Modelling soil water content in different tillage systems and soil types using machine learning

Abstract

Soil water availability is one of the major challenges in many rainfed crop production systems of the Global South. Soil water conservation practices are being promoted to enhance climate change adaptation for rainfed cropping systems of southern Africa. However, the cost and time required to develop and test appropriate modelling and simulation tools can be enormous. The objectives of this study were to: (i) test the performance of the decision tree, adaptive boosting (AdaBoost), support vector machine, neural network, stochastic gradient descent, k-nearest neighbours, random forest and linear regression machine learning models in predicting soil water under different tillage practices, soil types and depths, and (ii) assess the soil water classification and prediction capabilities of 8 models under different tillage practices, soil types and depths. The neural network, random forest and decision tree models had the best soil water prediction capabilities. The neural network, random forest and decision tree models were the best algorithms (RMSE = 15.801–16.369; MAE = 11.997–12.315; R2 = 0.822–0.835) for predicting and classifying soil water from different soil types and depth intervals. The support vector machine learning model was the weakest algorithm (RMSE = 36.177; MAE = 30.84; R2 = 0.133) for predicting and classifying soil water. All the algorithms poorly predicted and classified soil water based on tillage practices. All the models closely predicted soil water at 300 and 900 mm depths but poorly predicted soil water at 600 mm depth intervals. Based on this study, the neural network model is the best machine learning tool for predicting soil water in clay and sandy soils under semi-arid agroecological conditions.