Building Extraction From High Resolution Remote Sensing Images Based On Semantic Segmentation

With the rapid development of remote sensing imaging technology,more and more high-resolution remote sensing images with better quality have been available for high accuracy mapping.The extraction of specific targets from high-resolution remote sensing images plays an important role in urban development planning,automatic map updating,and land resources surveys.As an important artificial feature,the efficient,accurate,and automatic extraction of building and road has been a critical and challenging problem for a long time.At present,there have been some achievements in building extracts for high-resolution images.These achievements include those technologies based on pre-defined geometric features,as well as those existing deep learning-based techniques.Although these techniques show a relatively high extraction accuracy,but they still have some shortcomings.Firstly,the extraction based on pre-defined features mostly relies on the existing theories and knowledge on human vision,which is less reliable and efficient due to the difficulty of the associated algorithm developments.The extraction methods based on deep learning technology have certain defects in the use of samples in the same region for training tests.The distribution characteristics and architectural styles of buildings in different regions are varying.Training only one region's sampling for a model causes a low generalization ability.Besides,many studies only made an overall performance evaluation for the algorithm,but no specific context comparisons.Based on the above problems,this paper expands the existing semantic segmentation technology on multiple types of samples,and analyzes and improved several problems of building extraction.The main research contents and conclusions of this paper are as follows:(1)Through the data pre-processing of Inria aerial imagery,18000 training sets,3840 verification sets and 360 test sets were obtained.In order to carry out a comprehensive performance evaluation of the models focusing on building extraction,several morphological evaluation indicators are proposed.(2)This paper uses the traditional full convolutional network(FCN-8S)network structure based on VGG-16 for building segmentation,and attempts to configure the network according to the performance of the network under different hyperparameters.To this end,the following four scenarios are defined:roads are confused with buildings;buildings are obscured by trees or in the shadows of other objects;different types of buildings such as residential buildings and commercial buildings;and small buildings.We visually analyzed the testing results at these four scenarios and made a quantified evaluation Through quantitative and quantitative analysis,it is found that the FCN-8S-based building extraction algorithm can successfully identify and extract different types of buildings,including residential buildings and commercial buildings.For the other three scenarios,the extraction effect is not as good as the first scenario.In order to solve the problem of confusion of roads and buildings,we expanded the binary segmentation model by adding a road class to strengthen the ability of clarifying similar targets.The experiment proves that this strategy can solve this problem well.(3)Based on the analysis of the problems in the building extraction using FCN-8s network model,a new solution is proposed.The RefineNet network is used for building extraction,and ResNet-101 is used as the encoder for feature extraction.Through the analysis of the testing results,it is found that the network structure solves the problem of shaded buildings and missed houses better than before.We are also regret that for the scenario of occlusion of trees,the semantic segmentation network can only randomly succeed on the complete extraction of a building shape.The research shows that the building extraction using the semantic segmentation algorithm proposed in this paper solves the following issues to a certain extent:1,confusion between building and road,2 buildings are shaded by the nearby objects,3 small buildings are missed,4 residential buildings and commercial buildings are blurring Although there is still a way from the perfect extraction,but it is deserved to make effort,and it is only the time to make success.