EVALUATION OF FEATURE SELECTION METHODS FOR VEGETATION MAPPING USING MULTITEMPORAL SENTINEL IMAGERY

With the recent advances in remote sensing technologies for Earth observation (EO), many different remote sensors (e.g., optical, radar) collect data with distinctive properties. EO data have been employed to monitor croplands and forested areas, oceans and seas, urban settlements, and natural hazards. The spectral, spatial, and temporal resolutions of remote sensors have been continuously improving, making geospatial monitoring more accurate and comprehensive than ever before. To tackle this issue, various variable selection methods (e.g., filter, wrapper, and embedded methods) are nowadays used to reduce data complexity, and hence improve classification accuracy. Therefore, the goal of this research was twofold. Firstly, to assess the performance of the random forest (RF) classifier in a large heterogeneous landscape with diverse land-cover categories using multi-seasonal Sentinel imagery (i.e., Sentinel-1; S1 and Sentinel-2; S2) and ancillary data. Secondly, to compare RF variable selection methods to identify a subset of predictor variables that will be included in a final, simpler model. Using mean decrease accuracy (MDA) as a feature selection (FS) method, an original dataset was reduced from 114 to 34 input features, and its classification performance outperformed all-feature (114 features) and band-only (36 features) model with an OA of 90.91%. The most pertinent input features for vegetation mapping were S2 spectral bands (14 features), followed by the spectral indices derived from S2, texture features, and S1 bands. This research improved vegetation mapping by integrating radar and optical imagery, especially after applying FS methods which removed redundant and noisy features from the original dataset. Future research should address additional feature selection methods (i.e., filter, wrapper, or the embedded) for vegetation mapping, combined with advanced deep learning methods.