Machine learning for groundwater management across U.S. agricultural regions

Challenge

Challenge

Groundwater depletion across 450,000 km² of U.S. farmland threatens food production. Traditional numerical models could not keep pace with changing climate and irrigation demand at that scale.

Approach

Approach

We built a hybrid artificial neural network integrating climate data, streamflow, and irrigation demand, optimized with genetic algorithms and screened with mutual information theory. Monte Carlo analysis quantified uncertainty. Validation ran against independent USGS observations and traditional numerical models.

Deliverables

What the client received

• Predictive model covering 900,000+ USGS monitoring wells
• Critical stress zone maps for water managers
• Peer-reviewed methodology (Water Resources Research, 2017)

Outcome

Outcome

Model accuracy exceeded R² 0.85 across 900,000+ monitoring wells over a 33-year calibration period (1980-2013). Published as Sahoo et al. (2017), Water Resources Research, doi: 10.1002/2016WR019933, now over 150 citations.