Identifying tree species in mining landscapes of the middle Yellow River: 5-year phenological trajectory divergence time-series from Sentinel-2 data

Accurate quantitative information on tree species is essential for achieving sustainable forest development and ecological conservation in the Yellow River Basin. Vegetation-based phenology studies provide a critical foundation for monitoring inter-annual variability and long-term trends in vegetation dynamics. However, in mining-forest composite zones, extreme habitat fragmentation and highly artificial vegetation cover exacerbate spectral confusion in tree species identification. Furthermore, the classification potential of multi-temporal data and the extraction and utilization of valuable phenological features from multi-temporal data remain underexplored. In this study, the Huodong mining area was selected as the research site—a representative coal mining region in the middle reaches of the Yellow River. Using medium-resolution, dense time-series Sentinel-2 data spanning 2019–2023 to derive vegetation phenology and canopy structural characteristics. Phenological Trajectories Divergence Time Series (PTDTS) of tree species were comprehensively extracted from four dimensions: temporal, spatial, frequency, and dynamic features, and were used as input variables to compare the classification performance of different machine learning algorithms, ultimately selecting the optimal model for tree species classification. Analysis of multidimensional time-series data, interspecies phenological differences were assessed across different years, and the influence of environmental variables and feature combinations on classification performance was evaluated. The results showed that the optimal feature combination achieved an overall accuracy of 85.71%. Notably, using only spectral features and PTDTS variables alone could reach an accuracy of 85.62%, demonstrating that spectral-temporal phenological trajectory features remain the dominant explanatory information, even when topographic features and Sentinel-1 data are used to supplement our analysis. Additionally, the study revealed five years of continuous data, improving classification accuracy by 8.12% compared to single-year data. These findings underscore the importance of long-term time-series imagery for tree species mapping and the capability of Sentinel-2 data to support accurate regional tree species classification based on phenological trajectory analysis.