Predicting forest inventory variables with remote sensing data : the role of number and locality of the training plots

Field measurements of national forest inventories (NFIs) and auxiliary data derived from remote sensing can be used to model forest inventory variables. Because NFI observations typically cover large areas, training data for modeling is often limited by the number and proximity of the available plots. Determining suitable thresholds for these, however, is complicated and depends on various factors. In this study, we examine how the number and locality of the NFI plots affect the modeling accuracy. With locality, we refer to the spatial proximity of the training plots to the target location, which is often considered a proxy of their representativeness. We used field measurements from the Finnish NFI, collected mainly from managed, conifer-dominated boreal forests, and modeled growing stock volumes and trees’ dominant heights with Sentinel-2 and airborne laser scanning (ALS) data. We then compared prediction errors derived from varying numbers of training plots and modeling attempts where we purposely excluded the nearby plots within a specific radius. As a special case of training plots’ representativeness, we also assessed if models estimated using ALS features were transferable over the production area borders, i.e. scanned with similar quality standards but at different times and using different scanners. Our results indicated that a set of 100 training plots could be regarded as the minimum preferred count regardless of the auxiliary data. Locality affected modeling bias, which increased when the exclusion radius was extended to 50 km or beyond. This bias connected particularly to models’ abilities for adapting to local conditions which resulted in regional under- and overestimates, but these effects were only mildly observable through global statistics over all the applied validation data. Differences between ALS production areas were generally small, but they varied by inventory variable and were more pronounced for features derived from first echoes than from last or all echoes.