Ensemble Machine Learning Approach Improves Predicted Spatial Variation of Surface Soil Organic Carbon Stocks in Data-Limited Northern Circumpolar Region.

Umakant Mishra,William J Riley,Forrest M Hoffman,Sagar Gautam

doi:10.3389/fdata.2020.528441

Abstract

Various approaches of differing mathematical complexities are being applied for spatial prediction of soil properties. Regression kriging is a widely used hybrid approach of spatial variation that combines correlation between soil properties and environmental factors with spatial autocorrelation between soil observations. In this study, we compared four machine learning approaches (gradient boosting machine, multinarrative adaptive regression spline, random forest, and support vector machine) with regression kriging to predict the spatial variation of surface (0–30 cm) soil organic carbon (SOC) stocks at 250-m spatial resolution across the northern circumpolar permafrost region. We combined 2,374 soil profile observations (calibration datasets) with georeferenced datasets of environmental factors (climate, topography, land cover, bedrock geology, and soil types) to predict the spatial variation of surface SOC stocks. We evaluated the prediction accuracy at randomly selected sites (validation datasets) across the study area. We found that different techniques inferred different numbers of environmental factors and their relative importance for prediction of SOC stocks. Regression kriging produced lower prediction errors in comparison to multinarrative adaptive regression spline and support vector machine, and comparable prediction accuracy to gradient boosting machine and random forest. However, the ensemble median prediction of SOC stocks obtained from all four machine learning techniques showed highest prediction accuracy. Although the use of different approaches in spatial prediction of soil properties will depend on the availability of soil and environmental datasets and computational resources, we conclude that the ensemble median prediction obtained from multiple machine learning approaches provides greater spatial details and produces the highest prediction accuracy. Thus an ensemble prediction approach can be a better choice than any single prediction technique for predicting the spatial variation of SOC stocks.

Highlights

High latitude permafrost region soils store large stocks of soil organic carbon (SOC) due to multiple cryopedogenic processes operating over long time scales (Ping et al, 2008; Tarnocai et al, 2009; Hugelius et al, 2014; Ping et al, 2015)
We found that SOC stock predictions from two machine learning approaches (GBM and RF) and regression kriging have comparable prediction accuracies
The uncertainty in surface SOC stocks was less than 20% in about half of the study area

Summary

Introduction

High latitude permafrost region soils store large stocks of soil organic carbon (SOC) due to multiple cryopedogenic processes operating over long time scales (Ping et al, 2008; Tarnocai et al, 2009; Hugelius et al, 2014; Ping et al, 2015). Current earth system models show large uncertainty both in baseline SOC stock representations and their release to the atmosphere under changing climate (Mishra et al, 2013; Schuur et al, 2015; McGuire et al, 2016). Reliable estimates of the magnitude and spatial variation of permafrost region SOC stocks are essential to better understand the environmental controls and to reduce the uncertainty in predicting permafrost region carbon -climate feedbacks. The magnitude of SOC stored in the soil per unit of land area is highly variable in permafrost region soils (Mishra and Riley, 2015; Mishra et al, 2017), as SOC stocks depend on various environmental factors such as soil type, land use, topographic features, and climatic conditions, which are site specific. Knowledge of soil and site-specific environmental controllers is essential to make reliable spatial predictions of SOC stocks

Methods

Results

Discussion

Conclusion