Forest-volume maps can carry their uncertainty instead of hiding it
WHAT THE EVIDENCE SAYS
Using field plots and UAV-LiDAR metrics, researchers compared forest stock-volume models and found a probabilistic CB-NGBoost approach performed best among the tested models on an independent holdout, with R² of 0.74 and RMSE of 71.24 m³/ha. SHAP analysis indicated that variables driving the central estimate were not identical to those driving predictive uncertainty.
WHY IT MATTERS / OUR READ
For inventories and carbon or harvest decisions, an uncertainty surface can be as useful as the map itself. It helps separate places where the model predicts a value from places where the model has enough support to justify confidence in that value.
CAVEATS / UNCERTAINTY
The field dataset contained 30 original plots and only six original plots in the independent test split. The uncertainty treatment emphasized aleatoric uncertainty, prediction intervals were not perfectly calibrated, and transfer to other forest types or sensors was not tested.
SOURCE
Forest Ecosystems — primary study
View original evidence ↗
A UAV-LiDAR model in southern China paired probabilistic stock-volume estimates with interpretability, showing that predictors of the mean estimate were not the same as predictors of uncertainty.
Read the Wild Lands assessment →