A Weakly-Supervised Change Detection Technique for SAR Images Based on Deep Learning and Synthetic Training Data Generated by an Ensemble of Self-Organizing Maps

This paper proposes a novel approach to improve accuracy of weakly-supervised change detection in Synthetic Aperture Radar (SAR) imagery. The method is based on the idea to use an initial small size labeled dataset for synthetic training data generation (STDG) using an Ensemble of Self-Organizing Maps (ESOM). The resulting synthetic data set substitutes the initial small set of labeled authentic data and it is used to train a simple deep neural network (DNN) classifier. The proposed DNN architecture consists of 6 layers: an input layer, two Long Short-Term Memory (LSTM) layers, and three Fully Connected (FC) layers. The proposed method is evaluated on a TerraSAR-X image acquired in the Fukushima region, Japan, before and after tsunami. We have used only 200 labeled data (100 for the "change" class and 100 for the "no-change" class). As benchmark method, we have considered a Support Vector Machine (SVM) classifier. The experimental results confirmed the effectiveness of the proposed approach.