Assessing Various Scenarios of Multitemporal Sentinel-2 Imagery, Topographic Data, Texture Features, and Machine Learning Algorithms for Tree Species Identification

Abstract

Accurate information about forests, including the identification of tree species, can be achieved by utilizing combinations of various datasets, analyzed over different temporal scales, and employing advanced classification algorithms. Free Sentinel-2 (S-2) satellite imagery, along with other auxiliary data, can serve as valuable sources of geospatial data. In the present study, several scenarios and time intervals were evaluated for tree species identification. Within each scenario, the data were classified using supervised machine learning algorithms random forest (RF) and gradient tree boosting (GTB). The data combinations targeted four scenarios: S-2 bands (scenario 1); S-2 bands and topographic data (scenario 2); S-2 bands and texture features (scenario 3); S-2 bands, topographic data, and texture features (scenario 4). Each scenario was applied for spring, summer, autumn, and long-term intervals. The identified tree species included spruce, beech, fir, larch, pine, mixed species, and other broadleaf species. The best results in tree species identification were obtained in scenario 4. The findings showed that GTB outperformed RF algorithm, providing overall accuracies (OAs) between 96.40% (long-term, scenario 4) and 95.73% (spring, scenario 4). RF was placed second, reaching OAs ranging from 87.41% (long-term, scenario 4) to 84.02% (summer, scenario 4). The integration of topographic data in combinations led to the largest increase in OAs, reaching up to 24.05% percentage point (GTB, summer, scenario 2) in tree species identification. The contribution of texture features in tree species identification was marginal.