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

Wadoux, Alexandre M.J.‐C.; Minasny, Budiman; McBratney, Alex B.

doi:10.31223/osf.io/8eq6s

Cited by 15 publications

(19 citation statements)

References 79 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Different machine learning techniques have been used in a large variety of environmental analyses. Examples are random forests in digital soil mapping (Sekulić et al, 2020), support vector machines (SVMs) in earthquake‐triggered landslide susceptibility (Xu et al, 2012), groundwater spring mapping with artificial neural networks (Corsini et al, 2009) and convolutional neural networks (CNNs) in land use and land cover mapping (Zhu et al, 2017) and digital soil mapping (Wadoux et al, 2020).…”

Section: Introductionmentioning

confidence: 99%

Comparing geomorphological maps made manually and by deep learning

Meij

Meijles

Marcos

et al. 2022

Earth Surf Processes Landf

View full text Add to dashboard Cite

Geomorphological maps provide information on the relief, genesis and shape of the earth's surface and are widely used in sustainable spatial developments. The quality of geomorphological maps is however rarely assessed or reported, which limits their applicability. Moreover, older geomorphological maps often do not meet current quality requirements and require updating. This updating is time-consuming and because of its qualitative nature difficult to reproduce, but can be supported by novel computational methods. In this paper, we address these issues by (1) quantifying the uncertainty associated with manual geomorphological mapping, (2) exploring the use of convolutional neural networks (CNNs) for semi-automated geomorphological mapping and (3) testing the sensitivity of CNNs to uncertainties in manually created evaluation data.We selected a test area in the Dutch push-moraine district with a pronounced relief and a high variety of landforms. For this test area we developed five manually created geomorphological maps and 27 automatically created landform maps using CNNs. The resulting manual maps are similar on a regional level. We could identify the causes of disagreement between the maps on a local level, which often related to differences in mapping experience, choices in delineation and different interpretations of the legend. Coordination of mapping efforts and field validation are necessary to create accurate and precise maps. CNNs perform well in identifying landforms and geomorphological units, but fail at correct delineation. The human geomorphologist remains necessary to correct the delineation and classification of the computed maps. The uncertainty in the manually created data that are used to train and evaluate CNNs have a large effect on the model performance and evaluation. This also advocates for coordinated mapping efforts to ensure the quality of manually created training and test data. Further model development and data processing are required before CNNs can act as standalone mapping techniques.

show abstract

Section: Introductionmentioning

confidence: 99%

Comparing geomorphological maps made manually and by deep learning

Meij

Meijles

Marcos

et al. 2022

Earth Surf Processes Landf

View full text Add to dashboard Cite

show abstract

“…This could probably lead to non‐linearity in the HSI spectra. Compared with linear PLS models, ML techniques are advantageous because they do not depend on the assumption that data should be drawn from a given probability distribution (Wadoux et al, 2020). In terms of SVMR, its excellent performances in prediction accuracy and computational efficiency can be explained as follows.…”

Section: Discussionmentioning

confidence: 99%

“…A main concern about the developed SVMR approach might be its generalisation ability for other soil types. This concern arises mainly because ML techniques are data‐driven approaches that rely directly on calibration data for information extraction (Wadoux et al, 2020). To obtain an acceptable generalisation of SVMR in other regions, more soil types and more heterogeneous profiles should be collected in future studies.…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

A comparison of machine learning algorithms for mapping soil iron parameters indicative of pedogenic processes by hyperspectral imaging of intact soil profiles

