Assessing Site Effects and Geographic Transferability when Interpolating Point Referenced Spatial Data: A Digital Soil Mapping Case Study

TL;DR

This study evaluates how well models trained on soil data from one site predict data at another site, using advanced variable selection methods to improve spatial interpolation of soil carbon.

Contribution

It introduces a framework for assessing geographic transferability of site-specific models using LASSO-regularized regression with multiple environmental variables.

Findings

Site-specific models vary in transferability across locations.

LASSO-based variable selection improves model performance.

Models can inform soil carbon sequestration strategies.

Abstract

When making inferences concerning the environment, ground truthed data will frequently be available as point referenced (geostatistical) observations that are clustered into multiple sites rather than uniformly spaced across the area of interest. In such situations, the similarity of the dominant processes influencing the observed data across sites and the accuracy with which models fitted to data from one site can predict data from another site provide valuable information for scientists seeking to make inferences from these data. Such information may motivate a more informed second round of modelling of the data and also provides insight into the generality of the models developed and an indication of how these models may perform at predicting observations from other sites. We have investigated the geographic transferability of site specific models and compared the results of using different implementations of site specific effects in models for data combined from two sites. Since we have access to data on a broad collection of environmental characteristics that each held potential to aid the interpolation of our geostatistical response observations we have investigated these issues within the framework of a computationally efficient method for variable selection when the number of explanatory variables exceeds the number of observations. We have applied Least Absolute Shrinkage Selection Operator (LASSO) regularized Multiple Linear Regression (MLR) as fitted by the computationally efficient Least Angle Regression algorithm. The response variable in our case study, soil carbon, is of interest as a potential location for the sequestration of atmospheric carbon dioxide and for its positive contribution to soil health and fertility.