FOREST COVER MAPPING AND PINUS SPECIES CLASSIFICATION USING VERY HIGH-RESOLUTION SATELLITE IMAGES AND RANDOM FOREST
Keywords: Worldview-2, Machine Learning, National Forestry Inventory, forest cover, Pinus classification
Abstract. Advances in remote sensing technologies are generating new perspectives concerning the role of and methods used for National Forestry Inventories (NFIs). The increase in computation capabilities over the last several decades and the development of new statistical techniques have allowed for the automation of forest resource map generation through image analysis techniques and machine learning algorithms. The availability of large-scale data and the high temporal resolution that satellite platforms provide mean that it is possible to obtain updated information about forest resources at the stand level, thus increasing the quality of the spatial information. However, photointerpretation of satellite and aerial images is still the most common way that remote sensing information is used for NFIs or forest management purposes. This study describes a methodology that automatically maps the main forest covers in Galicia (Eucalyptus spp., conifers and broadleaves) using Worldview-2 and the random forest classifier. Furthermore, the method also evaluates the separate mapping of the three most abundant Pinus tree species in Galicia (Pinus pinaster, Pinus radiata and Pinus sylvestris). According to the results, Worldview-2 multispectral images allow for the efficient differentiation between the main forest classes that are present in Galicia with a very high degree of accuracy (91%) and ample spatial detail. Pinus species can also be efficiently differentiated (83%).