Systematic comparison of five machine-learning methods in classification and interpolation of soil particle size fractions using different transformed data

Abstract

Abstract. Soil texture and soil particle size fractions (psf) play an increasing role in physical, chemical and hydrological processes. Digital soil mapping using machine-learning methods was widely applied to generate more detailed prediction of qualitative or quantitative outputs than traditional soil-mapping methods in soil science. As compositional data, interpolation of soil psf combined with log ratio approaches was developed to improve the prediction accuracy, which also can be used to indirectly derive soil texture. However, few reports systematically analyzed and compared the classification and regression, the accuracies of original (untransformed) and log ratio approaches, and the performance of direct and indirect soil texture classification using machine-learning methods. In this total, a total of 45 evaluation models generated from five different machine-learning models combined with original and three log ratio approaches–additive log ratio, centered log ratio and isometric log ratio (ALR, CLR and ILR, respectively), to evaluate and compare the performance of soil texture classification and soil psf interpolation. The results demonstrated that log ratio approaches modified the soil sampling data more symmetrically, and with respect to soil texture classification, random forest (RF) and extreme gradient boosting (XGB) showed notable consequences. For soil psf interpolation, RF delivered the best performance among five machine-learning models with lowest root mean squared error (RMSE, sand: 15.09 %, silt: 13.86 %, clay: 6.31 %), mean absolute error (MAE, sand: 10.65 %, silt: 9.99 %, clay: 5.00 %), Aitchison distance (AD, 0.84) and standardized residual sum of squares (STRESS, 0.61), and highest coefficient of determination (R2, sand: 53.28 %, silt: 45.77 %, clay: 53.75 %). STRESS was improved using log ratio approaches, especially CLR and ILR. There is a pronounced improvement (21.3 %) in the kappa coefficient using indirect soil texture classification compared to the direct approach. Our systematic comparison helps to elucidate the processing and selection of compositional data in spatial simulation.

A review of time domain reflectometry (TDR) applications in porous media

“…To reduce subjective interpretation and to maintain constant standards, there is a need to develop new interpretation methods or tools. Machine learning or artificial intelligence has been used for graphical interpretation or target detection in various disciplines of earth science (Arabameri et al, 2020;Nanda et al, 2020;Rizvi et al, 2019;Saha et al, 2020;Zhang and Shi, 2020;Zhou et al, 2019). The machine learning may be used to facilitate the interpretation of TDR waveforms in order to determine a variety of physical properties and processes in porous media.…”

Digital mapping of soil texture classes using Random Forest classification algorithm

“…Different machine learning models were adopted by researchers for the prediction of textural classes. The most common models are support vector machine (Gomez et al, 2019;Wu et al, 2018;Zhang & Shi, 2019), k-nearest neighbour (Zhang & Shi, 2019), multinomial logistic regression (Camera et al, 2017), extreme gradient boosting (Zhang & Shi, 2019), artificial neural networks (Bagheri Bodaghabadi et al, 2015;Taalab et al, 2015) and random forest model (Camera et al, 2017;Ramcharan et al, 2018). Hengl et al (2017) reported that random forest and gradient boosting outperformed other linear models in the prediction of soil texture fractions for large data sets.…”

Mapping of Soil Total Nitrogen Content in the Middle Reaches of the Heihe River Basin in China Using Multi-Source Remote Sensing-Derived Variables

“…Gomes et al [41] found that the RF model was superior to Cubist in SOC mapping in Brazil while Sorenson et al [54] reported that Cubist performed better than RF in SOC prediction. Although these studies have different comparison results, other digital soil mapping studies in the HRB have also found that the prediction performance of RF models was better than SVM [55,56]. Based on these results, no single machine learning technique is most suitable for all landscapes.…”