Urban settlement tree detection and carbon sequestration mapping using probability map–augmented deep learning: a case study in Seoul

Urban settlements are critical yet underrepresented components of national greenhouse gas inventories. This study presents a segmentation-based GeoAI framework—aligned with Tier 2A methodologies—for mapping tree canopy and estimating carbon sequestration at high spatial resolutions in Seoul, South Korea. First, we evaluate the seasonal generalization of transformer-augmented U-Net models trained on summer, winter, and combined aerial RGB imagery, demonstrating that multi-season training mitigates performance collapse under off-season conditions (mIoU winter: 0.31 → 0.39). Next, we integrate Sentinel-2-derived Tree Probability Maps (TPMs) as a fourth input band, further improving winter-season segmentation performance (mIoU: 0.39 → 0.49) and reducing omission errors and boundary fragmentation in shadowed or sparse-canopy areas. Predicted canopy areas are converted to carbon sequestration maps using nationally certified coefficients and aggregated to 1 m, 10 m, and 100 m resolutions to align with block-, neighborhood-, and district-level planning units. Validation against the national street-tree inventory demonstrates strong spatial correspondence and high predictive confidence, with the 4-band model producing sharply right-skewed probability distributions at known tree locations. The proposed framework thus offers a scalable, season-resilient solution for urban tree canopy monitoring and carbon accounting, with immediate applicability to municipal and national reporting systems.