No Silver Bullet for Digital Soil Mapping: Country-specific Soil Organic Carbon Estimates across Latin America

Abstract

Abstract. Country-specific soil organic carbon (SOC) maps are the baseline for the Global SOC Map of the Global Soil Partnership (GSOCmap-GSP). This endeavor requires harmonizing heterogeneous datasets and building country-specific capacities for digital soil mapping (DSM). We identified country-specific predictors for SOC and tested the performance of five predictive algorithms for mapping SOC across Latin America. The algorithms included: support vector machines, random forest, kernel weighted nearest neighbors, partial least squares regression, and regression-Kriging based on stepwise multiple linear models. Country-specific training data and SOC predictors (5 × 5 km pixel resolution) were obtained from ISRIC-World-Soil-Information-System. In general, temperature, soil type, vegetation indices and topographic constraints were the best predictors for SOC, but country-specific predictors and their respective weights varied across Latin America. We compared a large diversity of country-specific data scenarios and were able to explain ~ 53 % of SOC variability (range

Machine learning for digital soil mapping: Applications, challenges and suggested solutions

“…A few studies (less than five) use boosted regression tree (Yang et al, 2016;Beguin et al, 2017). In addition, a number of studies use neural networks (Lamichhane et al, 2019) algorithms (Aitkenhead & Coull, 2016;Guevara et al, 2018), such as artificial neural networks (Dai et al, 2014). A relatively small number of studies use alternative algorithms such as support vector machines (Guevara et al, 2018), k -nearest neighbours (Mansuy et al, 2014) or generalized boosted regression (Tziachris et al, 2019;Gomes et al, 2019).…”

A note on knowledge discovery and machine learning in digital soil mapping

“…That is, to infer causal relationships between soil properties and forming factors and processes from the association of the former with the covariates (e.g. in Bui, Henderson, & Viergever, 2006;Guevara et al, 2018;Hengl et al, 2017;Ma, Minasny, Malone, & Mcbratney, 2019;Wiesmeier, Barthold, Blank, & Kögel-Knabner 2010;Wilford & Thomas, 2013). However, there are reasons to question the validity of our common practice of soil knowledge discovery when using ML, as previously demonstrated by Shmueli (2010) and Fourcade, Besnard, and Secondi (2018).…”

Spatial biases of information influence global estimates of soil respiration: How can we improve global predictions?

“…Recent studies have used ML (e.g., Jian, Steele, Thomas, et al, 2018; Vargas et al, 2018; Warner et al, 2019; Zhao et al, 2017) which can help improve local to large‐scale estimates by relating Rs observations to ancillary vegetation, terrain, climate, and survey data. Some ML algorithms are particularly effective in modeling biogeophysical factors due to their ability to account for nonlinear relationships and low sensitivity to autocorrelation among predictor variables relative to more traditional linear models (Guevara et al, 2018; Reichstein et al, 2019; Vargas et al, 2018). ML approaches have been shown to outperform mechanistic and semi‐empirical models in modeling carbon processes in which point‐based observations are used for regional to global estimates (Reichstein et al, 2019), which shows promise for the use of ML algorithms in future environmental modeling and benchmarking (Bond‐Lamberty, 2018).…”