Consequences of spatial structure in soil-geomorphic data on the results of machine learning models

In this paper, we examined the degree to which inherent spatial structure in soil properties influences the outcomes of machine learning (ML) approaches to predicting soil spatial variability. We compared the performances of four ML algorithms (support vector machine, artificial neural network, random forest, and random forest for spatial data) against two non-ML algorithms (ordinary least squares regression and spatial filtering regression). None of the ML algorithms produced residuals that had lower mean values or were less autocorrelated over space compared with the non-ML approaches. We recommend the use of random forest when a soil variable of interest is weakly autocorrelated (Moran's I < 0.1) and spatial filtering regression when it is relatively strongly autocorrelated (Moran's I > 0.4). Overall, this work opens the door to a more consistent selection of model algorithms through the establishment of threshold criteria for spatial autocorrelation of input variables.

Automated summary

In this study, we investigated the impact of inherent spatial structure in soil properties on the accuracy of machine learning approaches in predicting soil variability. We compared the performance of four machine learning algorithms and two non-machine learning algorithms. The results showed that none of the machine learning algorithms outperformed the non-machine learning approaches in terms of residual values and spatial autocorrelation. We recommend using random forest for weakly autocorrelated soil variables (Moran's I < 0.1) and spatial filtering regression for relatively strongly autocorrelated variables (Moran's I > 0.4). This research provides a framework for selecting appropriate model algorithms based on spatial autocorrelation criteria for input variables.