Building Change Detection Of Remote Sensing Imagery Based On Multi-Scale Siamese Network

The timely and accurate monitor on earth's surface change allowing human beings to control and utilize natural resources better,and exploring relationship between human and nature.Buildings are man-made places where people manufacture and stay,so they become critical research objects of change detection analysis,and serve as significant reference to studies of human geography and economics.Within the wide observing range,the periodicity of remote sensing imagery makes it become fine data resources of change detection research.Recently,the rise of resolution of remote sensing imagery offering change detection task more possibility and higher demand.Firstly,within equal size,very-high-resolution remote sensing imagery contains more information which increases the amount of data to be handled.Then,nowadays most change detection networks based on deep learning employ high-level features even segmentations of bi-temporal images,ignoring low-level features containing semantic and detailed information.Lastly,the pixel-level accuracy measurement in the present moment cannot meet demands in practice.To solve problems above,after reviewing deep learning methods,we discuss their application on change detection problem.This research mainly contents:(1)In order to lessen the information lost during feature extraction,we design a novel end-to-end multi-scale siamese atrous convolutional neural network,which implementing multi-scale feature fusion.After fusion,the extracted feature becomes more comprehensive so that less changed building would be left out and the segmentation is even better.(2)Applying cost function strategy which is suitable for unbalanced segmentation task,improves the stability and learning ability of the network and lowers the bad impact caused by lack of change samples,e.g.building change detection problem.Two different cost functions of different characters are used,one of them assigns higher weight to changed samples,and the other one plays the role to stabilize the training procedure.Building extraction and change detection use different cost functions to meet their specific need.(3)Apart from pixel-level measurement,we evaluate the results of change detection in object-level and obtaining the number of changed buildings.It is of more practical significance when combing two kinds of measurement,offering more explicit target to our later research.We train and test our network on two open-resource datasets WHU and LEVIRCD.The fact is that our method outperformed other recent methods of the world in both pixel-level and object-level,the Io U metric of the two datasets reaching 78.36 and 75.62,and our method is mostly appropriate for applying building change detection on veryhigh-resolution remote sensing imagery.