Selecting appropriate machine learning methods for digital soil mapping

• Published in: Applied Mathematical Modelling Vol. 81; pp. 401 - 418
• Main Authors: Khaledian, Yones, Miller, Bradley A.
• Format: Journal Article
• Language: English
• Published: New York Elsevier Inc 01.05.2020; Elsevier BV
• Subjects: Algorithm selection; Algorithms; Artificial neural networks; Datasets; Digital mapping; Digital soil mapping; Learning theory; Machine learning; Model evaluation; Soil mapping; Soil properties; Soils; Spatial distribution; Support vector machines; Algorithm selection; Model evaluation; Digital soil mapping; Machine learning
• ISSN: 0307-904X; 1088-8691; 0307-904X
• DOI: 10.1016/j.apm.2019.12.016

•Selecting the appropriate machine learning method depends on the digital soil mappers’ purpose.•Artificial neural network is strong with large sample sizes, but is black box.•Cubist produces interpretable results; however, Random Forests’ results are semi interpretable.•R2 is more sensitive to outliers than RMSE.•Independent validation is necessary to evaluate the predictive power of the model. Digital soil mapping (DSM) increasingly makes use of machine learning algorithms to identify relationships between soil properties and multiple covariates that can be detected across landscapes. Selecting the appropriate algorithm for model building is critical for optimizing results in the context of the available data. Over the past decade, many studies have tested different machine learning (ML) approaches on a variety of soil data sets. Here, we review the application of some of the most popular ML algorithms for digital soil mapping. Specifically, we compare the strengths and weaknesses of multiple linear regression (MLR), k-nearest neighbors (KNN), support vector regression (SVR), Cubist, random forest (RF), and artificial neural networks (ANN) for DSM. These algorithms were compared on the basis of five factors: (1) quantity of hyperparameters, (2) sample size, (3) covariate selection, (4) learning time, and (5) interpretability of the resulting model. If training time is a limitation, then algorithms that have fewer model parameters and hyperparameters should be considered, e.g., MLR, KNN, SVR, and Cubist. If the data set is large (thousands of samples) and computation time is not an issue, ANN would likely produce the best results. If the data set is small (<100), then Cubist, KNN, RF, and SVR are likely to perform better than ANN and MLR. The uncertainty in predictions produced by Cubist, KNN, RF, and SVR may not decrease with large datasets. When interpretability of the resulting model is important to the user, Cubist, MLR, and RF are more appropriate algorithms as they do not function as “black boxes.” There is no one correct approach to produce models for predicting the spatial distribution of soil properties. Nonetheless, some algorithms are more appropriate than others considering the nature of the data and purpose of mapping activity.