An Object-Oriented Binary Change Detection Method Using Nearest Neighbor Classification

Threshold selection is a critical step in using binary change detection methods. The threshold determines the accuracy of change detection results but is highly subjective and scene-dependent, depending on the familiarity with the study area and the analyst’s skill. Nearest neighbor classification is a nonparametric classifier, which was applied to remove the threshold. In order to find the most suitable feature to detect construction and farmland changes, a variety of single and multiple variables were explored. They were regional similarity (RSIM), brightness difference images (BDIs), multi-band difference images (MDIs), multi-band ratio difference images (MRDIs), a combination of RSIM and BDIs (RSIMBD), a combination of RSIM and a optimum band difference and a optimum band ratio difference (RSIMDR), MDIs and MRDIs multiple variable groups. All were tested for two study sites of the bi-temporal SPOT 5 imagery, the results indicated that RSIM, RSIMDR, RSIMBD were significantly better than other single and multiple variables.