• A province-wide inventory of 12.29 M visually interpreted forest stands was used. • The first 10 m map separating natural and planted forests in Guangxi was produced. • Object-based random forest outperformed pixel-based models in all evaluations. • Topography and vegetation traits (LCC, LAI) were key predictors of forest type. • Spatially stratified sampling showed coverage matters more than local sample size. Accurately distinguishing between natural forests and planted forests underpins effective biodiversity monitoring and climate policy, yet their spectral similarity and limited reference data challenge large-scale classification efforts. This study focuses on Guangxi Province, China, combining Sentinel-2 imagery, multi-source ecological indicators—including vegetation traits such as leaf chlorophyll content and LAI—and over 12 million visually interpreted forest stands to map the distribution of natural and planted forests. We assessed pixel- and object-based classification models under various sample scenarios, with accuracy independently validated using additional samples interpreted from high-resolution Google Earth imagery. Model evaluation relied primarily on random splits, supplemented with spatially stratified validation across ecological sub-regions to assess spatial generalization. Results show that the object-based random forest classifier achieved the highest accuracy (OA: 84.52%, F1: 84.04%), with topographic and vegetation functional traits as key predictors. Spatially, natural forests dominate mountainous zones, while planted forests prevail in flatter areas. This work delivers the first 10 m resolution forest type map for Guangxi based on a uniquely large training and validation dataset and demonstrates that combining functional traits with robust sampling improves classification performance. Our framework supports scalable forest monitoring and contributes to improved management and conservation strategies.
Guo et al. (2026) studied this question.