Large-scale mapping and uncertainty assessment of dominant tree species in southern China based on Sentinel-2 time...

The absence of large-scale maps of dominant forest species hinders research on species diversity, forest disturbance, and carbon sinks, while also constraining effective forest management. In southern China, challenges such as the timeliness and number of sample data, complex environmental heterogeneity, and the confounding effects of understory vegetation introduce potential taxonomic uncertainties in forest species mapping. To address these issues, this study proposes a comprehensive mapping framework that incorporates inter-annual sample updating and augmentation using the Continuous Change Detection and Classification (CCDC) algorithm and absolute thresholding based on class probability to construct a reliable training dataset. Predictor variables were derived from Sentinel-2 time-series and Digital Elevation Model (DEM) data. To mitigate the effects of regional heterogeneity, the Extreme Gradient Boosting (XGB) classifier was applied within sub-regions. Classification uncertainty associated with understory interference was assessed using binary contour mapping and correlation analysis. The effectiveness of inter-annual sample transfer was validated with a minimum update accuracy of 88 %, confirming the robustness of the CCDC-based approach. Moreover, incorporating augmented samples through thresholding improved classification accuracy by 2-6 percentage points in regions such as Anhui (AH), Yunnan (YN), Hunan (HN), and Hainan (HNA), with accuracies exceeding 80 %. The overall classification accuracy increased by 2 percentage points, surpassing 75 % globally. A significant negative correlation between canopy cover and classification uncertainty emphasized the influence of understory complexity on model performance. This study generated the first 10-m resolution map of dominant tree species across southern China, covering 85 % of the region's forests. The reproducibility of the proposed mapping framework and the accessibility of the resulting dataset provide valuable resources for filling temporal and spatial gaps in species-level forest information.