Classification of Dominant Tree Species and Vegetation Types Based on Sentinel-2 Time-Series Multispectral Data in the Eastern Qilian Mountains

Precise mapping of forest vegetation, including both dominant tree species and other vegetation types, is essential for forest management and biodiversity conservation, yet traditional high-resolution imagery is limited in spatial coverage. The high revisit frequency of Sentinel-2 provides dense time-series observa tions that capture phenological differences among species across broad areas. In this study, we compiled all cloud-free Sentinel 2 images from 2023 for the Eastern Qilian Mountains, China. Twelve thematic classes were defined in this study, including six dominant tree taxa (five tree species and one Picea genus group), five vegetation-type classes (mixed forests, shrubland, and grassland), and one non-vegetation class. Vegetation phenology metrics were derived using two methods: a curve-fitting approach (TIMESAT) and a harmonic analysis of time series (HANTS). We integrated these phenological features with the original spectral bands, Sentinel-1 SAR backscatter, texture metrics, and topo graphic variables as inputs to multiple classifiers (Random Forest, Support Vector Machine, Convolutional Neural Network, Vision Transformer, and Extreme Gradient Boosting). We also applied Shapley Additive Explanations (SHAP) to interpret model feature importance. Without phenological inputs, the best model achieved 92.5% overall accuracy. Incorporating phenological variables from TIMESAT increased accuracy to 96.1%—slightly higher than using HANTS-derived features (95.5%). SHAP analysis revealed that time-series phenological and spectral variables con tribute most to species discrimination, while SAR backscatter and topographic metrics provide crucial complementary information. These findings demonstrate that exploiting Sentinel-2's temporal dynamics alongside multi-source data markedly improves tree species classification in complex mountainous forests, and high light the added value of advanced phenology extraction methods and explainable machine learning techniques.