Classification of Landsat images based on spectral and topographic variables for land cover change in Zagros forests

Background/Question/Methods

When classifying remote sensing images, the selection of appropriate prediction variables and classification algorithm should consider the specific physical and ecological characteristics of the study areas. When working with time series data there is a need to select a unified classifier which is independent from the sensor used. Images from the Landsat MSS sensor have resulted in poor classification outputs as compared to TM and ETM sensors, due to the lower spatial and spectral resolutions. This study focused on two objectives: (1) select the best predictor variables for the classification of Zagros forests given the characteristics of the study area, available variables, and vegetation indices; and (2) evaluate the application of the Random Forest (RF) algorithm as a unified technique for the classification of datasets acquired from different sensors. Three images of the same study area were acquired from the Landsat 5 TM sensor in 2009, the Landsat 7 SLC-on ETM sensor in 1999, and the Landsat 2 MSS sensor in 1975. The images were corrected geometrically and radiometrically, and the RF algorithm was applied for variable selection and classification. A test of equivalence was used to compare the overall accuracy of the classified maps from the three sensors.

Results/Conclusions

High overall classification accuracies were achieved for all three images (97.90 % for MSS, 95.43% for TM, and 95.29 for ETM).The variable selection procedure identified and used 16, 19, and 10 predictor variables for the classification of MSS, ETM, and TM images, respectively. Slope, Normalized Difference Vegetation Index (NDVI), and elevation were determined as the most important predictor variables for all three images.  The ETM and TM derived maps had equivalent overall accuracy, but significantly higher overall accuracy was obtained for the MSS derived map. The selected predictor variables were consistent with ecological reality. The high and equivalent classification accuracies obtained from different years makes the RF algorithm suitable for the classification of Landsat datasets for the studies of land cover change in Zagros semi-arid forests.