Hasan et al. (2025) Satellite data and physics-constrained machine learning for estimating effective precipitation in the Western United States and application for monitoring groundwater irrigation

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-09-22
• Authors: Md Fahim Hasan, Ryan Smith, Sayantan Majumdar, Justin Huntington, Antônio Alves Meira Neto, Blake Minor
• DOI: 10.1016/j.agwat.2025.109821

Research Groups

• Department of Civil and Environmental Engineering, Colorado State University, Fort Collins, USA
• Division of Hydrologic Sciences, Desert Research Institute, Reno, USA

Short Summary

This study developed a physics-constrained machine learning framework to accurately estimate effective precipitation for irrigated croplands in the Western United States at a spatial resolution of approximately 2 kilometers and a monthly temporal scale from 2000 to 2020. The framework's effective precipitation estimates were successfully integrated into a water balance model to monitor groundwater irrigation, showing good skill with an R² of 0.78 and a PBIAS of –15 % when compared to in-situ pumping records.

Objective

• To develop a physics-constrained machine learning framework that accurately predicts monthly and growing season effective precipitation for irrigated croplands of the Western United States at approximately 2 kilometers spatial resolution from 2000 to 2020.
• To analyze the factors influencing effective precipitation dynamics in irrigated landscapes.
• To assess effective precipitation estimates by applying them in a water balance model to estimate groundwater pumping for irrigation in seven basins of the Western United States and validate these estimates against in-situ records.

Study Configuration

• Spatial Scale: Irrigated croplands across 17 states of the Western United States, with a spatial resolution of approximately 2 kilometers (2200 meters). Validation was conducted over seven specific basins.
• Temporal Scale: Monthly and growing season estimates from 2000 to 2020.

Methodology and Data

• Models used:
  • Physics-Constrained Machine Learning (PCML) framework, consisting of a monthly effective precipitation model and a water year effective precipitation fraction model.
  • Distributed Random Forests (DRF) algorithm (from Light Gradient Boosting Machine package) for both PCML steps.
  • Water balance model for estimating consumptive groundwater use and pumping.
• Data sources:
  • Satellite: OpenET ensemble actual evapotranspiration (ET) product (30 meters), IrrMapper, LANID, and AIM-HPA irrigated cropland datasets, Crop Data Layer (CDL), Tree cover (Sexton et al., 2013).
  • Modeled/Reanalysis: Precipitation, ASCE Grass Reference ET (RET), maximum relative humidity, downward shortwave radiation (Abatzoglou, 2013), daylight duration (Thornton et al., 2022), available water capacity (Boiko et al., 2021), sand content, field capacity (Hengl, 2018a,b), elevation (Farr et al., 2007), maximum temperature (PRISM), aridity index, precipitation intensity. USGS HUC12-scale surface water irrigation dataset (Haynes et al., 2023).
  • In-situ: Annual groundwater pumping records from state agencies.

Main Results

• The monthly effective precipitation model achieved strong performance on the holdout dataset with an R² of approximately 0.923, a Root Mean Square Error (RMSE) of approximately 9.14 millimeters per month, and a Mean Absolute Error (MAE) of approximately 6.46 millimeters per month.
• Effective precipitation estimates show an eastward increasing trend across the Western United States, consistent with climatic gradients. Median monthly effective precipitation ranged from 12 millimeters per month in arid regions (e.g., Harquahala INA, Arizona) to 39 millimeters per month in semi-arid regions (e.g., GMD3, Kansas).
• Feature importance analysis revealed that reference ET, daylight duration, maximum relative humidity, and precipitation-related variables are the primary drivers of effective precipitation, while soil properties have relatively lower importance.
• Groundwater pumping estimates derived from the framework showed good agreement with in-situ records, with an R² of 0.78 and a PBIAS of –14.95 % for volumetric comparisons, and an R² of 0.61 and a PBIAS of –17.07 % for depth comparisons.
• The framework successfully filled critical data gaps by providing pumping estimates for years with missing in-situ records in basins like the Republican River Basin and Diamond Valley.
• Comparison of PCML effective precipitation estimates with the ET Demands soil water balance model in the Oregon Hydrographic Area showed strong agreement (R² ≈ 0.84, RMSE ≈ 53 millimeters per year).

Contributions

• Developed a novel two-step physics-constrained machine learning framework for effective precipitation estimation, incorporating water balance principles to ensure physical realism.
• Generated a high spatio-temporal resolution (approximately 2 kilometers, monthly) effective precipitation dataset for the entire Western United States from 2000 to 2020, addressing a critical data gap.
• Provided insights into the complex interplay of climatic, hydrologic, and soil factors governing effective precipitation dynamics in irrigated landscapes through feature importance and accumulated local effects (ALE) analysis.
• Validated effective precipitation estimates indirectly by demonstrating their successful application in a water balance model to accurately predict groundwater pumping, complementing existing in-situ records.
• Presented a flexible, computationally efficient, and easily implementable framework that leverages cloud computing resources for large-scale remote sensing data processing.
• Highlighted a new application of OpenET satellite-based ET products for quantifying effective precipitation and consumptive use of irrigation water, enhancing water management capabilities.

Funding

• National Aeronautics and Space Administration (NASA) (award # 80NSSC21K0979)
• U.S. Geological Survey Water Resources Research Institute (grant G22AC00584-00)

Citation

@article{Hasan2025Satellite,
  author = {Hasan, Md Fahim and Smith, Ryan and Majumdar, Sayantan and Huntington, Justin and Neto, Antônio Alves Meira and Minor, Blake},
  title = {Satellite data and physics-constrained machine learning for estimating effective precipitation in the Western United States and application for monitoring groundwater irrigation},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.109821},
  url = {https://doi.org/10.1016/j.agwat.2025.109821}
}