Research On Impervious Surface Extraction Of High Resolution Remote Sensing Image Based On Deep Convolutional Neural Network

In view of the intricate information in the impervious surface,the atmosphere and clouds will also affect it during satellite photography,so the identification and extraction of impervious surfaces in remote sensing images are misaligned.The traditional method has higher extraction accuracy,but in the process of recognition,it is manually judged,which significantly reduces the work efficiency.Aiming at the inefficiency in the recognition and extraction of impervious surfaces,this paper proposes a method for extracting impervious surfaces of high-resolution remote sensing images based on deep convolutional neural networks: using the high-resolution Google Earth remote sensing image of Kunming,Yunnan Province as the data source(Resolution 2m),using 14700 slice data as training data and 5041 slices as verification data,select 5 deep convolutional neural networks such as Deeplabv3 model,U-Net model,CS-Net model,PSPNet model,SegNet model The algorithm extracts impervious surface information from high-resolution remote sensing images,and compares and analyzes the results of the five algorithms.According to the analysis results,find the optimal method and optimize the parameters to improve the extraction effect of the impervious surface.Research indicates:(1)Analysis of the results of impervious surface extraction based on five deep convolutional neural network algorithms,including Deeplabv3 model,U-Net model,CS-Net model,PSPNet model,and SegNet model.The results show that the U-Net model has The highest accuracy is 76.80%,the Kappa coefficient is 0.73,and the F1 value of the impervious surface is 0.43.The overall accuracy of impervious surface extraction of Deeplabv3 model,CS-Net model,PSPNet model and SegNet model is71.44% 68.00%,71.00%,and 71.00% are all lower than the U-Net model.Comparing the 5 models,the Deeplabv3 model has the smallest error rate and the largest F1 value is the CS-Net model.(2)Based on the improvement of the U-Net model,the ResNet101 structure is fused,and the parameters are improved based on this.The performance of the modelis compared according to the reference index of the verification data extracted from the impervious surface.It is found that the effect of the ResNet101 model is not obvious.However,after improving the parameters on this basis,the accuracy and F1 values have improved.We can compare the two indicators of the impervious surface extraction effect comparison and the overall accuracy of impervious surface extraction after the parameter improvement of the U-Net neural network model.The parameter-improved U-Net convolutional neural network’s impervious surface extraction effect has been enhanced in some ways.The research on the extraction of the impervious surface information of high-resolution remote sensing images based on five kinds of deep convolutional neural network algorithms can provide a certain reference for related research.(3)Using supervised classification algorithm and unsupervised classification algorithm to extract the impervious surface,compared with the convolutional neural network model U-Net to extract the overall accuracy and Kappa coefficient of the impervious surface,it can be seen that the convolutional neural network model is more general than the traditional BP neural network.The accuracy and Kappa coefficient are high,but compared with other supervised classification and unsupervised classification algorithms,the model needs to be further improved.The research on the extraction of impervious surface information of high-resolution remote sensing images based on five deep convolutional neural network algorithms can provide a certain reference for related research.