Assessing the spatial sensitivity of a random forest model: Application in gridded population modeling

Gridded human population data provide a spatial denominator to identify populations at risk, quantify burdens, and inform our understanding of human-environment systems. When modeling gridded population, the information used for training the model may differ in spatial resolution than what is produced by the model prediction. This case arises when approaching population modeling from a top-down, dasymetric approach in which one redistributes coarse administrative unit level population data (i.e., source unit) to a finer scale (i.e., target unit). However, often overlooked are issues associated with the differing variance across the scale, spatial autocorrelation and bias in sampling techniques. In this study, we examine the effects of intentionally biasing our sampling from the source to target scale within the context of a weighted, dasymetric mapping approach. The weighted component is based on a Random Forest estimator, which is a non-parametric ensemble-based prediction model. We investigate issues of autocorrelation and heterogeneity in the training data using 18 different types of samples to show the variations in training, census-level (i.e., source) and output, grid-level (i.e., target) predictions. We compare results to simple random sampling and geographically stratified random sampling. Results indicate that the Random Forest model is sensitive to the spatial autocorrelation inherent in the training data, which leads to an increase in the variance of the residuals. Sample training datasets that are at a spatial scale representative of the true population produced the best fitting models. However, the true representative dataset varied in autocorrelation for both scales. More attention is needed with ensemble-based learning and spatially-heterogeneous data as underlying issues of spatial autocorrelation influence results for both the census-level and grid-level estimations.