the BAM module can enhance the attention to abstract features and location information from both channel and space aspects,and the DF combined loss function can improve the convergence speed and overall robustness of the model during training,and the parts are combined with each other to effectively improve the accuracy of sea-land segmentation.The improved BAM-Res2UNet-DF network model achieves an average detection accuracy MIoU of 96.67%,which is 1.28%better than the original U-Net network.The experimental results show that the proposed network can be better applied to the research of high-resolution remote sensing sea-land image segmentation,and has certain reference value for improving the sea-land segmentation accuracy.