To address the problem that existing remote sensing images are diverse and contain a large number of complementary features between different satellite remote sensing images,but existing methods usually cannot fully utilize multiple features,in this paper,we adopt the Massachusetts road dataset,manually screen the samples,ex-pand the features using the SAR amplitude intensity images of Sentinel-1 satellite,and adopt the Unet architecture with Resnet50 as the encoder.Firstly,separate the heterogenous image feature extraction process by two-step opera-tion.Secondly,train the fusion features by iterative decoder.In order to enhance the image matching,we adopt CFOG features for heterogenous remote sensing image alignment,and use Tversky Loss as the loss function of nega-tive balance samples to assist road network extraction to achieve road segmentation.The road extraction results with high accuracy and precision are obtained.The results show that low-resolution SAR images also contain features that high-resolution optical images do not have,and the deep learning method can better fuse the two heterogenous features,improve the accuracy of road segmentation,and have a better effect on noise suppression.The experimental and analytical results of this paper can provide reference for application research in related fields.
NETL
NSTL
万方数据
