Abstract: In order to effectively preprocess the traditional remote sensing image change detection, we proposed a change detection method of building remote sensing image based on random patches and DeeplabV3+. This method builds a DeepLabV3+ semantic segmentation network based on the ResNet50, which is a feature extraction network, crops the random patches of 224 pixels×224 pixels in the image and label them as the network input to train the building extraction network,and then, modify the input layer of the building extraction network to six channels. The two-phase remote sensing images are converted into a 6-channel non-RGB image through matrix operation, which are used for network training and validating the change detection accuracy. In Experiment 1, the Mahalanobis distance classification method was used to detect the change by ENVI5.3 software. In Experiment 2, the improved U-Net network and random patches were used to complete the network training and accuracy verification. Experiment 3 used the training data and verification data of Experiment 2, and used random patches and DeepLabV3+ network to train the change detection network and verify the accuracy.The results of Experiment 1, 2, and 3 show that the average intersection-over-union of this method is 24.43%, 83.14%, and 89.90% respectively, and the boundary matching score is 61.47%, 80.24%, and 96.51% respectively.

Key words: random patches, DeepLabV3+network, semantic segmentation, building change detection