Linking land cover changes to carbon storage in Hunan Province, China: Insights from multi-task learning

Long-term land cover dynamics are a key regulator of regional carbon storage, yet most assessments still rely on a few temporal snapshots and therefore overlook how land transitions unfold over time. Here we reconstruct annual land cover in Hunan Province, China, from 1990 to 2023 using a multi-task deep learning framework, and combine it with climate-corrected carbon accounting, driver attribution and future scenario simulation. Although total carbon storage increased from 2122.74 Mt in 1990 to 2209.98 Mt in 2023 in Hunan, this net gain masked pronounced spatial heterogeneity and distinct carbon responses to land cover transition frequency and pathways. Trajectory analysis further showed that different change frequencies and transition pathways produced distinct carbon consequences, with return-type pathways tending toward near-neutral responses and directional pathways leading to clearer carbon gains or losses. The dominant drivers of carbon storage were also not stationary, with vegetation condition prevailing in 2000 and 2020, whereas population density became most important in 2010. Interactions among vegetation condition, topography, and population density mainly shaped spatial heterogeneity in carbon storage. Historical-trend simulation further indicates that regional carbon storage growth may decelerate sharply by 2030, with only a marginal increase of 1.35 Mt relative to 2020. These findings are particularly concerning because the potential to offset the carbon costs of continued urban expansion through land conversion alone may be increasingly limited, making the protection of existing high-carbon ecological spaces more urgent.