Mapping species at an individual-tree scale in a temperate forest, using sentinel-2 images, airborne laser scanning data, and random forest classification

Detailed information about tree species composition is critical to forest managers and ecologists. In this study, we used Sentinel-2 imagery in combination with a canopy height model (CHM) derived from airborne laser scanning (ALS) to map individual tree crowns and identify them to species level. Our study area covered 140 km2 of a mainly mixed temperate forest in the Veluwe area in The Netherlands. Ground truth data on tree species were acquired for 2460 trees. Tree crowns were automatically delineated from the CHM model. We identified the delineated tree crowns to species and phylum level (angiosperm vs. gymnosperm) using a random forest (RF) classification. The RF model used multitemporal spectral variables from Sentinel-2 and crown structural variables from the CHM and was validated using an independent dataset. Different combinations of variables were tested. After feature reduction from 25 to 15 features, the RF model identified tree crowns with an overall accuracy of 78.5% (Kappa value 0.75) for tree species and 84.5% (Kappa value 0.73) for tree phyla whilst using the combination of all variables. Adding crown structural and multitemporal spectral information improved the RF classification compared to using only a Sentinel image from one season as input data. The producer’s accuracies varied between 43.8% for Norway spruce (Picea abies) to 95.3% for Douglas fir (Pseudotsuga menziesii). The RF model was extrapolated to generate a tree species map over a study area (140 km2 ). The map showed high abundances of common oak (Quercus robur; 35.5%) and Scots pine (Pinus sylvestris; 22.8%) and low abundances of Norway spruce (Picea abies; 1.7%) and Douglas fir (Pseudotsuga menziesii; 2.8%). Our results indicate a high potential for individual tree classification based on Sentinel-2 imagery and automatically derived tree crowns from canopy height models.