Research On High-resolution Remote Sensing Image Building Change Detection Method

With the continuous development of remote sensing satellite technology,the spatial resolution and temporal resolution of remote sensing images used by people's livelihood have been greatly improved.Remote sensing change detection is through multiple observations of the same area to obtain changes in the area during the observation interval.Change detection has always been one of the most important studies in the field of remote sensing,and it has great influence in various fields.With the continuous advancement of urbanization,urban management is becoming more and more important.The timely detection of illegal buildings is one of the important tasks of city managers;natural disasters are a huge blow to humans,and China has also established a special disaster emergency response department.In order to respond to the raid of disasters and carry out rescue work in time.Remote sensing building change detection provides convenience for the above work.The development of deep learning technology has promoted the research process of computer vision in all directions.This article mainly studies and analyzes the problem of building change detection in highresolution remote sensing images,and classifies the remote sensing scenes to be detected into two categories: 1)The buildings represented by foreign villages are sparsely distributed and simple and regular scenes 2)Complex scenes with dense buildings,uneven distribution,and uneven building heights represented by domestic cities.In simple scenarios,there are problems with incomplete boundaries and irregularities in the prediction results of building changes;in complex scenarios,there are still a lot of false detection problems in high-rise buildings in existing methods.The deep convolutional neural network model is used to detect the changes of pixel-level buildings in two phases in two scenes.This article includes the main work in the following three aspects:(1)In order to solve the problem that the current building change detection work fails to combine object-level and pixel-level features in simple scenarios,a dual task constrained deep Siamese Convolutional Network(DTCDSCN)model is designed.DTCDSCN simultaneously executes the change detection task and the building semantic segmentation task,which effectively fuses the object-level semantic feature information and the pixel-level change feature.The prediction result has high pixel accuracy,complete contour of the changing building result and less noise.The effectiveness of DTCDSCN is verified on the WHU building change detection data set.(2)In view of the complex scenes,the difference in high-rise buildings' view angles leads to the problem of excessive pixel deviation between the two images,and a multi-scale constrained convolutional change detection network(MSCM-CDNet)model is designed.DTCDSCN uses the method of multi-scale supervision to force the model to consider larger receptive field features in this scenario,and to promote the fusion of the multi-scale features of the model to reduce the large number of false detection problems caused by the difference in perspective of high-rise buildings.Experiments were conducted on the Sense Time change detection data set and the results were analyzed to verify that MSCM-CDNet effectively reduces false detections.(3)The application of DTCDSCN and MSCM-CDNet to the actual problem of multi-type change detection of buildings further verified their effectiveness.Taking xview2 Challenge: Evaluation of building damage after natural disasters as the actual application scenario,a solution with DTCDSCN and MSCM-CDNet as the core was proposed.Full quantitative and visual analysis was performed on the x BD data set,and finally the fifth place in the public ranking of the competition was achieved.