Semantic Segmentation Of Warping Dams In Remote Sensing Image Based On Deep Learning

Warping dam plays an important role in the control of soil and water loss in the Loess Plateau.In the construction of warping dam and the control of soil and water loss,we need to obtain the location of warping dam and other related information.Recently,with the popularity of big data,the breakthrough of deep learning network structure,and the rapid improvement of computer computing ability,deep learning has once again entered the field of vision of researchers.This paper will build a deep learning model and extract warping dams from high-resolution remote sensing images to provide technical support for the prevention and control of soil erosion.In this paper,Hulu River and Lanni River basins in the Loess Plateau are taken as the research areas,and the warping dam data sets and labels are constructed by using high-resolution remote sensing images in a geographic information system.After comparing five basic models,including FCN,SegNet,U-Net,PSPNet,and DeepLab-V3+,through experiments,U-Net network is selected for improvement and optimization,and DU-Net and DSU-Net are proposed based on U-Net network,and the feasibility of improvement is compared through experiments.The main work of this paper are as follows:(1)The warping dam data set is constructed based on high-resolution google satellite map.Based on DEM,ArcGIS was used to extract the watershed image and cut the grid data of the image and label to a fixed size after labeling the region of interest.We use a sliding window method to expand the data,and construct three kinds of data sets,which are ready for the follow-up research.(2)The performance of FCN,SegNet,U-Net,PSPNet,and DeepLab-V3+models in this semantic segmentation task is compared and analyzed.The overall accuracy of F1 Score and MIoU from U-Net are 95.78%,66.96%,and 72.96%,respectively.The overall accuracy and MIoU of U-Net are slightly higher than DeepLab-V3+,and the training time of DeepLab-V3+is 1.95 times of U-Net,and the comprehensive segmentation effect shows that U-Net is better.Then we study the influence of network input size,batch size and sample size on the U-Net segmentation model,and find that the accuracy is positively correlated with the input size.The increasement of batch size reduces the total training time,and the large sample data set also plays a role in improving the accuracy.Finally,we determine that the input size is 480×480,and the batch size is 4 and use 2304 images for training,in order to carry out the next stage of research.(3)The method of improving U-Net is proposed.In order to solve the problem that some features are lost and the utilization rate of features is not high in the feature extraction phase of U-Net,a new model DU-Net is proposed by using DenseNet169 as a backbone of U-Net without full connection layer.Another model,DSU-Net,is also proposed by adding a SE Block after each density block to further improve the network.The experimental results show that the overall accuracies of the proposed DU-Net and DSU-Net are 0.80%and 0.74%higher than that of U-Net.Meanwhile,the F1 Scores from the DUNet and DSU-Net are increased by 7.20%and 8.52%,and the MIoUs are increased by 4.63%and 5.81%respectively.According to the segmentation results,the DU-Net and DSU-Net are more accurate in the division of noise,shadow and road.In particular,the DSU-Net model shows better performance than other models in warping dam segmentation from high-resolution remote sensing images,which proves the feasibility of our improved method.