Potential value of combining CNN, GEDI and multi-source remote sensing data to improve the estimate of aboveground forest carbon storage in northeast China

Quantifying forest biomass carbon (C) stock is critical for determining the regional carbon balance, but, a lack of both field observations spanning large climatic gradients and proper upscaling methods which take the spatial pattern into account, means there is little knowledge regarding forest C stock at high spatial resolutions. Here, we address this problem by combining a deep convolutional neural network (CNN) algorithm with in situ measurements, Global Ecosystem Dynamics Investigation (GEDI) observations, and Landsat and PALSAR-2 images to develop a new, spatially explicit estimate of forest aboveground carbon density (ACD) circa 2020 at a 30 m spatial resolution for northeast China, home to nearly one-third of China’s forested area. The result yields a high coefficient of determination (R²) of 0.83 and a relatively low root mean squared error (RMSE) of 5.28 MgC ha^-1, and is superior to traditional pixel-based and in situ based methods. Through linking in situ measurements with nearly 0.13 million GEDI observations, we obtained important samples across spatially variable environmental conditions, and in remote and rugged regions (when increasing the number of GEDI samples, RMSE decreased by 73.5%). CNN was able to extract important spatial patterns and performed well in capturing the spatial variation of forest carbon density. We also propose a CNN-based perturbation method to rank variable importance, which shows that the distribution of forest C storage is mainly determined by precipitation and forest age. Based on the proposed method, the local forest aboveground biomass C stock is estimated to be 3.52 ± 0.10 PgC, with an age-related forest aboveground biomass C sink of 7.94 TgC year^-1 before 2060. Terrestrial ecosystem models generally underestimate the regional C stock, partially because of biases in forest age simulations. The study highlights the importance of using deep learning methods to gain further process understanding of forest carbon dynamics under climate change.