Deep learning and transfer learning applied to sentinel-1 DInSAR and Sentinel-2 optical satellite imagery for change detection

This paper discusses Deep Learning (DL) and Transfer Learning (TL) state of the art techniques applied to a binary classification task for change detection in satellite imagery. A blob detection algorithm is applied to a Differential Interferometric Synthetic Aperture Radar (DInSAR) generated displacement map. The blobs are classified as either positive, corresponding to uplift or subsidence, or negative, corresponding to noise. The novel dataset consists of Sentinel-1 DInSAR processed georeferenced images of displacement, phase, coherence and RGB Sentinel-2 optical satellite imagery of the blobs. TL via Feature Extraction (FE) is applied using numerous DL models with weights pretrained on the ImageNet dataset to generate feature maps after removing the last predictive layer. A Logistic Regression classifier is then applied to the features. Fine-Tuning (FT) and Random Initialisation (RI), training from scratch, are also applied to ResNet-50 and EfficientNet B4 architectures. The best performing model (85.76%) is the ResNet-50 using FE. Small ensembles of some models are also investigated. An ensemble of ResNet-50, ResNeXt-50 and EfficientNet B4 has an accuracy of 84.83%. TL via FE with the ResNet-50 has an accuracy of approx. 9% and 8% higher than when using it for TL via FT or RI respectively. The EfficientNet B4 obtained an accuracy of 82.29% for FE, 66.35% for FT and 50.00% (as good as a random guess) for RI.