Urbanization and afforestation are associated with competing carbon stock dynamics in Xiong ’an New Area: A decade of evidence from multi-temporal land cover mapping

The temporal sequence of annual land-cover change and carbon storage is a useful ecological indicator for tracking carbon-sink function in rapidly urbanizing areas. Existing studies, however, rely on medium- to low-resolution products and static single-date estimates, which cannot capture how carbon-sink function evolves over time. The period during which newly planted trees are spectrally hard to distinguish overlaps with the early stage of carbon accumulation, so misclassification can systematically underestimate carbon-sink areas. Xiong’an New Area, where rapid urbanization and large-scale afforestation occur together, is a suitable case for studying this problem. We propose TS-Carbon-Framework, which combines a three-frame semantic segmentation network (TriFuseNet), the InVEST carbon-storage model, and error-propagation analysis into a processing chain that turns multi-temporal Sentinel-2 imagery into annual carbon-storage indicators. TriFuseNet separates newly planted vegetation from cropland using the cumulative spectral difference between a historical frame taken before planting and the current frame. On the 8-class segmentation dataset for Xiong’an New Area, the mIoU reached 45.33%, 2.44 percentage points higher than the best comparison method, and the Trees IoU reached 35.21%, 6.99 percentage points higher than the best single-frame comparison method (SegFormer-B4). Total carbon storage ranged from 1.938 × 1 0 7 to 2.658 × 1 0 7 Mg C. The systematic bias of TriFuseNet was − 2.43 % , and after correction the carbon storage of the Trees class was 10.32% higher than the pixel-count estimate. This updatable indicator system can serve as a reference for regional ecological monitoring and carbon-sink management.