Just out todayAugmented reality: The AR trainees were faster on the model they trained on, and no better anywhere elseVirtual reality: They matched the exercise dose exactly, and the VR advantage fell below the line that mattersVirtual reality: Both groups used the same headset. One watched a video first and picked up 10.8 more blocks

Climate & energy tech/Senegal

Dakar can forecast tomorrow’s air quality without a single real-time ground sensor

A 24-hour PM2.5 forecast for Dakar, built from open satellite data with no real-time ground sensor, scored R-squared 0.931 on a year it had never seen.

What the paper found

Researchers at Cheikh Anta Diop University built an operational 24-hour PM2.5 forecasting system for the HLM district of Dakar, Senegal, using six years of daily measurements and open-access satellite data. CatBoost achieved R-squared 0.931, RMSE 11.50 and MAE 7.51 micrograms per cubic metre on an independent 2024 test set, and the system requires no real-time ground PM2.5 input.

Grade 12 reading level6 min read

What happened

Air quality forecasting is built on sensor networks. A dense grid of monitors feeds a model that predicts tomorrow from today, and the approach works well in cities that have the grid.

Dakar does not. Neither do most African cities, and the pollution there is worse. The HLM district – low-income housing, urban-industrial – had a mean PM2.5 concentration of 274.71 micrograms per cubic metre across 2019 to 2024. For reference, World Health Organization guidance is in the single digits for annual exposure.

A team in the Department of Physics and the Department of Mathematics and Computer Science at the Faculty of Science and Technology, Cheikh Anta Diop University in Dakar, built a 24-hour forecasting system that does not need the sensor network.

Tested on a year it had never seen, it scored R-squared 0.931.

The test

Daily PM2.5 concentrations from 2019 to 2024 were used to develop and evaluate the system. 2024 was held out entirely as an independent test set – the model never saw it during training.

Four ensemble machine learning models were compared: Random Forest, Extra Trees Regression, XGBoost and CatBoost.

The predictors were all open-access: ERA5-Land meteorology, aerosol optical depth from the Copernicus Atmosphere Monitoring Service, and Sentinel-5P tropospheric columns for nitrogen dioxide, carbon monoxide and sulphur dioxide.

SHAP analysis was then used to identify which predictors carried the signal, which is where the most interesting finding came from.

The numbers

Measure Result
Site HLM district, Dakar, Senegal
Period 2019-2024, daily
Mean PM2.5 274.71 ug/m3
Best model CatBoost
R-squared, 2024 test set 0.931
RMSE 11.50 ug/m3
MAE 7.51 ug/m3
Dominant predictor aerosol optical depth
Then lagged PM2.5, relative humidity, seasonality, precipitation
Real-time ground PM2.5 required none

The result

CatBoost achieved the highest performance on the independent 2024 test set: R-squared 0.931, RMSE 11.50 micrograms per cubic metre, MAE 7.51.

SHAP analysis identified aerosol optical depth as the dominant predictor, followed by lagged PM2.5, relative humidity, seasonality and precipitation. The authors interpret this as reflecting the combined influence of Saharan dust transport, hygroscopic growth, wet deposition and Harmattan-monsoon dynamics.

The consequential finding follows from that: external forcing dominated over local emission persistence, confirming that accurate 24-hour predictions can be issued without any real-time ground-based PM2.5 input.

Relying exclusively on open-access data and requiring minimal computational resources, the framework is presented as scalable to other Sahelian urban contexts with similar monitoring constraints.

How this compares to what came before

Air quality forecasting for African cities has mostly meant transplanting models built where monitoring is dense. Those models assume access to a current measurement to anchor the prediction, and they inherit the pollutant mix of the cities they were built in – typically traffic and industry dominated.

Both assumptions fail in Dakar. There is no reliable current measurement, and the dominant driver is dust blown off the Sahara rather than anything emitted locally.

Building the model where it will be used, with the data actually available there, is what produced both the workable architecture and the physical insight. A model developed in Europe and applied to Dakar would have been structured around local emissions and would have needed a sensor reading it could not get.

What it means

The deployment implication is direct. A city with one working monitor – or none – can forecast tomorrow’s air quality from freely available satellite data. That is a different proposition from “build a sensor network first”, which has been the standing recommendation and has not happened.

The physical finding reshapes what intervention means. If external forcing dominates local emission persistence, then local traffic and industrial controls will not solve Dakar’s PM2.5 problem. Warning, shelter and health system preparedness are the levers that remain, and a 24-hour forecast is exactly what those need.

The flip side is a real limitation. A model that forecasts without real-time ground input also cannot correct itself when a local source fires – a fire, an industrial release, an unusual traffic event. It predicts the dominant regime well and will miss the exception.

