climate-weather/ai in a supporting role/Toxics 2022 · v2
Statistical model estimates what Makkah's air would have held without lockdown
Researchers compared air pollution in Makkah, Saudi Arabia, before and during the 2020 COVID-19 lockdown. A supervised model trained on 2019 readings predicted what pollutant levels the weather alone would have produced, giving a baseline to subtract from what was measured.
spectrum · one line per step, placed by what the step does · bright lines used AI
Modelling the Effect of COVID-19 Lockdown on Air Pollution in Makkah Saudi Arabia with a Supervised Machine Learning Approach
Toxics, 2022
doi:10.3390/toxics10050225 · record aix-00291 v2 · checked 2026-11-05
- AI was for
- Forecasting
- Model family
- Linear model
- Checked by
- Held-out
- Code
- not reported
AI processed or interpreted data, but the main finding does not rest on it.
What this research was about
Air pollution levels swing about from day to day even when the sources of that pollution do not change. Wind disperses dirty air or traps it, humidity and sunshine alter the chemistry, and traffic follows its own daily and weekly rhythm. That makes a simple before-and-after comparison awkward. When the COVID-19 lockdowns cut traffic and other activity in 2020, researchers wanted to know how much of the resulting change in air quality came from reduced emissions rather than from a spell of different weather.
This study looked at Makkah in Saudi Arabia, using hourly readings from five air quality monitoring stations for 2019 and 2020. Three pollutants were tracked: nitrogen dioxide (NO2), which comes largely from vehicle exhaust, PM10 (airborne particles small enough to be inhaled), and ozone (O3), which forms in the air through reactions driven by sunlight. The researchers took two routes to the same question, one a plain comparison of measured 2019 and 2020 figures for March to August, the other adjusted for weather.
Where AI came in
The weather-adjusted route used a generalised additive model, a form of statistical model that learns a flexible curve for each input rather than a straight line. It was fitted in the R package 'mgcv', separately for each pollutant, and described by the authors as supervised machine learning, meaning it learned from examples where both the conditions and the resulting pollutant level were known. The inputs were relative humidity, wind speed, wind direction, temperature, hour of day, day of week and day of year.
The 2019 data were divided, with three quarters used to fit the models and the remaining quarter held back to check their predictions. The models were then refitted on the whole of 2019 and applied to the weather and calendar conditions of the March to August 2020 lockdown period, producing a 'business as usual' estimate: what the pollutant levels would have been had activity carried on as before. Subtracting that estimate from the measured levels left the weather-corrected change. The model stood in for the counterfactual year that could not be observed.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
Hourly air quality and meteorology data from five monitoring stations in Makkah, Saudi Arabia were used to assess how the COVID-19 lockdown affected O3, NO2 and PM10. Generalised additive models were trained on 2019 data and used to predict business-as-usual concentrations for the March-August 2020 lockdown period, which were then subtracted from observed concentrations. Comparing measured 2019 and 2020 concentrations gave site-averaged changes of -12% for PM10, -58.66% for NO2 and +68.67% for O3; the weather-corrected comparison gave +21.96% for O3, -13.40% for NO2 and -9.66% for PM10. The authors attribute the rise in O3 to its inverse relationship with NO2.
How AI was used
A supervised generalised additive model was fitted in R with the 'mgcv' package, separately for O3, NO2 and PM10, using smooth functions of relative humidity, wind speed, wind direction, temperature, hour of day, day of week and day of year as predictors. The 2019 hourly data were split into a 75% training and a 25% testing portion; models were fitted on the training portion and scored on the held-out portion with RMSE and R-squared. The models were then refitted on the complete 2019 record and applied to the March-August 2020 lockdown period to predict concentrations under a business-as-usual scenario. These predictions were subtracted from the observed lockdown concentrations at each of the five sites to give weather-corrected changes, which were set alongside a separate non-model comparison of measured 2019 and 2020 concentrations for the same months. The 'openair' package was used for correlation coefficients, RMSE values and plots, and 'lubridate' to construct the temporal variables.
The shape of the work
Structural · the record, drawn
Collect air quality and meteorology data
Obtaining raw data, whether by measurement, download or retrieval.
Hourly measured data of several air pollutants and meteorological variables for years 2019 and 2020 were collected at five air quality monitoring stationswhere the paper describes this · verbatim
Compare measured 2019 and 2020 lockdown-period concentrations
Extracting understanding from model behaviour.
In the first approach, the mean concentration was calculated for the lockdown period in 2019 and 2020.where the paper describes this · verbatim
Add temporal variables and split 2019 data
Cleaning, filtering, normalising or labelling data already obtained.
Firstly, the model was validated by splitting the 2019 hourly data into 25% randomly selected testing data and 75% randomly selected training data.where the paper describes this · verbatim
Fit GAMs on training split
Fitting model parameters, including fine-tuning an existing model.
The model was fitted using the training data and validated using the independent testing datawhere the paper describes this · verbatim
Evaluate models on held-out split
Testing outputs against ground truth.
root mean squared error (RMSE) and coefficient of determination (R-squared) were calculated by comparing measured and predicted concentrationswhere the paper describes this · verbatim
Refit models on full 2019 data
Fitting model parameters, including fine-tuning an existing model.
The model was refitted using the entire year 2019 data for training the modelwhere the paper describes this · verbatim
Predict business-as-usual concentrations for lockdown period
Running a trained model over new data to predict, classify or score.
The trained models were reused to predict the pollutant concentrations for the BAU scenario.where the paper describes this · verbatim
Derive weather-corrected change from observed minus predicted
Extracting understanding from model behaviour.
The concentrations predicted by the models were subtracted from the observed concentrations during the COVID-19 lockdown period.where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The model supplied weather-corrected (business-as-usual) baseline concentrations for one of two parallel approaches; the reported direction of change in NO2, PM10 and O3 is also obtained from the non-model comparison of measured 2019 and 2020 data.
GAM is an interpretable supervised machine learning approachwhere the paper describes this · verbatim
the 2019 data were split into randomly selected training (75%) and testing data (25%)where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of Generalised additive model (GAM, R 'mgcv')Which version of the model was used is not stated.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
- What step 7 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00291, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Statistical model estimates what Makkah's air would have held without lockdown. Record aix-00291, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00291/