Change Detection In Urban Area Based On Siamese Convolutional Neural Network With High-Spatial-Resolution Remote-Sensing Images

Change detection is the process of finding the changes of the earth’s surface by comparing two or more remote sensing images acquired at different times in the same geographic area.Traditional change detection methods require manual design features,which is a time-consuming and laborious task and requires strong professional knowledge.And it is difficult to design a universal feature that applies to all types of categories.The multi-layer nonlinear mapping of the deep neural network makes it capable of fitting any function,so it can construct a high dimension classification surface and complete the task of pattern recognition with high quality.This paper proposes a method for end-to-end change detection in urban area based on a deep neural network,which can avoid the process of manually designing features and improve the accuracy of change detection.This paper mainly includes the following three aspects:(1)Designed a Siamese convolutional neural network to make change detection of two-phase high spatial resolution remote-sensing in urban area.SCNN consists of two branch networks and one decision network,because that the change detection needs to process the two-phase images at the same time.The two branches in the bottom network extract features of two-phase images respectively.Branch outputs are concatenated and given to the top decision network.Decision network plays the role of similarity measure of two-phase images and outputs change detection result.In this paper,the image segmentation block is used as the basic unit of change detection,in which way,deep neural network can fully extract high-level feature of remote sensing imagery.At the same time,it can realize "full coverage" change detection.The overall accuracy of urban area change detection based on SCNN in Wuhan has reached more than 88.57%.(2)Proposed data augmentation techniques especially for change detection.In the change detection,there are some differences between images acquired at different times due to the different acquisition conditions.The data augmentation techniques for image classification,such as rotation,translation,and noise,are not applicable to change detection.For this problem,this paper proposes different data augmentation methods for changed and unchanged samples,respectively.For unchanged samples,the approach of "iterative training SCNN-sample selection-sample expansion" are proposed.For changed samples,two temporal images at different positions are combined into new changed sample pairs.These methods can expand the dataset and avoid overfitting effectively.(3)Proposed a change detection method based on model-transfer to identify the changes in other cities.There are some differences between cities due to the different architectural style and city planning,thus a model trained in a city cannot be directly used in another city.Based on conservative training and layer transfer strategies,this paper migrates the SCNN model trained in Wuhan City to Xianning City to identify the changes.The overall accuracy of the change detection in Xianning City based on conservative training and layer transfer reach more than 83.07%and 86.98%,respectively.