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

Wadoux, Alexandre M.J.‐C.; Samuel‐Rosa, Alessandro; Poggio, Laura; Mulder, Vera Leatitia

doi:10.1111/ejss.12909

Cited by 77 publications

(62 citation statements)

References 21 publications

(26 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…variables that are indicative for potential drivers of soil organic C, and having a wide coverage of these potential drivers in the training data facilitates the interpretability of Random Forest models. & Random Forests can only reveal associations, not causality (Wadoux et al 2020).…”

Section: Limits Of Random Forest Algorithms In Explaining Soil Organic Cmentioning

confidence: 99%

Origin of carbon in agricultural soil profiles deduced from depth gradients of C:N ratios, carbon fractions, δ13C and δ15N values

2020

View full text Add to dashboard Cite

Aims Agricultural soils in Germany store 2.54 Pg of organic carbon (C). However, information about how and when this C entered the soils is limited. This study illustrates how depth profiles of organic matter can shed light on different entry paths of organic C. Methods Machine learning was used to explain total organic C (TOC), C:N, particulate organic C (POC), δ13C and δ15N values down to 100 cm depth based on pedology, geology, climate and management-related variables from the German Agricultural Soil Inventory. We estimated TOC turnover rates based on the relationship between the proportion of maize (only C4 plant) in crop rotations and soil δ13C values. Results In the upper 30 cm of cropland, fresh photosynthates added on average 0.2 to 0.8 Mg C ha− 1 year− 1. Organic fertiliser was another source of topsoil C, especially in grassland. Sandy sites in north-west Germany contained historic C from past heathland and peatland. One third of German agricultural land was found to be on colluvial and alluvial deposits, in which allochthonous C from upstream and upslope areas evidently increased the TOC content of subsoils. In and below hardpans, TOC content and C:N and POC:TOC ratios were low, indicating restricted root-derived C input. Conclusions Our data indicate that ongoing management in German agricultural soils mainly affects topsoil C, while C storage in subsoils reveals significant legacies from allochthonous, buried or translocated C inputs. Specific attention should be focused on the sustainable loosening of hardpans that could result in a slow, but significant increase in subsoil C stocks.

show abstract

Section: Limits Of Random Forest Algorithms In Explaining Soil Organic Cmentioning

confidence: 99%

Origin of carbon in agricultural soil profiles deduced from depth gradients of C:N ratios, carbon fractions, δ13C and δ15N values

2020

View full text Add to dashboard Cite

show abstract

“…It is possible to assess the importance of each covariate by shuffling the values of that covariate amongst the observation locations and calculating the reduction in prediction accuracy. However, Schmidt et al (2020) and Wadoux et al (2020) advise caution when inferring causal relationships from random forest models. Wadoux et al (2020) demonstrate that photographs of soil scientists projected across their study area can be utilised by a random forest to accurately map the soil carbon content.…”

Section: Machine Learningmentioning

confidence: 99%

How is Baseflow Index (BFI) impacted by water resource management practices?

Bloomfield¹,

Gong²,

Marchant³

et al. 2021

Preprint

View full text Add to dashboard Cite

Abstract. Water resource management (WRM) practices, such as abstractions and discharges, may impact baseflow. Here the CAMELS-GB large-sample hydrology dataset is used to assess the impacts of such practices on baseflow index (BFI) using statistical models of 429 catchments from Great Britain. Two complementary modelling schemes, multiple linear regression (LR) and machine learning (random forests, RF), are used to investigate the relationship between BFI and two sets of covariates (natural covariates only and a combined set of natural and WRM covariates). The LR and RF models show good agreement between explanatory covariates. In all models, the extent of fractured aquifers, clay soils, non-aquifers, and crop cover in catchments, catchment topography and aridity are significant or important natural covariates in explaining BFI. When WRM terms are included, groundwater abstraction is significant or the most important WRM covariate in both modelling schemes and discharge to rivers is also identified as significant or influential, although natural covariates still provide the main explanatory power of the models. Surface water abstraction is a significant covariate in the LR model but of only minor importance in the RF model. Reservoir storage covariates are not significant or are unimportant in both the LR and RF models for this large-sample analysis. Inclusion of WRM terms improves the performance of some models in specific catchments. The LR models of high BFI catchments with relatively high levels of groundwater abstraction show the greatest improvements, and there is some evidence of improvement in LR models of catchments with moderate to high discharges. However, there is no evidence that the inclusion of the WRM covariates improves the performance of LR models for catchments with high surface water abstraction or that they improve the performance of the RF models. These observations are used to formulate a conceptual framework for baseflow generation that incorporates WRM practices. It is recommended that information on WRM, particularly groundwater abstraction, should be included where possible in future large-sample hydrological data sets and in the analysis and prediction of BFI and other measures of baseflow.

show abstract

“…Opening the "black box" is then necessary but not straightforward (see next section on interpretability), and is often reduced to the analysis of which environmental covariates are the most often used by the model to make a prediction (see for example Mahmoudabadi et al (2017) that ML algorithm should not be used for obtaining new soil knowledge because the ML algorithm aims at predicting a pattern rather than finding causal relationships. Wadoux et al (2019c) suggest to use calibrated ML models as a "hypothesis discovery" tool, in which the mechanisms conveyed by the calibrated ML model are supplied to the researcher for possible explanations of the soil process, which can then be confronted to experiments and principles of soil genesis. The challenge that then arises, noticed by Gahegan (2019) is the conversion of the mechanisms of the ML model (the model "knowledge") from a data language to a human one.…”

Section: Machine Learning and Pedological Knowledgementioning

confidence: 99%

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

Wadoux¹,

Minasny²,

McBratney³

2020

Preprint

Self Cite

View full text Add to dashboard Cite

The uptake of machine learning (ML) algorithms in digital soil mapping (DSM) is transforming the way soil scientists produce their maps. Machine learning is currently applied to mapping soil properties or classes much in the same way as other unrelated fields of science. Mapping of soil, however, has unique aspects which require adaptations of the ML algorithms. These features are for example, but not limited to, the inclusion of pedological knowledge into the ML algorithm, the accounting of spatial structure present in the soil data, or the desire to increase our scientific understanding of the distribution and genesis of soil from a calibrated ML model. Tackling these challenges is critical for machine learning to gain credibility and scientific consistency in soil science. In this article, we review the current applications of machine learning in digital soil mapping and suggest improvements. We found a growing interest of the use of ML in DSM. Most studies focus on obtaining accurate maps and disregard the characteristics of soil data, such as spatial autocorrelation. Only a few studies account for existing soil knowledge or quantify the uncertainty of the predicted maps. We then discuss the challenges related to the application of ML for soil mapping and offer solutions from existing studies in the natural sciences. The challenges are organized as follows: sampling, resampling, accounting for the spatial information, multivariate mapping, uncertainty analysis, validation, integration of pedological knowledge and, interpretation of the models. We conclude that for future developments, machine learning should incorporate three core elements: plausibility, interpretability, and explainability, which will trigger soil scientists to move beyond model prediction and towards explanation of soil processes.

show abstract

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

Cited by 77 publications

References 21 publications

Origin of carbon in agricultural soil profiles deduced from depth gradients of C:N ratios, carbon fractions, δ13C and δ15N values

Origin of carbon in agricultural soil profiles deduced from depth gradients of C:N ratios, carbon fractions, δ13C and δ15N values

How is Baseflow Index (BFI) impacted by water resource management practices?

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

Contact Info

Product

Resources

About