Remote sensing image segmentation method based on AD-TransUnet

In order to solve the problems of rich information of ground objects, complex environment, unclear target segmentation and incorrect target classification caused by inconsistent size and uneven light in the shooting process, a remote sensing image segmentation algorithm based on improved TransUnet was proposed. Firstly, the empty convolutional space pyramid pooling is integrated into the feature coding downsampling stage, so that the network can better capture the feature information of different scales, Secondly, Depthwise Separable Convolution is added in the feature decoding and upsampling stages of each layer to reduce the number of model parameters and increase the spatial dimension information and depth dimension information. Finally, the loss function is optimized to make the model obtain better accuracy on the dataset than before. According to the experimental results, the mIoU value and F1-socre value of AD-TransUnet reach 52.3% and 64.1% on the LoveDA dataset, respectively, which are 1.9% and 2.1% higher than those of TransUnet. Then, The results show that the AD-TransUnet model achieves semantic segmentation of remote sensing images with better accuracy.