Remote Sensing Image Semantic Segmentation Algorithm Based on Improved PSPNet

Because remote sensing images usually contain a large range of ground material changes and complex environmental background, and have large differences in target scales, the segmentation effect of small targets or details is poor. To address these problems, a bibranched semantic segmentation structure is proposed, where the model uses MobileNetV2 as the backbone network. The main branch extracts multi-scale features through the pyramid pooling module, while the auxiliary branch introduces innovative SAFNet modules to enhance the processing power of complex backgrounds and multi-scale targets more effectively. The SAFNet module generates feature maps at four different scales through multi-level downsampling and optimizes feature expression using the 11 convolution and Squeeze-and-Excitation channel attention mechanisms, dynamically adjusting the importance of each channel. To enhance the ability of the model to identify multiscale targets, bilinear interpolation is used to adjust the feature resolution and introduce learnable fusion weights to weight features at different scales to more effectively fuse multiscale information. Through the comparison experiment with the existing popular semantic segmentation model, the overall accuracy of 98.72%, 62.1% mIoU and 77.13% mPA are obtained on iSAID data set. This method can effectively solve the problem of multi-scale fusion and inaccurate object segmentation in remote sensing images, and is suitable for semantic segmentation tasks of remote sensing images.