WRBSNet: A Novel Sea-Land Segmentation Network With a Wider Range of Batch Sizes

Based on the high-resolution remote sensing images and deep learning models, more and more detailed spatial information around the coastlines can be extracted automatically nowadays. As so far, most of the sea-land segmentation networks choose batch normalization (BN) as the normalization layer. However, the accuracy of segmentation results is easily affected by the batch size, and the error increases markedly when the batch size is small. Aiming at this problem, this letter proposes a new sea-land segmentation network wider range of batch sizes (WRBSNet) using the XBNBlock(GN)_P2 module. We also construct a high-resolution sea-land segmentation dataset to evaluate the performance of the WRBSNet under different batch sizes. Experiments show that the WRBSNet has a wider available range of batch sizes and also has a better performance. When the batch size is 2, the aAcc, mIoU, and mAcc of WRBSNet's optimal segmentation results can reach more than 0.89, and when the batch size is 4, 8, and 16, the corresponding three metrics can reach 0.97.