Amref Health Africa, ProRail and Analytics for a Better World are developing a decision-support tool to identify more promising locations for new boreholes in Kenya’s arid and semi-arid lands. Current geological assessments lead to successful drilling in around 60% of cases. The project aims to use satellite, geological and other spatial data to increase that success rate to above 90%, while reducing the cost of unsuccessful drilling.
Why is finding water underground so difficult?
In Kenya’s arid and semi-arid lands (ASALs), reliable access to water can depend on boreholes: deep, narrow wells drilled into the ground to reach underground water.
Choosing where to drill is difficult. A borehole can require substantial investment, but even after geological surveys and specialist assessments, drilling does not always reach a reliable water source. Amref estimates that the current approach succeeds in around 60% of cases.
An unsuccessful borehole means that money, equipment and staff time have been used without creating a functioning water source. For communities where reliable water infrastructure is already limited, it can also delay access to safe water.
Part of the difficulty is the data. Information that may indicate where groundwater is more likely to be found — including previous boreholes, terrain, climate, geology and hydrology — can be incomplete or spread across different sources. The project therefore began by bringing these datasets together and assessing whether they were reliable enough to inform drilling decisions.
Why is this relevant to predictive and prescriptive analytics?
The first analytical question is predictive: based on what is known about a location, how likely is drilling there to succeed?
The project uses geospatial machine learning, which means training a computer model to find patterns between previous drilling outcomes and information linked to particular locations, such as terrain, climate and water-related characteristics.
But the practical question goes further: if Amref is considering several possible locations, which ones should be investigated first?
Rather than producing a definitive instruction to drill at a particular point, the model is being developed as a screening and decision-support tool. It can rank locations according to their estimated potential and uncertainty, giving specialists another source of evidence when deciding where to carry out more detailed assessments.
This keeps expert judgement central. The model narrows the search as it does not replace geological expertise or field assessment.
How did Amref, ProRail and ABW develop the approach together?
The project was developed during the first edition of the Analytics for a Better World Impact Accelerator Program, bringing together Amref Health Africa’s knowledge of water programmes, field conditions, and engineering with ProRail’s data and analytics expertise.
The collaboration began by understanding how borehole locations are currently selected and what information is available. A substantial part of the first phase involved finding, combining and assessing data from previous boreholes together with spatial datasets on factors such as climate, terrain and hydrology.
The team gradually refined the problem into a machine-learning classification task. Rather than trying to predict with certainty whether water would be found, the model would identify locations that appear more or less promising and prioritise them for further investigation.
Frequent meetings and workshops allowed Amref’s domain knowledge and the technical analysis to develop together. The team identified early alignment between technical and non-technical perspectives as particularly important, because assumptions about the data could be checked against how borehole decisions are actually made in practice.
What does the tool do?
The tool combines historical borehole information with publicly available spatial data, including information related to climate, terrain and hydrology.
The machine-learning model looks for patterns associated with previous drilling outcomes and uses them to score potential locations.
A higher score does not mean that a borehole is guaranteed to succeed. Instead, it allows practitioners to distinguish locations where the available evidence suggests stronger potential from locations where uncertainty is higher.
The project was designed in three stages. First, historical data was used to develop and validate the model. Second, the model was translated into an interface that practitioners can test and refine. The final stage is intended to take the prototype into field use and compare its recommendations with actual drilling outcomes.
A new site for a borehole from Amref Health Africa's project gallery
Why did data quality become part of the project?
The project showed that developing a machine-learning model was only part of the challenge.
Historical borehole information was fragmented, and considerable work was required to understand which data could be combined and whether it was sufficiently consistent for modelling. Data availability and quality became some of the largest constraints during development.
At the same time, the team found that useful information did not always require expensive proprietary datasets. Combining open-source datasets created opportunities to build a more consistent view of possible drilling locations while keeping future operating costs lower.
The project therefore also pointed towards a longer-term organisational change, namely bringing previously fragmented borehole and spatial information into a more structured workflow that can be updated and reused for future decisions.
Why does human judgement remain important?
Groundwater conditions are complex, and available data cannot capture everything that determines whether drilling will succeed.
For that reason, the team deliberately positioned the model as a decision-support tool rather than a replacement for specialists. Its role is to prioritise locations for further investigation, while geological expertise and knowledge of local conditions remain part of the final decision.
This also affected how the project was developed. Regular interaction with domain specialists allowed assumptions in the model to be compared with field knowledge and helped the team focus on information that practitioners could realistically use.
The intended change is therefore not to automate borehole planning completely. It is to give specialists a stronger evidence base before committing resources to an expensive drilling decision.
What could increasing drilling success mean in practice?
The project’s ambition is to increase successful drilling from around 60% to more than 90%.
If that target is achieved in field testing, fewer drilling attempts would end without a viable water source. That could reduce the amount of programme funding spent on unsuccessful boreholes and allow the same resources to support more water infrastructure or other health and development activities.
For communities in arid and semi-arid areas, the longer-term relevance is more direct. Each successful borehole can contribute to more reliable access to safe water.
However, these downstream effects have yet to be measured. The next stage will focus on establishing whether locations prioritised by the model actually produce higher drilling success rates when used in real field decisions.
How was the solution designed to continue beyond the accelerator?
Sustainability was considered through both the technology and knowledge transfer.
The model was built using open-source tools and publicly available datasets, reducing dependence on expensive software or data licences. The team also produced documentation and planned knowledge-transfer sessions so that Amref can understand, maintain and adapt the approach rather than depending indefinitely on the original development team.
The underlying machine-learning pipeline was also designed to be retrained with new data. That meant the approach could potentially be adapted when more borehole results become available or when it is applied in another region.
What can this mean for future water planning?
The immediate next step is field validation.
Amref can now compare locations prioritised by the model with actual drilling outcomes and establish whether the tool improves success rates relative to the current approach. That testing will also show where predictions are less reliable and which additional data could improve them.
If the approach performs well, the same framework could be retrained for other arid and semi-arid regions rather than assuming that patterns identified in one area apply everywhere.
The longer-term opportunity is therefore larger than predicting individual boreholes. By combining historical drilling experience with spatial data in a reusable system, Amref could build a stronger evidence base for deciding where limited water-infrastructure resources are most likely to result in a functioning and sustainable water source.
