Developing machine learning models to predict water retention properties of soils of Sri Lanka

Abstract

Measuring soil water retention properties, such as volumetric water content (VWC) at field capacity (FC) and permanent wilting point (PWP), is a labor-intensive process, resulting in insufficient data for accurate irrigation scheduling. This study aimed to develop machine learning (ML) models to predict VWC at 10 (VWC10), 33, (VWC33), and 1500 kPa (VWC1500) of Sri Lankan soils using a data set of an extensive soil survey. Soil properties such as soil texture, bulk density (BD), soil organic carbon (SOC) were used with Multiple Linear Regression (MLR), K-Nearest Neighbor (KNN), Random Forest (RF), and Support Vector Regression (SVR) models. The predictive accuracy was evaluated using the coefficient of determination (R2), Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Square Error (RMSE) across the entire dataset and various textural classes. Among the models tested, MLR outperformed ML models, achieving a satisfactory accuracy (R2 = 0.63) in predicting VWC10 and VWC33 with the sand, silt, and bulk density combination. Furthermore, the combinations of sand and VWC10 and sand and VWC33 notably boosted the predictive accuracy of VWC1500 across all tested models (RMSE 3-4 %), with SVM and RF emerging as the top performers with R2 values of 0.85 and 0.90, respectively. Nevertheless, when silt, SOC, and BD were added to the sand and VWC10 or sand and VWC33 combination, there was only a slight improvement in the predictive accuracy of VWC1500. Results of the study revealed that the SVM and RF could be used to predict the VWC1500 of soils of Sri Lanka with higher accuracy using combinations of easily measured parameters like sand content and VWC10 or sand content and VWC33. Future research will aim to enhance the prediction models for VWC10 and VWC33 by including a larger data set covering diverse soil types.