Sensitivity of support vector machine to features selection for land cover classification based on sentinel-2 data over Riyadh urban arid area, Saudia Arabia

Abstract

This study aimed to investigate the effect of three different combinations of features on land cover (LC) classification accuracy with support vector machine (SVM) and Sentinel-2 data over Riyadh city as an urban arid area. The three combinations included the original spectral bands, the spectral indices with spectral bands, and the selected features after applying recursive feature elimination (RFE). The results showed that with constant sample size, adding the spectral indices had a negative influence on SVM performance accuracy metrics. On the other hand, applying RFE as a feature selection improved the accuracy of LC by nearly 2% in the overall accuracy index and by 6% in the f1-score index. In addition, the feature selection approach decreased the processing time and the number of features for accurate LC classification by removing irrelevant and redundant features. In conclusion, the study showed the importance of applying feature selection with SVM for producing optimal LC classification in the selected urban arid study area.