Study On High Resolution Remote Sensing Image Change Detection Method Based On Cascaded U-Net And Domain Adaptation

Change detection of remote sensing images is an important task in remote sensing image processing.Change detection detects changes in ground objects from remote sensing images taken at different times of the same location.Change detection has a wide range of applications in urban planning,land use,environmental protection,disaster assessment and other fields.With the continuous development of earth observation technology,sources of remote sensing image have become more diverse,and remote sensing data from different distributions have increased the difficulty and challenge of change detection.This thesis mainly studies the change detection of optical remote sensing images,that is,the binary change detection task for remote sensing image pairs taken at different times from the same location.With the development of deep learning techniques,change detection models have significantly improved their detection performance.This thesis summarizes and analyzes several existing change detection models and compares them through experiments.Most existing change detection models use network structures similar to U-Net.However,these change detection networks struggle to accurately segment the edges of changed objects and need improvement in detecting small and complex-shaped objects.Additionally,for detecting larger objects,a single U-Net model may have insufficient receptive field.Furthermore,the inconsistency in data fields due to different data sources,lighting conditions,and seasons causes a large domain gap between two temporal images in change detection datasets,affecting the detection performance of change detection models.And current domain adaptation methods in change detection do not fully utilize the invariant information contained in bi-temporal images which are captured at the same location.Moreover,existing domain adaptation methods lack flexibility and can only adapt from one fixed domain to another,requiring manual domain-based image division in the dataset.To address these two issues,this thesis focuses on the following two parts:1)To tackle the problems of existing change detection models,a cascaded U-Netbased change detection method is proposed.The method uses cascaded U-Net models to continuously optimize extracted features to produce better detection results.The method also utilizes patch embedding to allow the model to cascade more U-Net structures at a lower computational cost.Additionally,a training strategy is proposed to adjust the weight of different U-Net outputs during training to further improve detection accuracy.Through experiments on two datasets and quantitative and qualitative analyses of experimental results,the effectiveness of the proposed change detection method is verified.2)To address the issue of large domain gap between two temporal remote sensing images in change detection data,a domain exchange network is proposed.The network explicitly extracts domain-specific features and domain-invariant features of remote sensing images,exchanges the domain-specific features,and finally generates the exchanged image by combining both types of features from different images.The proposed network exchanges the domain of a pair of temporal images to generate images for each time point in both domains.Which enables change detection models to detect changes between images captured at different times but in the same domain.The effectiveness of the proposed method is verified through experiments,demonstrating that the proposed domain adaptation can improve the performance of change detection models.