ABSTRACT:

Efforts to monitor groundwater pumping for irrigation in the Western United States (Western US) are hindered by the lack of comprehensive pumping records. While previous studies have developed region-specific machine learning models using limited datasets, these models are often not transferable across regions, and a groundwater pumping dataset that goes beyond local and state boundaries remains missing. In this study, we develop a regional-scale, data-driven machine learning framework to address these limitations by integrating remote sensing datasets and in-situ pumping records from Arizona, Colorado, Kansas, and Nevada. Using gridded hydroclimatic and land use variables, including effective precipitation, fraction of irrigated croplands, and evapotranspiration, the model generates spatially continuous, high-resolution (2 km, annual) historical groundwater pumping estimates from 2000 to 2023 for groundwater-dominated basins of the Western US, while predicting total irrigation in conjunctive basins. The model demonstrates good predictive performance under randomized split, with an R2 = 0.62, NRMSE = 0.50, NMAE = 0.34, and PBIAS = 8.59% on the test set. Model evaluation over groundwater-dominated and conjunctive basins across the region shows satisfactory results. In addition, comparisons using spatial holdout analysis and power consumption-based pumping records in multiple basins indicate strong generalization capacity and spatial transferability within the study region. Our assessment identifies limited availability of in-situ pumping records and lack of surface water irrigation datasets as the primary constraints for further advancing such regional-scale frameworks. Overall, the findings highlight that regional transferability of machine learning models for predicting groundwater irrigation is achievable but contingent on holistic representation of the hydrologic system.

GEE asset - projects/ee-westus-pumping/assets/westus_pumping
GEE code example - https://code.earthengine.google.com/a439de4785b9123e7255c3471ff8ed15
GitHub repository - https://github.com/mdfahimhasan/WestUS_pumping

ABSTRACT:

This repository hosts the monthly effective precipitation datasets for the irrigated croplands of the Western United States at 2 km spatial resolution from 2000 to 2020.

Related article: Hasan, M. F., Smith, R. G., Majumdar, S., Huntington, J. L., Alves Meira Neto, A., & Minor, B. A. (2025). Satellite data and physics-constrained machine learning for estimating effective precipitation in the Western United States and application for monitoring groundwater irrigation. Agricultural Water Management, 319, 109821. https://doi.org/10.1016/j.agwat.2025.109821

GitHub repo: https://github.com/mdfahimhasan/WestUS_Peff

ABSTRACT:

Groundwater overdraft gives rise to multiple adverse impacts including land subsidence and permanent groundwater storage loss. Existing methods are unable to characterize groundwater storage loss at the global scale with sufficient resolution to be relevant for local studies. Here we explore the interrelation between groundwater stress, aquifer depletion, and land subsidence using remote sensing and model-based datasets with a machine learning approach. The developed model predicts global land subsidence magnitude at high spatial resolution (~2 km), provides a first-order estimate of aquifer storage loss due to consolidation of ~17 km3/year globally, and quantifies key drivers of subsidence. Roughly 73% of the mapped subsidence occurs over cropland and urban areas, highlighting the need for sustainable groundwater management practices over these areas. The results of this study aid in assessing the spatial extents of subsidence in known subsiding areas, and in locating unknown groundwater stressed regions.

ABSTRACT:

Groundwater overdraft gives rise to multiple adverse impacts including land subsidence and permanent groundwater storage loss. Existing methods have been unable to characterize groundwater storage loss at the global scale with sufficient resolution to be relevant for local studies. Here we explore the interrelation between groundwater stress, aquifer depletion, and land subsidence using remote sensing and model-based datasets with a machine learning approach. The developed model predicts global land subsidence magnitude at high spatial resolution (~2 km) and provides a first-order estimate of aquifer storage loss due to consolidation of ~17 km3/year globally. China, the United States, and Iran account for the majority of groundwater storage loss due to consolidation. The model quantifies key drivers of subsidence and has high predictive accuracy, with an F1-score of 0.83 on the validation set. Roughly 73% of the mapped subsidence occurs over cropland and urban areas, highlighting the need for sustainable groundwater management practices over these areas. The results of this study aid in assessing the spatial extents of subsidence in known subsiding areas, and in locating unknown groundwater stressed regions.