An End-to-End Deep Learning Change Detection Framework for Remote Sensing Images

Abstract

In view of the lack of overall design and research for remote sensing image deep learning change detection, a deep learning driven “data input - network design - model training - test time augmentation” end-to-end remote sensing image change detection framework is proposed. Firstly, change detection dataset is built from a pair of remote sensing images which cover the same location at different time. Secondly, while considering the change detection as a binary segmentation problem, the deep learning network for change detection is designed. Thirdly, the hyper-parameters are adjusted and the network is trained to get a change detection model. Finally, the blocks are predicted through change detection model, and the block results are merged to get the whole image. The experiment results of the new building detection show that the framework regards the change detection as a global problem, and can detect the change of the whole scene image directly, overcome the boundary error caused by the traditional block, and reduce the trivial process of general steps. The rate of false detection is low, but there are still some missed detections, mainly due to the imbalance of positive and negative samples, which has a negative effect on model training. It is necessary to refine the samples and optimize the model.