Zhao

Wang

et al. 2021

European J Soil Science

View full text Add to dashboard Cite

Soil iron (Fe) performs vital functions in the biogeochemical cycles of soil environments. The amount and profile allocation of various Fe parameters can be used as sensitive indicators of soil development and pedogenic processes. This study aimed to evaluate the potential of ground‐based hyperspectral imaging (HSI: 400–1010 nm) spectroscopy to predict and map six Fe parameters indicative of pedogenic processes: total Fe (Fet), dithionite‐citrate‐bicarbonate (DCB)‐extracted Fe (Fed), oxalate‐extracted Fe (Feo), weathering index (FeW), active ratio (FeA) and crystallinity ratio (FeC). In total, 17 intact soil profiles at a depth of 100 ± 5 cm were collected to acquire HSI images. Four non‐linear machine learning techniques, namely, random forest (RF), XGBoost, CatBoost and support vector machine regression (SVMR), were implemented and compared with linear partial least squares (PLS) to identify the models with the best performance for different soil Fe parameters. Our results indicate that the four non‐linear machine learning models outperformed PLS for most soil Fe parameters, with low root mean square error (RMSE) and high Lin's concordance correlation coefficient (LCCC) values. Based on the testing set, SVMR showed better performance over the other tested models for Fet (RMSEP = 2.645 g kg−1, LCCCP = 0.89), Fed (RMSEP = 0.972 g kg−1, LCCCP = 0.97), Feo (RMSEP = 0.273 g kg−1, LCCCP = 0.97), FeW (RMSEP = 0.035, LCCCP = 0.97), FeA (RMSEP = 0.033, LCCCP = 0.97) and FeC (RMSEP = 0.031, LCCCP = 0.97). According to the LCCCP values, soil Fet was predicted to be in substantial agreement by SVMR, and the other soil Fe predictions were considered to be in near perfect agreement. Moreover, SVMR required lower computational costs. Given these results, the combination of HSI spectroscopy and SVMR is recommended due to its more reliable estimation and profile mapping of the selected soil Fe parameters than that of PLS, RF, XGBoost and CatBoost. Highlights Hyperspectral imaging was used to map six soil Fe parameters of intact profiles. Nonlinear machine learning models were compared to select the most suitable model. Nonlinear techniques outperformed the linear PLS model in most cases. SVMR showed higher comprehensive performance than other models. Hyperspectral imaging combined with SVMR is recommended to map soil Fe in profiles.

show abstract

“…While valid and useful to obtain insights into complex models of soil variation, these methods are model-specific, i.e. they preclude comparison between models (Wadoux et al, 2020a). A number of "model-agnostic" or model-independent interpretation methods have recently been developed outside soil science, in the statistical and machine learning literature.…”

Section: Introductionmentioning

confidence: 99%

Beyond prediction: methods for interpreting complex models of soil variation

Wadoux¹,

Molnar²

2021

Preprint

Self Cite

View full text Add to dashboard Cite

Understanding the spatial variation of soil properties is central to many sub-disciplines of soil science. Commonly in soil mapping studies, a soil map is constructed through prediction by a statistical or non-statistical model calibrated with measured values of the soil property and environmental covariates of which maps are available. In recent years, the field has gradually shifted attention towards more complex statistical and algorithmic tools from the field of machine learning. These models are particularly useful for their predictive capabilities and are often more accurate than classical models, but they lack interpretability and their functioning cannot be readily visualized. There is a need to understand how these these models can be used for purposes other than making accurate prediction and whether it is possible to extract information on the relationships among variables found by the models. In this paper we describe and evaluate a set of methods for the interpretation of complex models of soil variation. An overview is presented of how model-independent methods can serve the purpose of interpreting and visualizing different aspects of the model. We illustrate the methods with the interpretation of two mapping models in a case study mapping topsoil organic carbon in France. We reveal the importance of each driver of soil variation, their interaction, as well as the functional form of the association between environmental covariate and the soil property. Interpretation is also conducted locally for an area and two spatial locations with distinct land use and climate. We show that in all cases important insights can be obtained, both into the overall model functioning and into the decision made by the model for a prediction at a location. This underpins the importance of going beyond accurate prediction in soil mapping studies. Interpretation of mapping models reveal how the predictions are made and can help us formulating hypotheses on the underlying soil processes and mechanisms driving soil variation.

show abstract

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

Cited by 15 publications

References 79 publications

Comparing geomorphological maps made manually and by deep learning

Comparing geomorphological maps made manually and by deep learning

A comparison of machine learning algorithms for mapping soil iron parameters indicative of pedogenic processes by hyperspectral imaging of intact soil profiles

Beyond prediction: methods for interpreting complex models of soil variation

Contact Info

Product

Resources

About