On Bayesian regression and regression forest for digital mapping of soil properties at field scale

CÓRDOBA, M.; BALZARINI, M.

Congreso; 31st International Biometric Conference; 2022

Information on soil properties variability at field scale is a key element for sustainable agriculture and soil conservation. The digital mapping based in a regression model trained from a soil variable sampled at some field sites and site-covariates, easier to obtain than soil data, allows economic studies of soil variability. The objective of this work was to compare the performance of approaches for modelling and digital soil mapping at field scale: 1. Spatial Bayesian regression models (BR), and 2. The machine learning algorithm, quantile regression forest (QRF). The used explanatory variables, eight topographic plus four vegetation indices, were easy to obtain and potentially correlated to soil variability. QRF is a tree-based ensemble method that provide information about the full conditional distribution of the response variable. A spatial prediction obtained by ordinary kriging on the prediction errors of the algorithm was added to the QRF predictions at unsampled sites. In BR, a posterior distribution of predicted values for the response variable conditional to the selected explanatory variables was obtained adding random site effects as Gaussian observations. The BR model was estimated by Integrated Nested Laplace Approximations (INLA) combined with Stochastic Partial Differential Equations (SPDE) to model spatiality. A first-order autoregressive model was assumed for the latent process that changes with time. Both, QRF and BR, were fitted for organic matter (OM), pH and phosphorus (P) as response variables. The database included 300 georeferenced site- measurements for each response across 2240 ha agricultural field collected in three years (2005, 2008 and 2011). Cross-validation of the k-fold type (with k = 10) was performed to estimate prediction error and several metrics to compare spatial predictions. A classical regression-kriging (RK) was fitted as reference. RB showed the best quantitative criteria for MO y pH, but not for P, the variable with the highest variance. Further, site-specific predictions based on RB presented less uncertainty. The RB estimated by INLA from spatio-temporal data represented a powerful methodological approach for fine-scale digital soil mapping.