“The automatic extraction of roads from remote sensing images has broad application prospects in fields such as smart cities, smart transportation, and autonomous driving. However, the automatic extraction of roads from high-resolution remote sensing images still faces challenges such as fragmentation and poor connectivity, making it difficult to extract complete roads. To this end, relevant experts have proposed an improved encoder decoder network SAMSNet (Split Attention and Multi Scale Attention Network). This network uses Split Attention Network (ResNeSt-50) as the encoder to extract semantic information of images across channels to achieve high-quality feature representation; Introducing cascaded parallel dilated convolution blocks to expand the receptive field and enhance the network's ability to perceive multi-scale contextual information; Introducing the Multi Scale Channel Attention Module (MS-CAM) in the skip connection section, while focusing on both global and local road information, to assist the network in identifying and detecting roads under extreme scale changes. And experimental validation was conducted on the DeepGlobe Road dataset, Massachusetts Road dataset, and GRSet dataset, comparing SAMSNet with nine other mainstream models. The verification results show that SAMSNet outperforms other comparison models in multiple evaluation metrics such as IoU and F1 score on three public datasets, achieving the best extraction results.”