Neural Networks and Geofactors Reveal Hidden Nitrate Threats in Ethiopian Groundwater

In the rolling highlands of Ethiopia’s upper Blue Nile basin, the water beneath the Wanka Watershed is quietly changing. Nitrate concentrations in shallow groundwater are rising as agriculture intensifies, and the health stakes could not be higher: elevated nitrate is linked to methemoglobinemia, the potentially fatal condition known as blue baby syndrome, as well as certain cancers and neural tube defects. Yet in this rural region, as across much of rural Ethiopia, routine water quality monitoring is nearly nonexistent. Laboratory analysis is expensive, logistically difficult, and often simply out of reach for remote communities. A new study published in Discover Sustainability offers a strikingly different approach, using geodetector-based geofactors and neural network methods to predict nitrate levels from a handful of environmental variables rather than costly repeated sampling.

The research, led by Asnakew Mulualem Tegegne of the University of Gondar with colleagues from Bahir Dar University, Care Ethiopia, and international partner institutions, set out to estimate the current status of nitrate concentration in a data-limited area. The team’s central insight was that the factors shaping nitrate in groundwater are not mysterious. They are written into the landscape itself: how much rain recharges the aquifer, how quickly water moves through it, what rocks and soils lie beneath the surface, and how people use the land above. If those relationships can be quantified, they can be exploited to map contamination without a dense monitoring network.

To do this, the researchers turned to geodetector, a statistical technique designed to measure the power of explanatory variables over a spatially patterned outcome. Geodetector works by stratifying the study area according to each candidate factor and comparing the variance of nitrate concentrations within and between strata. The output, known as the factor detector value, expresses each variable’s power of influence as a number between zero and one. The higher the value, the more strongly that geofactor controls the spatial distribution of nitrate. This makes geodetector unusually well suited to heterogeneous watersheds where relationships between land and water are nonlinear and where traditional regression assumptions may not hold.

The team assembled eight geofactors as input variables: groundwater recharge, static water level, transmissivity, lithology, mean air temperature, the Topographic Wetness Index, soil type, and land use. Forty-one groundwater nitrate concentration samples collected in 2013 served as the training dataset, while twelve additional samples taken from the same well locations in 2022 were reserved for validation. The factor detector results were revealing. Land use emerged as the single most powerful influence, with a value of 0.232, followed closely by recharge at 0.210. Transmissivity scored 0.119, soil 0.115, lithology 0.079, the Topographic Wetness Index 0.078, mean air temperature 0.021, and static water level just 0.015.

These numbers tell a coherent story about what drives nitrate contamination in agriculturally intensified highlands. Land use sits at the top because farming practices determine how much nitrogen is applied to fields and how much of it leaches downward. Recharge matters because rainfall percolating through soil is the vehicle that carries dissolved nitrate into the aquifer. Transmissivity, a measure of how easily water flows through the aquifer, controls how contaminants spread once they arrive. Soil type governs both infiltration and the natural attenuation capacity of the subsurface. By contrast, lithology, static water level, and temperature exert only weak influence, suggesting that in this watershed the nitrate problem is fundamentally a surface phenomenon driven by human activity rather than a geological one.

Alongside the geodetector analysis, the researchers trained a neural network to predict nitrate concentrations from the same geofactors. Neural networks are computational models loosely inspired by biological brains, consisting of layers of interconnected nodes that adjust their connection weights as they learn from training examples. Their strength lies in capturing complex, nonlinear relationships between inputs and outputs, exactly the kind of relationships that link landscape properties to water chemistry. The team evaluated both approaches against standard performance metrics, including root mean square error, mean absolute error, residual sum of squares error, and the coefficient of determination.

The results were impressive for both methods, but not identical. The geofactor approach achieved a root mean square error of 0.39, a mean absolute error of 0.26, a residual sum of squares error of 1.83, and a coefficient of determination of 0.995. The neural network produced corresponding values of 0.53, 0.33, 3.37, and 0.99. Both are acceptable by conventional standards, and both captured the spatial pattern of nitrate with high fidelity. But the geofactor method edged out the neural network on every metric in this particular watershed, suggesting that when the underlying explanatory variables are well chosen and the study area is relatively homogeneous, a simpler statistical stratification can outperform a more complex machine learning model.

The practical payoff of the study is a set of predicted nitrate concentration maps for the Wanka Watershed, expressed in milligrams per liter of nitrate-nitrogen. These maps allow managers and policymakers to identify hotspots of contamination and prioritize interventions, whether that means changing fertilizer application practices, protecting recharge zones, or relocating wells. Because the approach relies on freely available landscape data rather than repeated laboratory sampling, it could be replicated in other data-limited watersheds across Ethiopia and beyond, transforming how water quality is monitored in regions where conventional networks are unaffordable.

The study also carries broader implications for global water security. The United Nations Sustainable Development Goal 6, which calls for clean water and sanitation, and Goal 3, which targets good health and well-being, both depend on knowing whether the water people drink is safe. Groundwater is the main source of drinking water in rural and urban areas across the developing world, yet water quality remains poorly monitored in many regions. By demonstrating that geodetector-based geofactors and neural networks can estimate nitrate concentrations from landscape properties with high accuracy, the Wanka Watershed study offers a template for low-cost, scalable water quality assessment. As agricultural intensification accelerates across the global South, tools like these may become essential for protecting both human health and the long-term sustainability of the aquifers that billions depend upon.

Subject of Research: Environmental Science

Article Title: Monitoring groundwater nitrate concentrations using geodetector-based geofactors and neural network approach in the Wanka Watershed, Ethiopia

Article References: Tegegne, A. M., Akale, A. T., Setargie, T. A., Ayalew, A. D., & Lohani, T. K. (2026). Monitoring groundwater nitrate concentrations using geodetector-based geofactors and neural network approach in the Wanka Watershed, Ethiopia. Discover Sustainability. https://doi.org/10.1007/s43621-026-04869-0

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04869-0

Keywords: groundwater, nitrate contamination, machine learning, neural networks, geodetector, Ethiopia, water quality, Blue Nile basin, Monitoring, nitrate, concentrations, geodetector-based