Sentinel-2 vs. PlanetScope: Comparison and combination for tree species classification in two central European forest ecosystems

Abstract

The increasing rate of species extinction and declining environmental conditions necessitate a comprehensive understanding of habitats, including tree species diversity, which is a critical factor influencing forest ecosystem functions. Traditional methods of acquiring information on tree species, like forest inventories and field-based approaches, are often time-intensive, costly, and impractical for large-scale applications, making remote sensing a feasible alternative. This study compared and combined two multispectral remote sensing datasets, including Sentinel-2 (S2) and PlanetScope (PS), for tree species classification in two Austrian forest ecosystems: the riparian forests of the National Park Donau-Auen (NPDA), where nine tree species were distinguished, and the forests of the Biosphere Reserve Wienerwald (BRWW) where 12 species were investigated. Mono-temporal and multi-temporal data from S2 and PS were analyzed individually and in combination (S2 + PS). A robust reference dataset (835 samples in NPDA and 1283 in BRWW) and a Random Forest algorithm with recursive feature selection were used for classifications. When comparing mono-temporal datasets, S2 consistently outperformed PS, achieving the highest overall accuracies of 63.7 % for NPDA and 70.6 % for BRWW, compared to 58.1 % and 57.4 % with PS. Using multi-temporal S2 data further enhanced classification accuracy, reaching 78.3 % for NPDA and 83.3 % for BRWW, while multi-temporal PS data achieved 74.4 % and 77.7 %, respectively. Combining datasets in NPDA demonstrates an improvement of 1.8 and 5.7 percentage points compared to the sole use of S2 and PS multi-temporal data, respectively. In BRWW, the improvement was 1.3 and 6.9 percentage points. Classification accuracies were higher in BRWW, likely due to its larger reference dataset and the inclusion of more phenologically and morphologically distinct tree species. Overall, this study highlighted the superior performance of S2, particularly in mono-temporal analyses, the added value of combining S2 and PS datasets, and the well-known advantages of using multi-temporal datasets. Notably, the study fairly distinguished between three closely related Poplar species, including Populus alba, Populus × canadensis, and Populus nigra, in riparian forests of NPDA, which is also of great interest from a nature conservation perspective. The outputs of this study can provide helpful information for new satellite missions.