A framework for developing spatially transferable models of boreal forest biomass from airborne lidar

Abstract Forest inventory (FI) maps of aboveground biomass (AGB) and other FI attributes are often based on airborne laser scanning (ALS) models calibrated with limited ground FI plots. Being empirical, spatial transferability (i.e. similar performance in regions lacking ground data) is a key concern. Here, we propose and test a framework for developing spatially transferable models of AGB that could be used for large-area FI mapping. We employ a data-driven, best-subset selection (BSS) approach to identify top-performing models among thousands of candidates of linear (including some with the square root of AGB as response) and log–log (power) form and up to three predictors selected from hundreds of metrics derived from multispectral (3-channel) ALS point clouds acquired over 163 FI plots in four regions of the southern Taiga Plains of the Northwest Territories, Canada. After discarding models with high multi-collinearity, spatial transferability is tested via leave-one-region-out cross-validation (LORO-CV). In addition to low Root Mean Square Error (RMSE), we require negligible bias (< 1SE rule, i.e. bias below one standard error) in each omitted region. A final t-test on low-biomass stands (~10% lowest AGB plots) further removes any model systematically overestimating AGB in those stands, which are prevalent in the Taiga Plains. From more than 11 000 initial models, only a few satisfy those criteria. The recommended model—a linear, three-predictor solution—achieved a LORO-CV mean absolute bias of 2.0%, a LORO-CV RMSE of 19.3% and an in-sample RMSE of 20.5% (27.8 Mg ha−1, adjR2 = 0.82), slightly higher than the 19.8% (26.8 Mg ha−1) of the best BSS model, which did exhibit a regional bias of up to 5.5%. Linear models using intensity-weighted metrics from the 1064 nm channel outperformed square root, log–log, multispectral, or first-returns-only solutions. Just minimizing in-sample RMSE does not guarantee strong spatial transferability; enforcing near-zero regional bias does, with a small accuracy trade-off.