Forest aboveground carbon storage estimation and uncertainty analysis by coupled multi-source remote sensing data in Liaoning Province

• A multi-scale framework for estimating forest AGC stock was presented. • Confirmed the feasibility of sample expansion through UAV high-resolution images. • Considering spatial geographic correlation solves the spot pixel matching issues. • Presented a method for quantifying uncertainty propagation. Accurate mapping of large-scale forest aboveground carbon (AGC) stock is essential for understanding the role of forests in the global carbon cycle. Traditional forest resource inventory methods pose limitations due to their sparse spatial coverage and time-consuming data collection. While satellite sensors offer extensive spatial and temporal coverage, their low spatial resolution often introduces significant errors when ground measurements are directly matched with satellite image pixels. This study proposes a large-scale research framework for precise forest AGC stock mapping, which can effectively address these challenges by integrating unmanned aerial vehicle (UAV) imagery with spaceborne LiDAR data. Using Liaoning Province as the study area and based on field sample collection, UAV high-resolution images were used to generate expanded samples at the individual tree scale, and the forest AGC stock results at the spot scale were obtained through accumulation. Considering the number of expansion samples and the bias of the spaceborne LiDAR position caused by the terrain, we expanded the samples via GANs, and tested the effects of different spot radii on the model’s fitting accuracy via a random forest regression model. Finally, selecting 12.5 m as the optimal fitting radius of the model, we obtained the forest AGC stock results at the spot scale of the GEDI and ICESat-2 in the province. Owing to the difficulty of one-to-one matching between spaceborne LiDAR spots and satellite image pixels, geographical correlation was performed to extract the average pixel value of multiple pixels covered by the light spot based on an area-weighting method as the model’s input features. Combining this step with an ensemble machine learning algorithm, the final estimate of forest AGC stock in Liaoning Province was calculated to be 101.35 Tg, with an uncertainty of ±37.31 Tg. Our approach outperformed publicly available products, namely, AGBCY2021, ESA_CCI, and GEDI_L4B_V2, achieving an RMSE (%) of 19.86 %, and demonstrating the efficacy of the proposed method for quantifying uncertainty propagation in a multiscale analysis framework.