SegFormer-PAS Model Extraction of Arable Land in Pengzhou Area of Chengdu Plain
The improved SegFormer-PAS model is used to extract the cultivated land in Pengzhou of Chengdu Plain by using the 1 m resolution data fused by Gaofen Ⅱ as the data source.The model selects Mit-B0 as the encoder,and introduces the polarized self-attention(PSA)mechanism to better capture the global context information and solve the problem of insufficient extraction of cultivated land features,and adopts the transpose convolution instead of simple bilinear interpolation in the up-sampling part of the encoder to reduce the loss of spatial details.The experimental results show that the intersection-to-union(IoU),recall,accuracy and F1 score of the SegFormer-PAS model in the experimental area are 90.18％,91.86％,90.86％and 91.26％,respectively,which are improved compared with that of the benchmark model SegFormer-B0,and the results of SegFormer-PAS in the task of cropland extraction in Pengzhou of Chengdu Plain are better than the four classical semantic segmentation algorithms,namely SegFormer,U-Net,Unet++and HRNet.
万方数据
维普
