Assessing the robustness of variable selection methods when accounting for co-registration errors in the estimation of forest biophysical and ecological attributes

The estimation of forest attributes using active remote sensing data and sampling designs is a solid scientific discipline oriented to derive structural information from forest ecosystems. In this study, we examined the problem of selecting predictor variables derived from airborne laser scanning (ALS) when using stepwise and multiple linear regression, and when using kNN imputation using Random Forests (RF) to select predictor variables. Systems of forest attributes and ecological indicators were estimated using ALS statistics and ground data collected in Central-East Spain, on two Mediterranean pine forest areas. Co-registration errors were simulated by means of altering ground data coordinates. The larger the co-registration error, the higher the variation in the selected ALS-based predictors in the parametric approach, which was less affected than the RF-based selection. In both cases, the final performance of the models was not as affected by co-registration errors as the observed variation in the selection of variables: the impact was around 1% in the parametric approach while the value reached the 8% for both stand basal area and the Gini-index under the non-parametric method. The forest biophysical attributes of the set behaved differently towards altering the co-registration factor in both cases. To control the co-registration factor between remote sensing data sources and ground surveys is an essential step to reduce uncertainty in data interpretation and in the modeling of forest attributes.