Supervized Change Detection for SAR Imagery Based on Processing of a Low Size Training Data Set by an Ensemble of Self-Organizing Maps

摘要

This paper presents a new method to improve accuracy of supervised change detection in Synthetic Aperture Radar (SAR) imagery. The model is based on the idea to apply a low size labeled dataset to the input of an Ensemble of Self-Organizing Maps (ESOM) for training data generation (TDG). The resulted synthetic data set produced by ESOM substitutes the initial authentically labeled sample set and it is used to train a supervised change detection classifier. The proposed method is evaluated using a TerraSAR-X image of 400×400 pixels acquired in the Fukushima region, Japan, before and after tsunami. As change detection classifiers we have comparatively considered Support Vector Machine (SVM), Nearest Neighbor (NN), the three-Nearest Neighbors (3-NN), and Likelihood Bayes classifier. The experimental results have confirmed the effectiveness of the proposed approach using only 100 authentic labeled pixels.