How reliable are remote sensing maps calibrated over large areas? A matter of scale?

TL;DR

This study evaluates how the transferability of remote sensing models for forest attributes varies with spatial scale and sampling effort, highlighting the importance of model validity domains for accurate forest mapping.

Contribution

It provides an analysis of model transferability across different spatial scales and sampling efforts, emphasizing the impact on forest attribute map accuracy.

Findings

Local models exhibit bias and extrapolation issues when applied elsewhere.

Increasing sampling efforts reduces extrapolation problems.

Model transferability is crucial for reliable forest management maps.

Abstract

Remote sensing data are increasingly available and frequently used to produce forest attributes maps. The sampling strategy of the calibration plots may directly affect predictions and map qualities. The aim of this manuscript is to evaluate models transferability at different spatial scales according to the sampling efforts and the calibration domain of these models. Forest inventory plots from locals and regionals networks were used to calibrate randomForest (RF) models for stand basal area predictions. Auxiliary data from ALS flights and a Sentinel-2 image were used. Model transferability was assessed by comparing models developed over a given area and applied elsewhere. Performances were measured in terms of precision (RMSE and bias), coefficient of determination (R2) and the proportion of extrapolated predictions. Regional networks were also thinned to evaluate the effect of sampling efforts on models' performances. Local models showed large bias and extrapolation issues when applied elsewhere. Local issues of regional models were also observed, raising transferability and extrapolation concerns. An increase in sampling efforts was shown to reduce extrapolation issues. The outcoming results of this study underline the importance of considering models' validity domain while producing forest attribute maps, since their transferability is of crucial importance from a forest management perspective.