Evaluating the performance of change indices extracted from multi-temporal remote sensing images in detecting land use and land cover changes
In this paper, the performance of 8 change indices including Euclidean Distance (ED), Spectral Angle Mapper (SAM), Spectral Correlation Mapper (SCM), image regression, ERGAS, spectral-spatial correlation, mutual information (MI), and Jeffries-Matusita Distance (JMD) has been compared on two different datasets from the accuracy and computational time points of view. The first dataset includes a pair of bi-temporal images taken by Landsat TM5 and ETM+ sensors over the southern shores of Lake Urmia, and the second dataset is taken by Landsat TM4 and TM5 sensors overs the Maragheh city and sorrounding area. Implementing the mentioned indices on the first dataset indicates the SAM's significant superiority compared to the other indices. False alarm (FA), missed error (ME), and total error (TE) of the change map resulting from SAM are 3.40%, 13.91%, and 8.86%, respectively. The change map resulting from SCM is in the second order, its FA, ME, and TE values are almost twice the corresponding values derived from SAM. JMD, regression, MI, ED, and ERGAS indices were in the next ranks respectively with 20.17%, 20.61%, 20.84%, 21.22%, and 21.47% TE on their change maps. In the first dataset, the worst change detection result was obtained from the correlation index (TE=27.80%). In the second dataset, the best results have been obtained first from the ERGAS and then from the magnitude of change, which were at the top compared to others. The FA, ME, and TE values of the change map resulting from the ERGAS were 0.63%, 26.54%, and 7.5%, respectively, and the FA, ME, and TE values of the change map resulting from the ED were 0.63%, 32.23%, and 9.01%, respectively. In the lower ranks, SCM, SAM, regression, spectral-spatial correlation, and JMD have led to 11.41%, 12.62%, 14.45%, 17.34%, and 18.03% total errors in change detection, respectively. The worst result with 26.56% TE was achieved from the MI. It is noted should be that the significance of the differences in accuracy between the mentioned indices was tested and verified by McNemar’s test. In terms of computation time, the ED was the most efficient, while the MI was time-consuming on the analysed datasets.