Road Extraction From Remote Sensing Images Integrating Cost Sensitive And Convolutional Neural Network

Road extraction,as an important role in the extraction of remote sensing image objects,can provide semantic information and location information for roads.It is a special semantic segmentation task.The goal is to divide image pixels into two subsets of road objects and background regions.Each pixel in the road object area is assigned a uniform semantic label.Convolutional neural network(CNN)has the characteristics of few parameters,full utilization of local features,and strong feature extraction ability.It is the mainstream method used in road extraction at present.However,the class imbalance problem often appears in the CNN training set,which greatly reduces the performance of the CNN model.This paper adopts the method of integrating costsensitive learning and deep learning to propose an algorithm for the imbalance problem in road extraction from high-resolution remote sensing images.The main work and innovations are as follows:(1)Aiming at the problem of the lack of high-resolution remote sensing image data sets with semantic tags,this paper establishes an imbalanced high-resolution remote sensing image semantic segmentation data set.This dataset is annotated with highresolution remote sensing images downloaded from Bing Maps and combined with OSM vector data.It has a higher resolution of 0.28 meters and contains a large number of semantic labels for roads and buildings,including roads,buildings,and The proportions of the background samples were 79.1%,11.2%,and 9.7%,and the proportions of various types of samples were quite different.(2)In order to maximize the output probability of the true category,the traditional loss function does not consider the output probability that does not belong to the correct category,and assigns the same misclassification cost to different classes.It is impossible to distinguish between easily distinguishable samples and rare samples.The simple samples caused a lot of loss,resulting in the model tends to learn from simple samples,while ignoring rare and difficult samples.Aiming at this problem,this paper designs a cost-sensitive convolutional neural network algorithm with adaptive sample distribution characteristics.The algorithm first adds a balance factor to the sample by counting the sample characteristics of the training data,giving the minority class greater weight.Then use the cost matrix to modify the loss function,assign different misclassification costs to different classes,combine cost-sensitive learning with convolutional neural networks,and make the model cost-sensitive.Finally,in order to verify the effectiveness of the algorithm proposed in this paper,experiments were carried out on the Zimbabwe data set.On the test set,F1 reached 91.1%.Compared with CE and Focal loss,F1 increased by 2.6% and 0.3%,respectively.(3)Aiming at the problems of multi-scale,large intra-class differences and class imbalance in road extraction from high-resolution remote sensing images,this paper designs a cost-sensitive semantic segmentation network Rar Cnet based on road perception.First of all,Rar Cnet uses the Dense Net network in the encoding process to enhance the ability of road feature extraction.At the same time,it uses the feature pyramid network(FPN)structure for multi-scale feature fusion,builds road embedding branches on the backbone network,and inherits multi-scale context features.Solve the multi-scale problem.Then,using the symbiosis relationship between the geospatial scene and the road,the RS relationship module was designed to correlate road-related context information,and to enhance the distinction between road features.Using this symbiosis relationship to suppress background features,thereby reducing the interclass relationship difference.Finally,in order to reduce the imbalance between the road and the background,a cost-sensitive optimization is proposed,which gradually concentrates the network on difficult samples,thereby reducing the gradient contribution of a large number of simple samples in the background,so as to train the road and background balance.In order to verify the effectiveness of Rar Cnet,comparative experiments were conducted on the Mnih data set and Cheng-roads data set.The experimental results show that when Rar Cnet does not use cost-sensitive optimization,the F1 value reaches 75.2% and 83.2%,which is cost-sensitive.During optimization,the F1 value reached 79.7% and 90.0%,an increase of 4.5% and 6.8%respectively.