Automatic Change Detection in High-Resolution Remote Sensing Images by Using a Multiple Classifier System and Spectral–Spatial Features

Abstract

Change detection (CD) is an active research topic in remote sensing applications including urban studies, disaster assessment, and deforestation monitoring. In this paper, we propose an automatic method for CD in high-resolution remote sensing images that uses a novel strategy for the selection of training samples and an ensemble of multiple classifiers. As for the selection of training samples, our proposed method uses two groups of thresholds instead of just one threshold to enhance the quality of the selected training samples by allowing for their selection in an intelligent manner. In order to achieve high CD accuracy, spatial information such as texture and morphological profiles are utilized in conjunction with spectral information. Our multiple classifier system (MCS) exploits the extreme learning machine (ELM), multinomial logistic regression (MLR), and K-nearest neighbor (KNN) classifiers. To validate our newly proposed approach, we conduct experiments using multispectral images collected by ZY-3. The proposed method provides state-of-the-art CD accuracies as compared with other approaches widely used in the literature for CD purposes.