Application of SP-UNet++ in Remote Sensing Image Segmentation

In the field of image segmentation, U-Net and Fully Convolutional Networks (FCNs) have achieved great success, but these models have two limitations, their optimal depth is unknown a priori, and requires extensive schema search or inefficient integration of different depth models; Their hop connections impose an unnecessarily restrictive fusion scheme that forces aggregation only on the same scale feature maps of the encoder and decoder subnetworks. It is more prominent in the field of remote sensing. In order to overcome these two problems, this paper further adds the spatial pyramid pooling module called Spatial Pyramid Pooling U-Net++ (SP-UNet++) based on U-Net++. Automatically learning features at different depths and determining the depth with the best performance can greatly reduce the number of training parameters. Experimental results show that the overall recognition accuracy of SP-UNet++ is significantly improved compared with U-Net.