Research On Road Extraction Of High-Resolution Remote Sensing Images Based On Deep Learning

Roads are a fundamental component of geographic information and are crucial for intelligent cities,route planning,and vehicle navigation.The old road extraction technique relies on manual intervention to extract features,highlighting poor efficiency and unreliable accuracy,which cannot satisfy the demands of road information modernization.Convolutional neural networks have strong feature learning capabilities and can adaptively extract feature information at different levels,which is conducive to image segmentation,and deep learning encourages the development of this neighborhood.Road information can be automatically and intelligently extracted using the current deep learning segmentation method,although this power is limited in complex backgrounds.Thus,a new segmentation network called ACS-Unet was created to address the issues as mentioned earlier:（1）This research builds on U-Net and creates an ACS-Unet model with high efficiency information extraction to address the issue that the U-Net network model has low efficiency in extracting information against the backdrop of complicated roadways.In order to strengthen the acquisition of the channel’s useful information,suppress the response of irrelevant information,and also introduce the location information of the road,which is helpful for the subsequent decoding block operation,the coordinate attention module is first integrated in the jump connection.At the bottom of the encoder,a spatial hole convolutional pyramid module is added that can extract multi-layer feature data.This module can use various expansion rates to broaden the network’s receptive field,acquire multi-scale context data,and improve the model’s capacity for detailed extraction.The spatial attention module is used to enhance the decoder’s capacity for adaptive learning,boost the learning of various location-specific information,and enhance the model’s resistance to interference.（2）The created ACS-Unet model has been trained and improved.To address the issue of seriously unbalanced positive and negative samples of road and background in remote sensing photos,the segmentation network is optimized using the Diceloss mixed loss function and the transfer learning approach to load the pre-training weights of the model.The batch normalization operation is used to improve the information flow and computational efficiency of the model while avoid the problem of gradient anomalies.The gradient descent and model convergence are accelerated using the Adam optimization technique.（3）This paper studies the remote sensing road dataset,first using ENVI to preprocess the remote sensing data,obtain high-quality remote sensing images through a series of techniques like radiation calibration,atmospheric correction,orthorectification,and fusion,and then manually annotate and save the results from verifying the extraction effect of the model in various scenarios.The Quick Bird road dataset,consisting of 5400 photos,is then obtained after the data is augmented with horizontal flipping,mirroring,color modification,etc.,using MATLAB software.（4）The developed ACS-Unet semantic segmentation model is then used to extract road information from high-resolution remote sensing images.The successful segmentation experiments are conducted on the Massachusetts,Deep Globe,and Quick Bird datasets,respectively.The results show that the improved model outperforms FCN,U-Net,Seg Net,and Dense UNet in four evaluation indicators:Precision,Recall,F₁-score,and Io U.ACS-Unet can extract more full road conditions from complex remote sensing photos,which has strong practicability in varied road extraction tasks.