Advancing forest GHG inventory accuracy with a phenological classification framework: Toward an observation-based approach 3 in South Korea

Reliable, observation-based activity data are essential for improving national greenhouse gas (GHG) inventories in the LULUCF sector. However, in South Korea, the method used for current GHG estimations for the forest sector has several limitations. Therefore, we developed a phenological classification framework (PCF) that exploits seasonal reflectance dynamics in Sentinel-2 time series (10 m) to map coniferous and broadleaf forests across South Korea. Using a U-Net classifier trained on phenology-aware seasonal composites, we generated annual wall-to-wall maps for 2019–2021 and evaluated performance with independent visual interpretation. The PCF attained an overall accuracy of 83.13 % (kappa = 0.6755), and class-wise histogram distributions of DN values exhibited consistent, year-to-year separability between coniferous and broadleaf forests, supporting the reliability of phenology-driven discrimination beyond pointwise metrics. Applying the maps to mixed-forest areas revealed a pronounced, policy-relevant asymmetry that departs from the conventional 50:50 conifer–broadleaf allocation embedded in Approach 1 statistic. When propagated through the stock-difference method, the observation-based areas yielded carbon stock trajectories that diverged from national reports, underscoring how improved activity data can materially influence inventory outcomes. Emphasizing temporal (phenological) variation in freely available multispectral data can deliver scalable, operational, and repeatable Approach 3 activity data aligned with IPCC guidelines—without requiring hyperspectral sensors or dense ancillary inputs. Overall, the PCF provides a practical pathway to close gaps between statistical assumptions and observed forest dynamics, improving both scientific understanding and the credibility of national carbon accounting.