Road And Small Building Extraction Methods Based On Semantic Segmentation In Remote Sensing Image

Roads and buildings are important ground objects.Roads and buildings extraction in remote sensing images is a pre-step for many applications such as traffic navigation,municipal planning,and geographic information systems.The roads have the characteristics of slender and tortuous shape,complex connectivity,small proportion and large span,which makes the road extraction task full of challenges.The extraction of large-scale buildings that occupy a relatively high proportion has been studied thoroughly,but these methods are not applicable to scattered small buildings.In recent years,convolutional neural networks have been widely used in object detection,image segmentation and other fields,and are superior to traditional methods in performance and efficiency.In this paper,CNN-based semantic segmentation techniques are used to conduct research on roads and scattered small building extraction methods in high resolution remote sensing images.The main work is as follows:The key technologies related to this paper are summarized,including convolutional networks for image classification such as AlexNex,VGG,GoogLeNet,ResNet,and CNN-based semantic segmentation techniques,which mainly include variants of various encoders and context knowledge integration methods such as conditional random fields,multi-scale prediction and feature fusion.To tackle the long-span and connectivity problems of roads,the LinkNet variant(D)-(NL)-LinkNet with dilated convolution and non-local operations is proposed.Based on the study of the best choices of the optimal insertion position and pairwise function for non-local operations,the dilated convolutional and non-local operation modules are inserted into LinkNet.And according to their respective open or close states,three new networks are generated.These networks are trained on training data from a road extraction challenge,and test results obtained by online submission prove the effectiveness of the proposed methods.It is very difficult for conventional methods to extract small buildings on high-resolution satellite imagery,thus a new fully convolutional network ZF-FCN is proposed.ZF-FCN uses a small receptive field to obtain more local information,uses less max pooling operations to avoid violent down-sampling,and uses Lovász-Softmax loss to solve sample imbalance problem and better optimizes the IoU metric.A dataset is established in this paper,and the experiments are carried out on the augmented data with different cropping sizes,which prove that ZF-FCN outperforms both FCN and U-Net.