Development of a Machine Learning Framework for an Irrigation Decision Support System

Authors

Eric Wilkening, University of Nebraska-LincolnFollow
Derek M. Heeren, University of Nebraska-LincolnFollow
Yeyin Shi, University of Nebraska-LincolnFollow
Laila A. Puntel, University of Nebraska-LincolnFollow
Guillermo R. Balboa, University of Nebraska-LincolnFollow
Abia Katimbo, University of Nebraska-LincolnFollow
Kuan Zhang, University of Nebraska-LincolnFollow
Precious Nneka Amori, University of Nebraska-LincolnFollow
Bruno Lena, University of Nebraska-LincolnFollow

Document Type

Article

Date of this Version

8-2025

Citation

Journal of Natural Resources and Agricultural Ecosystems (accepted)

doi: 10.13031/jnrae.16162

Comments

Open access

License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY NC-ND 4.0)

Abstract

Currently used irrigation scheduling techniques often require significant human intervention and are time consuming, particularly for variable rate irrigation. This research aimed to develop a conceptual framework that incorporates disparate data sources leveraging both mechanistic and machine learning (ML) approaches. The overarching objective is to utilize the growing availability of data and technology to manage irrigation more precisely, which will minimize negative impacts of irrigation on our water resources. An initial irrigation machine learning model was proposed and tested in this study as a key component of an edge-cloud computing and federated learning decision-making framework. The specific objectives were (1) to develop and evaluate ML models based on linear regression and random forest to derive irrigation recommendations using the latest date (LD), which is used to trigger an irrigation event; (2) to compare ML to the SETMI model for irrigation recommendations; and (3) to propose a framework for implementing ML to support variable rate irrigation and real-time irrigation management. Input data included weather variables, weather-based parameters (e.g., growing degree days), soil properties, basic agronomic information, remote sensing imagery (e.g., for soil-adjusted vegetation index), soil water sensor data, and canopy temperature from stationary infrared thermometers. Both linear regression and random forest ML algorithms were trained and evaluated on their ability to generate irrigation recommendations for maize and soybean. The latest date (LD) was defined as the number of days until an irrigation event is necessary to avoid crop stress. The RMSE of the LD when irrigation will be needed was 7 days (linear regression) and 1.5 days (random forest). The three most important variables were found to be accumulated GDDs, solar radiation, and accumulated precipitation. Implementing the model on a different field site in a subsequent year was not successful (R² = 0.01), although it did show an increase in LD after precipitation and irrigation, as expected. This research highlights the potential of ML techniques in the science of irrigation scheduling and identifies next steps for continued technology development.