Answers are grounded in this paper and cite their sources.
Tree species classification is challenging in the heterogeneous forest-steppe ecosystem. We evaluated whether bi-sensor data fusion (Sentinel-2, PlanetScope SuperDove) and topographic variables improve tree species classification in Selenge Province, northern Mongolia. Random Forest (RF) classified five common tree species—Siberian larch, Manchurian birch, Scots pine, Siberian spruce, and Siberian pine—as well as burned forests and non-forest areas using spectral bands, vegetation indices, and terrain derivatives. Sentinel-2, PlanetScope, and FUSION models were assessed across eight dates (2021–2025) using overall accuracy (OA), Kappa, and F1, all derived from polygon-based 5-fold cross-validation to avoid spatial autocorrelation bias. The FUSION model achieved the highest accuracy in June 2023 (OA = 0.913, Kappa = 0.895) and September 2022 (OA = 0.909, Kappa = 0.891). The Sentinel-2 model performed stably (OA = 0.850–0.905, mean = 0.883, CV = 2.0%), while the FUSION model was slightly higher on average (mean = 0.889) but more variable (CV = 2.4%). McNemar’s test showed FUSION significantly outperformed both single sensors on all dates (p < 0.001), although the practical gain was small and phenology-dependent. The FUSION model showed the closest aggregate area agreement with the national forest inventory (<1% error). FUSION improved accuracy only under specific phenological conditions, whereas Sentinel-2 alone was temporally more stable and equally accurate under full leaf-on conditions, making it a simpler, more robust baseline.
No takes yet. Share an insight, caveat, or question.
Boldbaatar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: