Optimal Land-cover Classification Feature Selection in Arid Areas based on Sentinel-2 Imagery and Spectral Indices

Abstract

Adding spectral indices to Sentinel-2 spectral bands to improve land-cover (LC) classification with limited sample size can affect the accuracy due to the curse of dimensionality. In this study, we compared the performance metrics of Random Forest (RF) classifier with three different combinations of features for land cover classification in an urban arid area. The first combination used the ten Sentinel-2 bands with 10 and 20 m spatial resolution. The second combination consisted of the first combination in addition to five common spectral indices (15 features). The third combination represented the best output of features in terms of performance metrics after applying recursive feature elimination (RFE) for the second combination. The results showed that applying RFE reduced the number of features in combination 2 from 15 to 8 and the average F1-score indicator increased by nearly 8 and 6 percent in comparison with using the other two combinations respectively. The findings of this study confirmed the importance of feature selection in improving LC classification accuracy in arid areas through removing the redundant variable when using limited sample size and using spectral indices with spectral bands, respectively.