Variational-AutoEncoder Siamese Network for Change Detection on Satellite Imagery

Capturing changes in certain areas has assumed huge importance due to both man-made and natural phenomenon. Our novel research is aimed at urban classification using an unsupervised Variational Auto-encoder based Siamese network model for change detection on satellite imagery. It works well on Polarimetric Synthetic Aperture Radar (PolSAR) images of Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) as well as optical satellite images. The model encodes the images to get the latent feature vectors, on which the Siamese difference is performed and the resultant latent space is decoded to get the difference image. The metrics which help us determine how well our model predicts the mask are the recall score and the receiver operating characteristic curve. We have archived a recall score of 0.87 and the area under the receiver operating characteristic curve is 0.98