Land cover change detection thresholds for Landsat data samples

This paper presents the results of research on common change detection techniques. More specifically it looks into the optimization of threshold values for these investigated change detection techniques: image differencing, normalized image differencing, image ratioing, normalized variance differencing, normalized spectral Euclidean distance and Tasseled Cap parameters difference. The threshold values were optimized for the detection of land cover change/no-change based on the comparison with an existing validated classification of five broad land cover classes. For this study a sample set of 104 image pairs was selected, each of 20 × 20 km, cut from Landsat TM/ETM+ imagery series. An object based approach was applied for the land cover change detection. The results showed that the threshold of normalized variance difference had most stable values across the sample set, however applying optimized thresholds the achieved accuracy was comparable for all tested methods.