Importance ranking of data and model uncertainties in quantile regression forest-based spatial predictions when data are sparse, imprecise and clustered

Quantile regression forest (qRF) is a popular approach for producing maps in environmental studies because of its efficiency and of the advantage of providing a quantification of prediction uncertainty. A procedure for sensitivity analysis is developed to quantify the influence of two types of uncertainties on the qRF predictive performance: (1) the model uncertainties related to qRF modelling choices, i.e., the values of the hyperparameters, the selection of the covariates, and the inclusion of spatial Euclidean distance fields (EDFs); (2) data imperfections related to measurement errors and to the spatial distribution of the samples. This problem is further addressed by examining whether sparse, imprecise and clustered data, named SIC data, affects the importance ranking of the data and model uncertainties. By considering two different environmental variables, above ground biomass, and soil organic carbon stock for Western Europe, a series of repeated random experiments is defined to cover multiple situations (spatial distribution of the samples, number of clusters, number of samples outside the regions where the data are clustered, and performance scores). For the cases considered, the results confirm the utmost importance of the uncertainty related to the selection of covariates whether the data are SIC or purely randomly distributed. When the SIC data show a single cluster, the importance of including EDFs increases with the intensity of the data clustering. These results are however sensitive to characteristics of the SIC data spatial distribution (position of the cluster, sample distribution). When the SIC data show several clusters, including this type of spatial proxies is even as important as the selection of covariates. Finally, the results for above ground biomass highlight the need for a careful selection of the hyperparameter m try that controls the number of covariates during the construction of the random forest model. • The influence of data and model uncertainties on qRF performance is quantified. • Studied cases concern sparse, imprecise and clustered (SIC) data. • The selection of covariates is the utmost importance. • When data are clustered, the importance of hyperparameters is enhanced. • Importance of including spatial proxies depends on the SIC characteristics.