Explainable Machine-Learning Framework for Crop-Water Requirement Forecasting and Canal Allocation in Large-Scale Command Areas

Abstract

Forecasting water requirements for diverse crops across vast regions is fundamental to fulfill the critical need for water resource management. Accurate predictions of crop water requirements over a designated horizon facilitate the optimal development of an effective allocation mechanism to minimize water shortages in agricultural command areas with stream sources. Toward this goal, an interpretable machine-learning framework is developed to provide short-term predictions of water requirements focusing on different crops grown in specific time slots in a large-scale command area. Predictions of water demand are calculated per crop for the next 10 days, explained, and subsequently utilized in designing an allocation strategy that minimizes redundancy while maximizing effectiveness. Redundant-supply, shortfall-risk, and blind-spots indices help gauge performance. Along with a precise prediction mechanism, these indices are crucial for practitioners such as policymakers, irrigation authorities, and water resource engineers concerned with the management of large command areas.

The prediction mechanism is optimized using important features related to meteorology, soil, and crop suitable for establishing a robust model. Important features are highlighted using a combination of various explainable artificial-intelligence tools. A general implementation strategy for a large command area is discussed, indicating the usefulness of adopting such forecasting approaches and allocation principles in irrigation projects.