Performance of machine learning algorithms for forest species classification using WorldView-3 data in the Southern Alentejo region, Portugal

Abstract

Recent advances in remote sensing technologies and the increased availability of high spatial resolution satellite data allow the acquisition of detailed spatial information. These data have been used for monitoring the Earth's surface, namely monitoring land use land cover, quantifying biomass and carbon, and evaluating the protection and conservation of forest areas. O WorldView-3 is a high spatial resolution satellite (0.50m) with 8 multispectral bands (visible and infrared) which allows obtaining detailed data from the Earth's surface. This study aims to map the forest occupation by specie with two WoldView-3 images, and to evaluate the performance of machine learning classifiers (maximum likelihood, support vector machine and random forest) in two regions of Alentejo, south of Portugal. The main forest species are Quercus suber in one region and Quercus rotundifolia in another. The procedures performed were multiresolution image segmentation and object-oriented classification based on 4 bands (blue, green, red and near infrared). As auxiliary data, vegetation indices (NDVI and SAVI) and principal components were calculated. In the object-oriented classification process, the three classifiers were tested. The support vector machine classifier was the one that presented the best accuracy (kappa and overall accuracy), for both images, allowing to obtain good results in the identification of forest species. In the image dominated by Quercus suber, the values of kappa and overall accuracy were 90% and 95%, and for the image where Quercus rotundifolia predominated, 90% and 96% respectively. The methodology applied to the high spatial resolution satellite data showed very good results in the identification and mapping of main forest species. Higher precision values stand out for the image where the Quercus rotundifolia predominates, where there is less spectral variation, namely fewer land use classes, thus reducing errors between classes that may be spectrally similar.