An Improved TransUNet Landslide Extraction Method for High-resolution Remote Sensing Images
Remote sensing image landslide identification method is of great significance for emergency command.The TransUNet model has high computational complexity due to the attention mechanism in its architecture,and the skip connection cannot achieve alignment between adjacent feature maps,resulting in long model training time and inability to effectively utilize shallow layer high-resolution feature information.To solve the above problems,the TransUNet model architecture is improved and an improved Flow Alignment TransUNet(FATransUNet)model is proposed.Firstly,the Transformer module in the original structure is replaced by an Efficient Transformer module,which effectively reduces the computational complexity of the model.Secondly,a Flow Alignment Module(FAM)is introduced to replace the up-sampling operation in the original skip connection,feature splicing and decoding stages,which not only simplifies the operation process,but also effectively integrates the high-resolution information in the shallow layer.Experiments based on the open source Bijie landslide dataset show that the F1 score and mIoU of the FATransUNet model reach 91.4％and 91.1％,respectively,which are higher than the accuracy of other five models(FCN,U-Net,SegNet,DeepLabV3+and TransUNet).The interference of complex background on landslide extraction is effectively suppressed,and the extraction accuracy of landslides in high-resolution remote sensing images is improved.
deep learninglandslide extractionFATransUNetBijie landslide data
万方数据
维普
