DETAILED MAPPING OF RESIDENTIAL LAND USE IN QUEZON CITY USING SENTINEL-2 IMAGERY: AN ANALYSIS OF PIXEL-BASED IMAGE CLASSIFICATION USING SUPPORT VECTOR MACHINE

Abstract. Through the years, several studies have attempted to map human settlements using very-high-resolution (VHR) imagery and proprietary software. However, with limitations especially in terms of cost, researchers are now taking the open-source route. For this study, the researchers aimed to perform multi-level classification to delineate residential land use using Sentinel-2 processed with Orfeo Toolbox. The performance of pixel-based approaches with support vector machine (SVM), was applied to different multi-band combinations and varying SVM kernel types. Three sets of information were used – the spectral bands, normalized difference indices, and grey-level co-occurrence matrix (GLCM) measures. In the general land cover classification, except for models with sigmoid kernel, the outputs yielded overall accuracies (OAs) of at least 90%, with special bands and indices raster inputs. The linear kernel performed the best, yielding 93.17% overall accuracy. During the residential versus non-residential built-up cover classification, GLCM measures were added to the set of inputs. The RBF kernel worked best with an OA of 81.56%. The addition of GLCM improved the results, as compared with models with no textural measures. For the residential land use classification, the combination of spectral bands and GLCM, worked best for the pixel-based method, with the linear classifier obtaining the highest OA of 78.24%.