Comparison of various uncertainty modelling approaches based on geostatistics and machine learning algorithms

Gábor Szatmári,László Pásztor

doi:10.1016/j.geoderma.2018.09.008

Gábor Szatmári, László Pásztor

Open Access

https://doi.org/10.1016/j.geoderma.2018.09.008

Copy DOI

Abstract

We compared the suitability of several commonly applied digital soil mapping (DSM) techniques to quantify uncertainty with regards to a survey of soil organic carbon stock (SOCS) in Hungary. To represent the wide range of DSM techniques fairly, the followings were selected: universal kriging (UK), sequential Gaussian simulation (SGS), random forest combined with kriging (RFK) and quantile regression forest (QRF). For RFK two different uncertainty quantification approaches were adopted based on kriging variance (RFK-1) and bootstrapping (RFK-2). The selection of the potential environmental covariates was based on Jenny's factorial model of soil formation. The spatial predictions of SOCS and their uncertainty models were evaluated and compared using a control dataset. For this purpose, we applied the most common measures (i.e. mean error and root mean square error), furthermore, accuracy plot and G statistic. According to our results, QRF and SGS produced the best uncertainty models. UK and RFK-2 overestimated the uncertainty whereas RFK-1 produced the worst uncertainty quantification according to the accuracy plots and G statistics. We could draw the general conclusion that there is a need to validate the uncertainty models. Furthermore, great attention should be paid to the assumptions made in uncertainty modelling.

Highlights

Predictive soil maps suffer from different types of errors, where the most common error sources could be the measurements, digitization, typing, interpretation, classification, generalization and interpolation (Heuvelink, 2014)
The quantification, visualization and communication of the uncertainty of the digital soil mapping (DSM) products would be indispensable to stakeholders as it has already been stressed by the GlobalSoilMap.net initiative (Arrouays et al, 2014)
We examined at the control points that how many times soil organic carbon stock (SOCS) falls within the 90% prediction interval (PI)

Summary

Introduction

Predictive soil maps suffer from different types of errors, where the most common error sources could be the measurements, digitization, typing, interpretation, classification, generalization and interpolation (Heuvelink, 2014). The quantification, visualization and communication of the uncertainty of the digital soil mapping (DSM) products would be indispensable to stakeholders (e.g. policy makers, society etc.) as it has already been stressed by the GlobalSoilMap.net initiative (Arrouays et al, 2014). Various approaches (e.g. geostatistical and machine learning) are in use to model and quantify the uncertainty of DSM products. Most of these approaches apply a probabilistic framework within which the soil attribute of interest at a single location is regarded as a realization of a random variable. Vaysse and Lagacherie (2017) applied the kriging variance among others to construct uncertainty models to various DSM products in France. The most commonly applied geostatistical approach is the kriging variance that is jointly computed with the kriging prediction (Webster and Oliver, 2007). Vaysse and Lagacherie (2017) applied the kriging variance among others to construct uncertainty models to various DSM products in France. Kempen et al (2014) applied the regression kriging variance to produce the 90% prediction interval for topsoil clay map in the Netherlands

Objectives

Methods

Results

Discussion

Conclusion