Industrial change detection using deep learning applied to DInSAR and optical satellite imagery

Abstract

In order to detect industrial change in satellite imagery binary classification is investigated using traditional ML with feature extractors (HOG and LBP) with SVM, a simple CNN with only 2 convolutional layers called Simple ConvNet, state-of-the-art Deep Learning (DL) and Transfer Learning (TL) methods. Differential Interferometric Synthetic Aperture Radar (DInSAR) is used to generate interferograms and Terrain Corrected Displacement maps. A Sentinel-1B Synthetic Aperture Radar (SAR) image data stack from 28th Nov. 2017 to 5th Dec. 2018 is used with the images 12 days apart, where imagery was available. Blobs are detected in a displacement map which is generated using DInSAR and the Laplacian of Gaussian algorithm. The blobs are qualitatively verified using optical images from the Sentinel-2 satellite. If subsidence or uplift has indeed occurred then the blob is classified as positive. However, if uplift or subsidence has not occured then the blob is classified as negative which refers to noise. The blob detection algorithm has a high false positive rate. However, true positive blobs are detected corresponding to quarries, mines, construction, etc. A novel dataset is developed comprising of DInSAR processed Sentinel-1 displacement, coherence and phase georeferenced imagery and the corresponding Sentinel-2 RGB optical satellite images of the blobs. A variety of DL architectures that are pretrained on ImageNet, a computer vision performance benchmark dataset, are used for implementing TL with Feature Extraction (FE). For the EfficientNet B4 and ResNet-50 architectures TL with Fine-Tuning (FT) as well as classic DL i.e. training from scratch, Random Initialisation (RI), are also evaluated. Ensemble performance containing certain architectures is also evaluated. The best performing architecture and method (84.34%) is the ResNet-50 with TL via FE applied. It outperformed all the other models and methods including newer, deeper (due to more data being needed to train deeper networks) and ensemble models from an accuracy perspective. Two of the ensembles evaluated using FE have an accuracy of 84.40% and 84.27% respectively with the same F1 score as ResNet50. It is concluded that a correlation between an increase in model size inferring a lower FE accuracy depends on architecture and holds for ResNet and EfficientNet but not for ResNetV2 architectures. The ResNet50 which has slightly more parameters than the EfficientNet B4 performs better with RI and FT respectively. Most models using TL with FE independent of network size outperformed RI and FT with ResNet50 and EfficientNet B4 respectively on this dataset. Most of the DL and TL models outperformed the traditional ML models except for the EfficientNet B4 models for RI and FT. Simple ConvNet (83.37%) outperformed most of the models except for FE with ResNet101, ResNet50 and the ensemble models. FE with the ResNet50 only outperformed Simple ConvNet by 0.97%. Thus, simple CNNs should not be overlooked for small datasets. Displacement Time-series has been developed for all the pixels in the study area. Although there is some noise upward and downward trends can be seen corresponding to change. A velocity map was developed with uplift of 20cm to subsidence of 60cm over the 1 year period noting that the subsidence sometimes occurred in areas of low coherence where it is not accurate.