Examining model transferability for forest carbon stock estimation in subtropical ecosystems based on airborne Lidar and field measurements

Airborne Lidar has been widely recognized as a highly dependable data source for estimating forest carbon stock, but the lack of transferable modeling approaches limits its broader application. This study proposed a framework that leverages prior knowledge from training regions while minimizing the need for extensive local calibration. Two modeling scenarios were evaluated: a universal mixed-effects model integrating airborne Lidar and sample plots from multiple regions, explicitly accounting for forest type and regional effects. In addition, a newly developed prior-constrained backpropagation (PCB) algorithm was designed to calibrate models in transfer areas with a limited number of samples. The results indicate that (1) the mixed-effects modeling approach could be directly applied to untrained areas within the same province, achieving a relative root mean square error (RMSEr) in the range of 15.92%–19.3%; (2) the PCB algorithm enabled effective model calibration using fewer than 10 local samples, improving forest carbon stock estimation in transfer areas with RMSEr values below 20% for Masson pine and eucalyptus, and below 15% for Chinese fir; (3) the performance of optimal PCB varied depending on modeling variables, and improved the estimation accuracy using necessary variables for the model calibration in transfer areas. This research highlights how combining Lidar with advanced computational modeling can improve the spatial transferability of carbon stock estimation, especially in regions where sample collection is difficult but Lidar data are available.