Digital mapping for soil texture class prediction in northwestern Türkiye by different machine learning algorithms

• Published in: Geoderma Regional Vol. 31; p. e00584
• Main Authors: Kaya, Fuat, Başayiğit, Levent, Keshavarzi, Ali, Francaviglia, Rosa
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2022
• Subjects: agroecology; decision support systems; digital elevation models; Digital soil mapping; Environmental covariates; Fluvisols; lakes; Landsat; Machine learning; prediction; remote sensing; soil management; soil texture; Soil texture class; support vector machines; texture; topography; topsoil; uncertainty; Vertisols; Environmental covariates; Vertisols; Fluvisols; Digital soil mapping; Soil texture class; Machine learning
• ISSN: 2352-0094; 2352-0094
• DOI: 10.1016/j.geodrs.2022.e00584

Soil texture classes (STCs) influence the physical, chemical and biological properties of the soil, and accurate spatial predictions of STCs are essential for agro-ecological modeling. The purpose of this study was to assess the capabilities of environmental covariates derived from Landsat 8 OLI Science products and a digital elevation model (DEM), as well as three machine learning methods, to provide an accurate and reliable classification of soil texture classes. Estimation efficiency of soil texture classes was investigated using Decision tree (DT), Random forest (RF), and Support vector machine (SVM) algorithms in an area with Fluvisols and Vertisols as predominant reference soil groups in Northwestern Türkiye (Lake of Manyas). The models were validated using the leave-one-out, cross-validation technique. The best of the three machine learning models for soil texture classification was RF, with an overall accuracy of 0.63 and a kappa index value of 0.14, according to the accuracy evaluation. The RF algorithm generated a map whose findings were more consistent with the real environment using the confusion index (RF:0.30) and abundance index (RF:1.0) values as uncertainty criteria. The most important predictor of soil texture classes was the topographic wetness index (mean decrease accuracy: 3%) for tree-based learning algorithms, followed by other indexes based on satellite data. These findings will contribute to support sustainable soil management techniques in a location with a large degree of topsoil texture diversity. [Display omitted] •Three machine learning techniques were used for digital mapping of soil texture classes.•Overall accuracy bias was measured by confusion and abundance index values.•The support vector machine model overestimated the dataset's majority class.•Random forest maps had the best overall accuracy and abundance index.•Topographic wetness index was the most important predictor in the area.