Large-scale forest resource mapping with spatial gaps in the training data: Comparison of different modeling approaches

Abstract

Forest attribute maps are essential for supporting local decision-making regarding forest resource use. Such maps are produced by combining remote sensing and field data through various modeling approaches. When mapping across large areas, spatial gaps in field data used for model training are common. Our study evaluates the performance of three methods—k-Nearest Neighbor (k-NN), Random Forests (RF), and Multi-Layer Perceptron (MLP)—for forest resource mapping across Norway, Sweden, and Finland in an experimental setup with respect to availability of field data around the target area. Models were trained with sample plot sizes (N) ranging from 100 to 3000. RF consistently produced the most accurate predictions in terms of relative bias and RMSE. While spatial gaps in the training data (radius: 7–141 km) affected %RMSE of broad-leaved above ground biomass (AGB), they had minimal impact on %RMSE of both local and country-level predictions of total AGB and volume. For RF with N=3000, %RMSE of total AGB ranged between 53%–55% in Finland and Sweden, and 70%–72% in Norway across gap sizes. However, %bias increased for local predictions across the whole study region with larger gaps: RF with N=500 showed bias of −12%–12% (7 km gap) and −17%–28% (78 km gap). Similarly, country-level %bias of total AGB for Norway increased from −1.7% to −3.7% with larger gaps. In conclusion, spatial gaps in training data can significantly affect bias in predictions. Therefore, forest attribute maps should always be accompanied by metadata describing the training data used.