Applied remote sensing and explainable machine learning unveil vegetation carbon storage implications of forest vegetation changes in Yunnan, China

Forests are key terrestrial carbon pools, with carbon density varying across vegetation types. Since 2000, planted forests have expanded and natural forests have declined globally, yet how these dynamics affect regional carbon storage remains unclear. Yunnan, in southwest China, harbors forests ranging from tropical to alpine. To quantify how forest dynamics affect carbon storage in Yunnan, we integrated remote sensing, field surveys, and explainable machine learning. We mapped nine forest vegetation types and assessed both their spatio-temporal dynamics and associated carbon storage changes from 2000 to 2020. Explainable machine learning revealed that greater distance from croplands and grasslands was linked to higher forest stability, while natural forest loss and planted forest gain mainly occurred where annual precipitation exceeded 1200 mm. Vegetation, distance to croplands and grasslands, and climatic factors were most strongly associated with vegetation carbon density. Natural forests had a mean carbon density 52.67 Mg/ha higher than planted forests. Over two decades, remote sensing estimated a 19.85% decrease in natural forest area and a 9.18% increase in planted forest area (error-adjusted: 14.89% ± 0.73% and 8.26% ± 0.42%, respectively). Based on these maps, natural forest carbon storage declined by 7.69% ± 5.06%, while planted forest expansion and rising carbon density yielded a net 13.86% ± 4.88% gain in vegetation carbon storage. Overall, this integrated framework provides an assessment of forest vegetation dynamics and carbon storage consistent with ecological principles, and future studies should incorporate additional multi-source observations for more detailed ecological assessments.