Improving Forest Species Mapping Using Sentinel-2 Time Series

Forest management and inventories need to derive forest species information. Generally, this is done by a fastidious manual photointerpretation and an in-field validation. Remote sensing methods allow producing forest mapping at a large extent. This paper focuses on the contribution of Sentinel-2 multispectral time series for mapping fine forest species in order to improve forest inventories. First, a S2 monodate classification is studied with the seasonal effect. Secondly S2 time series were used, with different attribute combinations integrating the NDVI. Results showed a high accuracy improvement with S2 time series (up to +0.14). Two classifiers were tested; support vector machines and Random forests. Kappa reached 0.72 with a high per-class precision ranging between 0.77 and 0.97 for tree species while providing a low precision on maquis areas.