Assessing Machine-Learning Performances in Predicting Future Irrigation Demand for Large-Scale Rice Granaries Based on Global Climate Models

Future climate prediction at a local scale is one of the pressing challenges affecting water management-related mitigation plans. Modeling crop irrigation demands under possible climate change will require multiple monotonous and time-consuming steps. This study focuses on evaluating the performance of machine learning models such as support vector regression (SVR), random forest (RF), and a meta-ensemble model (meta SVR-RF) in predicting future rice irrigation demand for the Kerian irrigation scheme based on global climate model (GCM) data as alternative techniques. The research analyzes the accuracy of these models using historical climate records from 1976 to 2005 and future records from 2021 to 2080, with various statistical metrics employed to assess their precision, such as coefficient correlation (R2), Kling-Gupta efficiency (KGE), mean absolute error (MAE), and root mean square error (RMSE). Meta SVR-RF model exhibited superior performance during the training and testing phases compared to individual models with R2, KGE, MAE, and RMSE values 0.15% better than SVR and 1.25% better than RF, 2.82% better than SVR and 7.82% better than RF, 37.48% better than SVR and 53.13% better than RF, 29.86% better than SVR and 50.24% better than RF, respectively. While SVR performance improved in the testing phase by 0.10% of R2, 0.84% of KGE, and 6.83% of RMSE. The study identifies the potential factors affecting model performance, particularly for SVR, and emphasizes the benefits of ensemble techniques. Moreover, it concludes that machine learning models offer a reliable approach to estimating irrigation demand under future climate scenarios, simplifying the traditionally complex water balance calculation process, and highlighting their applicability in climate impact studies.