Object-based classification of vegetation species in a subtropical wetland using Sentinel-1 and Sentinel-2A images

Abstract

The classification of vegetation species is a fundamental technical task, necessary for the sustainable management of wetland ecosystems. GEographic-Object-Based Image Analysis (GEOBIA) and data mining have enabled classification and monitoring of wetlands with higher accuracies and lower costs. The objective of this study is to evaluate the performance of Random Forest (RF) and k-Nearest Neighbor (k-NN) data mining methods in vegetation species object-based classification in a subtropical wetland, integrating Sentinel-1 and Sentinel-2A images. In this work, 91.3% accuracy was reached for the object-based classification of vegetation in wetlands with the RF method, and 81.1% accuracy was reached with the k-NN method. Synthetic aperture radar (SAR) features obtained the two major importances, 9.7% (VHM) and 8.9% (VVM). The optical features, red edge and the two short-wavelength infrared bands resulted in values greater than 6%. We conclude that the integration of optical satellite images and SAR, together with the use of GEOBIA and data mining, was successful in classifying vegetation classes of wetlands.