Can learning regression features by computer vision improve the generalisation of geostastistical interpolators?

Recent approaches for large-scale mapping of continuous environmental variables by combining ground observations, remote sensing and machine learning have proposed incorporating computer vision capabilities into the model, so that potentially-complex regression features may be learned automatically from covariate datasets, such as of terrain elevation and other satellite imagery (e.g. see Kirkwood et al 2022; 'Bayesian deep learning for spatial interpolation in the presence of auxiliary information').

Here we present new research using national-scale land-surface geochemical data to explore and compare how the incorporation of computer vision for automatic feature learning affects the predictive performance of geostastistical interpolators both within and beyond the spatial extents of the study areas in which ground observations are collected. We attempt to characterise empirically how well the predictive performance of different models is preserved with increasing distance from training observations in order to provide insights into the value of incorporating computer vision capabilities into geostatistical models, compared to more traditional approaches.