Spatiotemporal prediction of forest litterfall in China by using multi-source data and Transformer-CatBoost model

Abstract Forest litterfall is a key contributor to soil carbon accumulation. However, existing studies have primarily foused on site-level observations or annual-scale assessments, while the intra-annual dynamics and spatial distribution of forest litterfall at the national scale remain poorly understood. In turn, this limitied comprehensive spatiotemporal assessments of forest carbon sequestration capacity. In this study, we compiled 4,223 monthly litterfall observations from 88 forest sites across China and integrated multi-source environmental variables to develop a Transformer-CatBoost hybrid prediction model for estimating the spatiotemporal patterns of forest litterfall across three representatibe years (2002, 2009 and 2018), corresponding to major stages of ecological restoration efforts in China. Model evaluation demonstrated strong predictive performance ( R 2 = 0.74), effectively capturing the nonlinear relationships driving litterfall dynamics. By incorporating national forest area changes in 2002, 2009, and 2018, the study further revealed the spatiotemporal evolution of forest structure under large-scale ecological restoration programs. Based on nationwide monthly-scale modeling results, we systematically characterized the spatial distribution and seasonal variation of litterfall production across China’s forests, with an anuual average of 547.04 ± 0.23 g m⁻ 2 (or 479.13 ± 0.20 g m⁻ 2 excluding January and December). Furthermore, using a fixed carbon conversion rate, we estimated national carbon content of forest litterfall at 290.4 Tg in 2002, 311.9 Tg in 2009, and 354.1 Tg in 2018, indicating a clear increasing trend. This study represents the nationwide, monthly-scale modeling and prediction of forest litterfall in China. Graphical Abstract