Concurrent Self-Organizing Maps for Supervised/Unsupervised Change Detection in Remote Sensing Images

This paper proposes two approaches to change detection in bitemporal remote sensing images based on concurrent self-organizing maps (CSOM) neural classifier. The first one performs change detection in a supervised way, whereas the second performs change detection in an unsupervised way. The supervised approach is based on two steps: 1) concatenation (CON); and 2) CSOM classification. CSOM classifier uses two SOM modules: 1) one associated to the class of change; and 2) the other to the class of no-change for the generation of the training set. The unsupervised change detection approach is based on four steps: 1) image comparison (IC), consisting of either computation of difference image (DI) for passive sensors or computation of log-ratio image (LRI) for active sensors; 2) unsupervised selection of the pseudotraining sample set (USPS); 3) concatenation (CON); and 4) CSOM classification. The proposed approaches are evaluated using two datasets. First dataset is a LANDSAT-5 TM bitemporal image over Mexico area taken before and after two wildfires, and the second one is a TerraSAR-X image acquired in the Fukushima region, Japan, before and after tsunami. Experimental results confirm the effectiveness of the proposed approaches.