AI Maps Hidden Gamma Radiation Hotspots at Nigerian Steel Waste Site

Beneath the rusting heaps of scrap metal at a steel dumpsite in Katsina State, northern Nigeria, an invisible hazard is scattered unevenly across the soil: gamma radiation emitted by naturally occurring and technologically concentrated radionuclides. Mapping that hazard has traditionally required dense networks of field measurements, expensive instruments, and laborious laboratory analysis. Now, a team of Nigerian researchers has shown that a well-trained machine learning model can do much of the work instead, predicting the absorbed gamma dose rate across a heterogeneous industrial site with remarkable accuracy and flagging the locations where human exposure risk is greatest.

The study, published in the journal Environmental Monitoring and Assessment, was carried out by Isaiah Eze Igwe and Usman Ibrahim Machika of the Federal University Dutsin-Ma, together with Chibueze Anslem Mgbechidimma of the University of Ibadan. Their target was the dumpsite of the NAK Steel company in Katsina State, an environment typical of many industrial waste grounds across the region, where slag, scrap, and contaminated soils create a patchwork of radiation levels that conventional single-point surveys can easily miss. The researchers assembled a dataset of 36 soil sampling locations across the site and used it to train and compare four supervised learning algorithms: support vector machine, linear regression, random forest, and decision tree.

The choice of algorithms was deliberate. Linear regression represents the simplest statistical baseline, fitting a straight-line relationship between input variables and the measured dose rate. Support vector machines, first introduced in the 1990s, are more flexible, capable of capturing nonlinear patterns by mapping data into higher-dimensional spaces. Decision trees split the data through a cascade of binary decisions, producing piecewise predictions that can adapt to sharp local changes. Random forests, an ensemble method pioneered by Leo Breiman in 2001, combine hundreds of decision trees, each trained on a random subset of the data, and average their outputs. This averaging tends to suppress the overfitting that plagues individual trees while preserving the ability to model complex, nonlinear spatial relationships.

All four models were trained and tested under identical preprocessing and validation conditions, ensuring a fair comparison. The results were decisive. Random forest outperformed every competitor across all performance metrics, achieving a root mean square error of just 0.0131 and a mean absolute error of 0.0101, the lowest prediction errors of the group. Its coefficient of determination, a measure of how much of the variability in the data the model explains, reached 0.902, meaning the model accounted for roughly ninety percent of the spatial variation in absorbed gamma dose rate. When the task was reframed as classifying locations into hotspot and non-hotspot categories, the random forest again led the field, with a classification accuracy of 0.89, a Cohen’s kappa of 0.77, and an area under the receiver operating characteristic curve of 0.93, all indicating strong and reliable discrimination.

The behavior of the other algorithms was just as instructive, though for different reasons. Linear regression and support vector machines both tended to smooth out spatial variability, producing predictions that underestimated localized anomalies. In a radiation context, that smoothing is not a harmless simplification; it can hide exactly the concentrated pockets of elevated dose that matter most for protecting workers and nearby residents. The decision tree model showed the opposite failure mode, generating abrupt spatial patterns suggestive of overfitting, where the model memorized the quirks of the training samples rather than learning the underlying structure of the radiation field. The contrast highlights a central lesson of modern machine learning applied to environmental data: the best model is rarely the most flexible or the simplest, but the one whose inductive bias matches the phenomenon being mapped.

With the best-performing model in hand, the researchers converted their predicted dose rates into two internationally recognized radiological risk indicators. The first was the annual gonadal dose, an organ-specific measure of radiation exposure to the reproductive organs, which is of particular concern because reproductive tissue is among the most radiosensitive in the human body. Predicted annual gonadal dose values across the site ranged from 0.03 to 0.27 millisieverts per year. The second metric was the excess lifetime cancer risk, a probabilistic estimate of the additional cancer burden that long-term exposure to the site’s radiation field would impose on a hypothetical resident. The ELCR estimates mirrored the spatial variability of the annual effective dose, rising and falling in step with the predicted dose rates across the dumpsite.

The spatial pattern of risk was far from uniform. The highest values of both dose and cancer risk clustered around a single location designated S11, identified as the principal hotspot of the site. This concentration is exactly the kind of feature that sparse or averaged survey campaigns can overlook, and it underscores why spatially resolved prediction matters. At a working dumpsite, where scavengers, workers, and nearby communities may spend hours in specific zones rather than wandering uniformly across the whole area, knowing where the hotspots lie transforms a vague site-wide average into actionable information: which zones to fence, which to remediate, and which to monitor most frequently.

The study is candid about its limitations. A dataset of 36 sampling points is small by machine learning standards, and the authors acknowledge that this constrains the generalization of their models beyond the specific site studied. Small training sets carry a real risk that a model’s apparent accuracy partly reflects the particular configuration of the validation split rather than genuine predictive power. Nevertheless, the framework’s value lies less in the specific numbers it produced for NAK Steel and more in the demonstration that a rigorous, low-cost predictive pipeline, from soil sampling through algorithm comparison to dose and risk conversion, can function in data-sparse regions where expensive radiation monitoring infrastructure is limited. The authors followed internationally accepted conversion models in moving from predicted dose rates to organ doses and lifetime risk, anchoring the machine learning output in established radiological protection practice.

The broader context makes the work timely. Industrial dumpsites across Nigeria and much of the developing world are increasingly recognized as potential hotspots of environmental radiation exposure, both because industrial byproducts such as slag can concentrate naturally occurring radioactive materials and because unregulated waste handling allows contamination to spread. Previous studies at Nigerian dumpsites and mining areas have documented elevated gamma dose rates and raised concerns for waste pickers, whose informal and unprotected work places them in direct and prolonged contact with contaminated soil. Yet predictive assessments of dose distribution in these settings have remained scarce, leaving regulators with little more than point measurements to guide decisions.

What this study offers is a practical template. By showing that random forest can reliably predict absorbed gamma dose rate in a heterogeneous industrial environment, and by chaining that prediction into organ dose and cancer risk estimates, the researchers have sketched a workflow that other teams can replicate at other sites with modest sampling budgets. The approach also aligns with a growing international trend of applying machine learning to environmental radioactivity, from mapping geogenic radon potential to modeling soil-plant transfer of radiocaesium. For data-sparse regions, where the choice has often been between sparse surveys and no assessment at all, predictive modeling offers a middle path: a way to extract maximum insight from limited measurements, direct future sampling to where it matters most, and give communities and regulators a clearer picture of the invisible radiation landscape they live and work within.

Subject of Research: Machine learning prediction of absorbed gamma dose rate and radiological risk assessment at an industrial steel dumpsite in Nigeria

Article Title: Machine learning prediction of absorbed gamma dose rate and radiological hazard assessment at a Nigerian steel dumpsite

Article References: Igwe, I. E., Machika, U. I., & Mgbechidimma, C. A. (2026). Machine learning prediction of absorbed gamma dose rate and radiological hazard assessment at a Nigerian steel dumpsite. Environmental Monitoring and Assessment, 198(10), Article 1046. https://doi.org/10.1007/s10661-026-15865-2

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15865-2

Keywords: machine learning, random forest, gamma dose rate, radiological risk assessment, industrial dumpsite, Nigeria, annual gonadal dose, excess lifetime cancer risk, environmental radiation monitoring, soil radioactivity, hotspot mapping, Katsina State