Siamese Network With Multi-Level Features for Patch-Based Change Detection in Satellite Imagery

We present a patch-based Siamese neural network for detecting structural changes in satellite imagery. The two channels of our Siamese network are based on the VGG16 architecture with shared weights and are used as feature extractors. Changes between the target and reference images are detected with a fully connected decision network trained on a large dataset of DIRSIG image chips. We experiment with features from different levels of the network to evaluate their combined effect on detection performance. We further incorporate bootstrapping in the training process to improve the network's ability to classify difficult samples. Our results show that our method achieved very good results on change detection accuracy that were best when combining features from two layers.