Classification of dominant forest tree species by multi-source very high resolution remote sensing data

In this study we classified dominant forest tree species in a Mediterranean study area that hosted forest stands dominated by seven tree species, both coniferous and broadleaf, plus two mixed formations, for a total of nine classes. Airborne laser scanning data with a point density of 10 pts/m2 and multispectral data (RGB and NIR) with 20 cm spatial resolution were taken using a helicopter. We divided the study area into a grid of quadrats of side 23 m and each quadrat was assigned to a dominant forest tree species class by visual inspection of remote sensing data. For each quadrat multispectral features and point cloud-derived metrics were extracted. For classification purposes, the quadrats were divided into training sites (35%) and test sites (65%). Two supervised classifiers were tested: Random Forest (RF) and k-NN. Several combinations of data sources were used. The accuracy of the supervised classifications was assessed against the visual one. The best accuracy of dominant forest tree species was obtained with RF using all data sources, achieving an overall accuracy of 71%. The overall accuracy increased up to 83% when forest categories (coniferous, broadleaf and mixed forest) instead of forest tree species were considered.