Research Article
Prediction of Irrigation Water Supply Using Supervised Machine Learning Models in Koga Irrigation Scheme, Ethiopia
@INPROCEEDINGS{10.1007/978-3-030-93712-6_5, author={Menwagaw T. Damtie and Seifu A. Tilahun and Fasikaw A. Zimale and Petra Schmitter}, title={Prediction of Irrigation Water Supply Using Supervised Machine Learning Models in Koga Irrigation Scheme, Ethiopia}, proceedings={Advances of Science and Technology. 9th EAI International Conference, ICAST 2021, Hybrid Event, Bahir Dar, Ethiopia, August 27--29, 2021, Proceedings, Part II}, proceedings_a={ICAST PART 2}, year={2022}, month={1}, keywords={Canal outlet Machine learning R environment MARS}, doi={10.1007/978-3-030-93712-6_5} }- Menwagaw T. Damtie
Seifu A. Tilahun
Fasikaw A. Zimale
Petra Schmitter
Year: 2022
Prediction of Irrigation Water Supply Using Supervised Machine Learning Models in Koga Irrigation Scheme, Ethiopia
ICAST PART 2
Springer
DOI: 10.1007/978-3-030-93712-6_5
Abstract
Estimating water supply through irrigation canal distribution systems is a crucial process for better water management in the irrigation schemes. This study aimed to develop an approach to predict the discharge delivered to unregulated irrigation canals (such as quaternary canals) from geometric and hydraulic information of regulated section of the system (such as tertiary and others) using machine learning approaches. The prediction performance of four Caret-based Supervised Machine Learning Models, namely; Multivariate Adaptive Regression Splines (MARS), Artificial Neural Networks (ANN), Random Forest (RF), and Radial Basis Support Vector Machines (SVM), were developed in the R programming environment, followed by variability assessment among canal outlets at Koga irrigation Scheme. Water delivery performance at quaternary canals showed a significant flow variation among the canal outlets. The comparative study of model prediction results showed identified MARS as the optimal model, both at the training stage (RMSE = 0.074 & R2= 0.86 with normalized data) and testing stage (RMSE = 3.89 & R2= 0.85 with rescaled data). Furthermore, the model building process and output equations of MARS were relatively interpretable compared to neuro and tree-based models, such as Artificial Neural Network and Random Forest. Thus, the MARS model was recommended to estimate the water supply to ungated irrigation canals as a function of flow rate information at gated distributary canal and other field data at lower components of irrigation schemes.
