An Overview for Philanthropists of an AI System to help Small Farmers: WellMapr™ / MODL
12/06/2026
12/06/2026
Vision & Method
The vision of this initiative is to deliver more food, clean drinking water, and sustainable irrigation to smallholder farmers in Ethiopia. The core method relies on an artificial intelligence system called “WellMapr™” to detect and map shallow groundwater. Currently, aid agencies lose extensive funding to dry wells—wasting resources on an estimated 70% failure rate in the lowlands and 50% in the highlands. WellMapr™ drives down these drilling costs and financial risks.
A parallel goal is to embed WellMapr™ within a Sustainable Knowledge Centre (SKC). The SKC will actively listen to small farmers and deliver digestible packets of information to support their overall agricultural development and coordinate with 500 water agencies across Ethiopia.
Context & Summary
In Ethiopia, over 47 million smallholder farmers live below the World Bank poverty line, experiencing long-term undernourishment, fear, and social unrest due to climate-change-induced droughts. Shifting rainfall patterns have drastically reduced crop yields, particularly for cereals. While Ethiopia possesses 6 million hectares of land suitable for shallow irrigation, less than 5% is currently utilized. Traditional oil-industry-derived geophysics could identify groundwater, but its prohibitive cost prevents widespread use.
WellMapr™ solves this problem by using low-cost AI to improve drilling accuracy for shallow wells (30 meters) on small farm plots (0.5 to 1.0 hectare). Access to reliable irrigation can increase crop yields threefold. In areas where shallow groundwater is unavailable, WellMapr™ maps can redirect finite philanthropic resources toward alternative, highly beneficial soil conservation practices. These methods include deep bed farming, zero tillage, and sand dams, which can still double crop yields.
Since 2020, this project has received over $1 million in support. It has pushed its median average error from +/- 18 meters in Q4 of 2025 to +/- 9 meters in Q2 of 2026 representing a doubling of accuracy in six months, due to focused technology sprints made possible with funding from FCDO, UK Aid, the Frontier Technology Hub, and the Czech Geological Survey. The underlying technology is completely dependent on local well data availability, but holds global potential for further development, offering a model to benefit 2.5 billion small farmers worldwide facing climate-driven weather disruptions.
Our Partners
The initial development team began working together in December 2019. The project brings together local expertise, international academic institutions, and leading technology companies:
Led by MapAid, these organizations are also developing WellLite™, a citizen-science water well data collection app designed to feed vital field data back into the main WellMapr™ system.
Costs & Plan
Our immediate target is to raise £64,700 for Phase 1, which focuses on field-based testing to determine the best way to code the WellMapr™ application.
Phase 2 involves deploying the application across a 23,000-square-kilometer area spanning three areas in southern Ethiopia: Sidama Region, Gamo, and Gofa Zones. Phase 3 will scale the technology across the entire country. Phase 4 focuses on data archiving and retrieval. Supported by Databricks, Phase 4 also includes building an Online Water Library (OWL) to scan and capture water data from a wide variety of historical water documents.
Introduction to the Technology
The concept for WellMapr™ emerged from discussions with senior members of the Ethiopian Ministry of Agriculture, who emphasized the urgent need for groundwater maps to aid vulnerable small farmers. Agriculture employs 80% of the Ethiopian workforce. Reliable water access from wells transforms these communities, saving rural women from spending hours walking to fetch impure water from rivers, while improving regional health, food security, and education. To prove the concept, the team focused initial development on the 5,000-square-kilometer Bilate sub-catchment within the Rift Valley, a region severely impacted by Ethiopia’s worst drought in 30 years.
Bilate Watershed Groundwater Map, color-coded with MODL predictions and black training well dots, superimposed on a gold map of Ethiopia.
The core technology powering the WellMapr™ app is a machine learning algorithm called Modeling Optimal Drilling Locations, or MODL. MODL analyzes data from a sample of existing wells to recognize patterns and predict water table depths in areas where no wells currently exist. The system will eventually generate these highly predictive maps for the entire Gamo, Gofa, and Sidama regions.
Water Table Depth Prediction
To train an AI to find water, it must learn from features that are both readily available and closely related to underground hydrology. MODL uses 20 public environmental variables, including seasonal precipitation, temperature, vegetation type, humidity, elevation, and soil type. This information is sourced from open databases provided by NASA, NOAA, and the USGS.
To build the initial model, the team gathered a dataset of wells within the Bilate watershed. Each well log contained GPS coordinates, soil types converted into numerical values, and the actual depth to the water table.
A major trap in machine learning is “overfitting,” where an algorithm memorizes the training data perfectly but fails when applied to new, unknown locations. To prevent this, the team split the well data into two groups: a training set (80% of the wells) to teach the algorithm, and a testing set (20% of the wells) kept completely hidden from the model. By testing the AI on data it had never seen before, the team achieved an accurate assessment of how WellMapr™ will perform when guiding real-world drilling teams in uncharted fields.
Technical Details
Data Collection, Model Construction, and Map Visualization Flowchart
The development process moves through three clear phases: Data Collection and Preparation, Model Construction and Evaluation, and Map Visualization.
During the evaluation phase, the team tested five popular machine learning algorithms: multiple linear regression, multivariate adaptive regression splines, artificial neural networks, random forest regression, and gradient boosting regression. Each model was fed the 20 environmental variables and evaluated using standard statistical metrics, such as R-squared and mean absolute error. To guarantee the results were not a random fluke, the training and testing split was shuffled and repeated fifteen times, averaging the final performance scores.
Gradient boosting regression emerged as the clear winner. This method works iteratively by
layering multiple simple models on top of one another, with each new layer specifically correcting the errors made by the previous one until predictions can no longer be improved.
Observed vs. Predicted Depth Scatterplot with Blue and Red Dots
When plotted visually, the predicted depths closely mirror the actual observed depths. The median absolute error on completely unseen test wells averaged roughly 19 meters. This means that half the time, the AI’s prediction is within 19 meters of the actual water level—an accuracy rate high enough to significantly reduce the risk of drilling dry wells.
The final stage takes this trained model and applies it to a high-resolution grid of the entire target region, dividing the landscape into precise 100-meter by 100-meter blocks. The AI predicts the water depth for each grid point, outputting a clear, color-coded map for field teams.
Future Outlook
WellMapr™ is highly adaptable and designed to scale. Because it starts with historical records, the model can be continually updated. Every new well dug in the field—whether it succeeds or fails—provides data that makes the AI smarter over time. For regions where data is scarce, advanced cross-validation techniques like Leave-One-Out Cross Validation (LOOCV) are used to extract maximum predictive power from very small datasets.
Ultimately, this project uses cutting-edge machine learning to solve a fundamental human crisis. By blending advanced data science with the deep expertise of local communities and international scientists, WellMapr™ removes the expensive guesswork from water exploration. Every successful well stabilizes a community, shields a family from climate shock, and builds long-term economic resilience.
Authorship
This overview was written under the leadership of Kathryn Laskey from George Mason University, with inputs from Global MapAid and Arba Minch University, reflecting the cross-border collaboration required to bring this technology to Ethiopia’s smallholder farmers.
