Mapping tree species composition in a Caspian temperate mixed forest based on spectral-temporal metrics and machine learning

The tree species composition (TSC) reflects a forest's tree species diversity and is relevant for forest planning, biodiversity conservation, and forest resources management. Yet, accurate information on tree species composition at landscape scale is largely missing, especially for mixed forests and remote areas. One reason being that mapping tree species is time-consuming, and costly, especially in mixed forests and remote areas. Here we develop a robust method for mapping TSC in a mixed temperate forest. Based on forest inventory plots and considering the frequency of dominant tree species in the inventory dataset, five species groups were defined: pure oriental beech, mixed oriental beech, pure common hornbeam, mixed common hornbeam, and mixed deciduous. The classification is based on three-year time series data of Landsat-8 (L8) and Sentinel-2 (S2) derived spectral-temporal features (STMs) and vegetation indices within the long-term, seasonal, and monthly time scales. Model performances of three Machine Learning (ML) algorithms, Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Tree (CART) were compared and revealed different classification accuracies (overall accuracies (OAs) between ∼ 70 % and 86 %). Highest OA was obtained using SVM regardless of the classification dataset (STMs and satellite time series). The comparisons between different time scales indicated that with both L8 and S2 time series the seasonal STMs produced higher accuracies than monthly and long-term STMs with S2 outperforming L8 across all time scales and with all tested ML algorithms. We conclude that the freely available satellite time series, spectral-temporal features, and ML algorithms are favourable for accurate TSC mapping