Florida’s farms, communities, and natural ecosystems share a vital resource: groundwater. Aquifers support our drinking-water supplies, agriculture, springs, wetlands and rivers. But in many parts of Florida, shallow and highly permeable aquifers also leave groundwater vulnerable to contaminants moving down from the land.

A key concern is nitrate from fertilizers, septic systems and other human activities. For agriculture, the challenge is finding the balance between producing food and protecting the water resources that farms and communities depend on. That requires knowing where groundwater is most vulnerable and where protection efforts can have the greatest impact.

New research led by Golmar Golmohammadi, assistant professor of watershed hydrology and biogeochemistry at SWES, explores how artificial intelligence (AI) can help answer that question. The study combines established groundwater science with modern deep learning to create high-resolution maps of nitrate vulnerability across Florida’s Surficial Aquifer System. It uses explainable AI to show what factors are driving those predictions.

Why groundwater vulnerability matters to Florida agriculture

Nitrate is a particular concern because nitrogen is essential for agriculture, but what plants don’t use can move through soil and eventually reach groundwater. Once it’s there, nitrate can be difficult and expensive to remove. And because groundwater connects with springs, streams, wetlands and other surface waters, its quality can affect entire watersheds.

For Florida agriculture, the challenge is not simply reducing fertilizer use or restricting production. It is understanding where the landscape is most likely to allow nutrients to move into groundwater.

“The goal is to give farmers, water managers and policymakers better information about where the groundwater is most vulnerable to contamination,” Golmohammadi said, “so protection and monitoring efforts can be more targeted and effective.”

Bringing AI into groundwater science

The study builds on DRASTIC, a groundwater-vulnerability assessment method used for decades. DRASTIC considers seven characteristics of the landscape and aquifer, including depth to groundwater, recharge, soil, topography, and the aquifer’s ability to transmit water.

DRASTIC is useful because it is straightforward and grounded in hydrogeologic principles. But real landscapes are complex. The importance of individual factors can vary across a region, and their interactions are difficult to capture with fixed weights.

The researchers used deep learning, a form of AI, to identify these complex spatial relationships. They tested two approaches—a convolutional neural network (CNN) and a U-Net—using the DRASTIC data along with nitrate observations. The models generated vulnerability predictions at 10-meter resolution, allowing them to capture finer-scale differences across Florida’s Surficial Aquifer System.

Both models performed well, but U-Net produced smoother, more spatially refined vulnerability patterns.

“AI is most useful to us when it helps reveal patterns that are difficult to see with conventional approaches,” Golmohammadi said. “The important point is not simply that the computer produces a map—it is that we can use explainable AI to investigate what is driving those predictions.”

Predicting vulnerability is only part of the challenge. For an AI tool to be useful in environmental decision-making, scientists and managers also need to understand why it produces a particular result.

The researchers addressed this using SHAP, an explainable-AI method that helps quantify how individual factors influence a model’s predictions. The analysis identified depth to water and topography as the most influential factors. Hydraulic conductivity and net recharge also played important roles.

Depth to water is understandable: when groundwater is close to the land surface, contaminants have less distance to travel before reaching the aquifer. Topography is more complicated in Florida. Even modest differences in elevation can correspond to differences in soils, drainage, infiltration and land use. Slightly elevated sandy areas can allow water to infiltrate rapidly, while some low-lying and poorly drained areas can retain water and provide conditions that reduce nitrate movement.

Golmohammadi notes that the findings illustrate an important advantage of AI.

“The models can identify relationships that may be difficult to capture through a single set of predetermined rules,” she said.

From maps to management

The research shows how detailed vulnerability assessments could support decisions on the ground.

High-resolution vulnerability maps could help water managers identify areas where groundwater monitoring or preventive measures deserve greater attention, including:

• septic-system upgrades;
• fertilizer-management strategies;
• groundwater monitoring and additional water-quality assessment; and
• protection of springs, wetlands and surface waters connected to vulnerable groundwater.

The maps do not predict that every location identified as “vulnerable” will have elevated nitrate. Instead, they identify areas where environmental conditions make contamination more likely.

“A vulnerability map is a screening and decision-support tool, showing us where to look more closely,” Golmohammadi said. “It does not and should not replace measurements in wells and groundwater monitoring.”

For agriculture and environmental management, that distinction is important.

The research also highlights how closely connected Florida’s water challenges are. For farmers, groundwater quality is tied to nutrient management and the long-term productivity of the water resources that support agriculture. In communities, it is connected to drinking-water supplies and septic systems. Environmental managers link groundwater quality to springs, wetlands, rivers, and aquatic ecosystems. Those connections are becoming increasingly important as Florida experiences population growth, land-use change, and climate-related pressures.

What comes next

The researchers emphasize that this work is an important step, not the final answer.

Future research could incorporate changing climate and hydrologic conditions, connect the models with real-time groundwater monitoring, and expand the approach to additional contaminants such as arsenic. Independent validation with new datasets will also be important as the models are developed for broader decision-support applications.

Golmohammadi’s team is also working to translate the statewide vulnerability maps into an interactive, web-based tool. That will let users explore groundwater vulnerability across Florida and see contributing factors in a given area.

“The next step is to move from a primarily static vulnerability map toward a more dynamic understanding of groundwater risk,” she said. “Doing that can help us account for changing climate conditions, new observations, and multiple contaminants.”

The broader goal is not to let AI make groundwater decisions on its own, but to give scientists, farmers, water managers, and policymakers better information for making those decisions.

For the full research article, see “Interpretable deep-learning DRASTIC for statewide nitrate-vulnerability mapping in Florida’s surficial aquifer system” in Progress in Disaster Science on ScienceDirect: https://doi.org/10.1016/j.pdisas.2026.100677

Featured image above of a South Florida landscape by Tom Wright, UF/IFAS Photography.

2