That is an acceptable trade when the dominant regime is 274 micrograms per cubic metre of Saharan dust. It would not be in a city where local sources matter more.

Business ideas from this paper

Sensor-free forecasting for underserved cities. The architecture is the product: open satellite inputs, modest compute, no ground network. Replicating it across Sahelian cities is a deployable service rather than a research programme.

Health system warning integration. A 24-hour forecast is useful only if it reaches clinics, schools and vulnerable households. Building that delivery layer is the part that turns a model into a public health intervention.

Minimal calibration networks. The model needs historical ground data to train, even if it does not need real-time input. A short-duration monitoring deployment to bootstrap a city’s model is a defined, affordable project.

Dust-aware infrastructure planning. If Saharan dust dominates, then air filtration, building design and solar panel cleaning schedules across the Sahel all have a forecastable input they do not currently use.

What could be wrong with it

One monitoring district in one city. The HLM district is specifically a high-exposure urban-industrial environment and is not representative of Dakar as a whole.

The R-squared of 0.931 comes from a single held-out year, 2024. Testing on an unseen year is the right design and one year is a narrow base; a different year with different dust dynamics could give a different figure.

A mean of 274.71 micrograms per cubic metre is extremely high, and a model trained across that range may transfer poorly to cleaner cities where the variation has a different structure.

The system cannot respond to local events, by design.

The scalability claim to other Sahelian contexts is a projection, not a demonstration – no second city was tested.

Data are not publicly deposited.

How sure can you be?

We score this 6 out of 10.

It is a peer-reviewed paper in Environmental Monitoring and Assessment, using six years of daily ground measurements with a properly held-out independent test year, from a named institution – Cheikh Anta Diop University in Dakar.

It loses points because all authors share one institution, there is no independent replication, the data are not openly deposited, and the validation rests on one site and one test year.

The design choices are good – independent test year, open inputs, SHAP analysis to explain the model. The base is narrow, and we would want a second city before treating the Sahelian scalability claim as established.

Do this today

If you work on air quality in a city with sparse monitoring, this architecture is replicable with open data and modest compute. The barrier is historical ground data for training, not infrastructure.

If you are commissioning a monitoring network, note that the forecast here needed historical data to train and no real-time input to run. That changes the cost profile of what you need to build.

If you live or work in the Sahel, the finding that external dust dominates local emissions is the planning fact. Local emission controls are worth doing for other reasons and will not move this number.

Source: Operational 24-h PM2.5 forecasting using ensemble methods and open-access satellite data: application to Dakar, Senegal, Environmental Monitoring and Assessment, May 2026. PMID 42120759 · doi.org ().

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.

What to remember
  • Built and tested entirely in Senegal, at Cheikh Anta Diop University in Dakar.
  • Mean PM2.5 across 2019-2024 was 274.71 micrograms per cubic metre - extremely high by any standard.
  • Aerosol optical depth was the dominant predictor, reflecting Saharan dust transport rather than local sources.
  • The system needs no real-time ground PM2.5 input, which is what makes it deployable where sensors do not exist.
Where this really is

Six years of daily PM2.5 measurements from a high-exposure district were used to build and test the system, with 2024 held out entirely as an independent test set.

Questions people ask

Why does this matter for African cities?

Chronic PM2.5 pollution threatens public health in African megacities while ground monitoring infrastructure remains sparse. A forecast that works without a dense sensor network is deployable where the need is greatest.

What does needing no real-time ground input mean?

The model forecasts from satellite and reanalysis data alone. It does not need today's sensor reading to predict tomorrow, which means it works in cities that have no working sensor today.

What is aerosol optical depth?

A satellite measurement of how much light atmospheric particles block. It was the strongest predictor here, which points to Saharan dust transport as the dominant driver rather than local emissions.

Is R-squared 0.931 reliable?

It is on one held-out year, 2024. Testing on a year the model never saw is the right design; a single test year means the figure rests on one season's conditions.

About the author

Mark Alex

Mark Alex is the founder and Managing Director of Real Biz Digital, a technology company operating out of Nairobi since 2018. He works in agentic AI and the Model Context Protocol, AI governance, enterprise software architecture and cybersecurity. He holds an MSc in Mechatronical Engineering from Obuda University in Budapest and a BSc in IT, Forensic Technology and Cybercrime, from USIU-Africa in Nairobi, and has published IEEE conference research on an AI-powered digital twin for greenhouse systems. He is the author of seven books. Between 2020 and 2024 he mentored more than 200 university students and interns in Nairobi. He writes every Just Out Tech article from the original research paper.