Abstract
Differential Interferometric Synthetic Aperture Radar (DInSAR), coherence, phase and 2 displacement are derived from processing SAR images to monitor geological phenomena and 3 urban change. Previously Sentinel-1 SAR data combined with Sentinel-2 optical imagery has 4 improved classification accuracy in various domains. However, the fusing of Sentinel-1 DInSAR 5 processed imagery with Sentinel-2 optical imagery has not been thoroughly investigated. Thus, we 6 explore this fusion in urban change detection by creating a verified balanced binary classification 7 dataset comprising 1440 blobs. Machine learning models using feature descriptors and non-deep 8 learning classifiers, including a two-layer Convolutional Neural Network (ConvNet2), are used 9 as baselines. Transfer Learning by Feature Extraction (TLFE) using various pre-trained models, 10 deep learning from random initialisation and Transfer Learning by Fine-Tuning (TLFT) are all 11 evaluated. We introduce a Feature Space Ensemble family (FeatSpaceEnsNet), an Average Ensemble 12 family (AvgEnsNet) and a Hybrid Ensemble family (HybridEnsNet) of TLFE Neural Networks. The 13 FeatSpaceEnsNets combine TLFE features directly in the feature space using Logistic Regression. 14 AvgEnsNets combine TLFEs at the decision level by aggregation. HybridEnsNets are a combination 15 of FeatSpaceEnsNets and AvgEnsNets. Several FeatSpaceEnsNets, AvgEnsNets and HybridEnsNets, 16 comprising a heterogeneous mixture of different depth and architecture models, are defined and 17 evaluated. We show that, in general, TLFE outperforms both TLFT and classic deep learning for 18 the small dataset used and that larger ensembles of TLFE models do not always improve accuracy. 19 The best performing ensemble is an AvgEnsNet (84.862%) comprised of a ResNet50, ResNeXt50 and 20 EfficientNet B4. This was matched by a similarly composed FeatSpaceEnsNet with an F1 score of 21 0.001 and variance of 0.266 less. The best performing HybridEnsNet had an accuracy of 84.775%. 22 All of the ensembles evaluated outperform the best performing single model, ResNet50 with TLFE 23 (83.751%), except for AvgEnsNet 3, AvgEnsNet 6 and FeatSpaceEnsNet 5. Five of the seven similarly 24 composed FeatSpaceEnsNets outperform the corresponding AvgEnsNet.