Publication Date

12-2025

Date of Final Oral Examination (Defense)

7-25-2025

Type of Culminating Activity

Thesis

Degree Title

Master of Science in Civil Engineering

Department

Civil Engineering

Supervisory Committee Chair

Mojtaba Sadegh, Ph.D.

Supervisory Committee Member

David Hoekema, Ph.D.

Supervisory Committee Member

Kevin Roche, Ph.D.

Abstract

Idaho is the third largest water user in the United States due to its extensive agriculture industry (Murray, 2018). Irrigation has allowed farmers to take advantage of the warm climate and rich soils of the Eastern Snake River Plain to develop one of the nation’s most productive agricultural regions. This area now contains one-third of Idaho’s irrigated acres (Idaho Department of Water Resources, 2024). It also contains critical hydrologic connections between the Snake River and the Eastern Snake Plain Aquifer.

Farming practices in Idaho have gradually shifted over time. Warming temperatures and increased irrigation efficiency are allowing an expansion of high-water use crops and higher economic value crops in the Eastern Snake River Plain. Historically, irrigation was achieved through flood and furrow methods, that allowed water to be distributed by gravity. Since the 1940s, farmers have increasingly transitioned to pressurized irrigation systems. Pressurized irrigation has improved irrigation efficiency at the expense of reduced groundwater recharge and lower return flows to the Snake River for downstream use. Return flows are the portion of water applied for irrigation that returns to surface water bodies systems after application, representing a critical component of regional water budgets. Improved efficiency often results in less return flows, higher crop yields, and additional cropping acreage rather than a reduction in water withdrawal, further pressurizing the already stressed water resources.

This thesis aimed to estimate irrigation return flows by developing generalizable machine learning models, and focused on Twin Falls Canal Company (TFCC), which is one of the largest canal companies in the Snake River Plain of southern Idaho. Estimating irrigation return flows is crucial for accurate water planning and budgeting and allows the canal company to optimize water distribution, reduce water waste, and ensure sustainable water management within their service area.

A deep neural network model, dubbed ReFlow-Net, was developed to predict return flows with 1, 7, 14, and 30 days lead time. The ReFlow-Net model uses climate, land cover, diversion flows, and observed return flow data to learn patterns and provide return flows. Phase 1 of the model receives inputs that stop at day n, and predictions are conducted at days n+1, n+7, n+14, n+30, achieving R^2 values of 93%, 60%, 62%, and 39%, respectively. Phase 2 saw major improvements, as it receives inputs to the day prior to that of return flow prediction (n, n+6, n+13, n+29), and achieves R^2 values of 94%, 86%, 82%, and 77%, respectively, noting that future weather forecasts are generally reliable up to two weeks and longer lead time forecasts are mainly a scientific exploration at the time of this writing. This research is one of the pioneering studies of deep learning-based return flow prediction and is expected to improve water budgeting for the region and support sustainable water management efforts by accounting for return flows as the study area continues to experience climate variability and change, and population growth.

DOI

https://doi.org/10.18122/td.2452.boisestate

Download

Share

COinS