Air Pollution: Kampala

Presented by Mike Smith, University of Sheffield
michaeltsmith.org.uk m.t.smith@sheffield.ac.uk @mikethomassmith








Why does it matter?

In 2012, harm equivalent to 2.6 million deaths, due to outdoor air pollution (mostly heart disease or stroke).

Summary: Air pollution is a big deal (similar scale as malaria?).

Note: We'll focus on PM2.5.

WHO press release 2014: http://www.who.int/mediacentre/news/releases/2014/air-pollution/en/

How bad is it?

Only two papers about Kampala's air quality (Stephan Schwander's group), and a new sensor at the US embassy.



(white board)

US Embassy sensor

  • Manchester average PM2.5 (at 4 sites): 8-16 $\mu g/m^3$ (2011-2015)
  • WHO Guidelines: 10 $\mu g/m^3$ (annual mean)
  • EU Guidelines: 18 $\mu g/m^3$
  • US Guidelines: 12 $\mu g/m^3$

US Embassy sensor: by hour

Other sensors in other locations in city agree with general trend.
Night time peak found in other tropical and subtropical cities.





US Embassy sensor

  • Provides good, reliable, calibrated data.
  • Only tells us about pollution at the Embassy (A).

Methods for estimating pollution

  • Land Use Regression
    • not probabilistic, typically linear regression, long term avgs
    • Ideally requires 40+ sensors... [expensive]
  • Dispersion models (include box and Gaussian plume models). Really depend on having an idea where pollution is from.


    • Hoek, Gerard, et al. "A review of land-use regression models to assess spatial variation of outdoor air pollution." Atmospheric environment 42.33 (2008): 7561-7578.Wu, Jiansheng, et al. "Applying land use regression model to estimate spatial variation of PM2. 5 in Beijing, China." Environmental Science and Pollution Research 22.9 (2015): 7045-7061.

Plan

Can we combine lots of low-cost sensors with a few precision devices to acheive similiar accuracies to a full-scale network?

PART 1: HARDWARE DEVELOPMENT

Project development

Began with trying the Shinyei sensor hooked up to a phone (2014).

Boda boda deployment

Engineer Bainomugisha. Ssematimba Joel hardware dev.
Switched to OPC sensor from AlphaSense.

From left to right: Ssekanjako John, the bodaboda driver; me; Engineer Bainomugisha.

Static site deployment

Static site deployment

Project development

unobtrusive hardware...

Example data

PART 2: COMBINING SENSORS

Combining Sensors

Using a GP model, with inputs:

  • latitude, longitude and time
  • time-of-day
  • distance from large roads (data from OSM)

Multiple output GPs: Make each sensor type one output


website

(visualisation built by Irene Michalaki at ARM)

Combining Sensors

Sensors aren't necessarily calibrated, but are still correlated

Calibration with Coregionalisation

  • Need to consider sensor proximity
  • Assume sensors are proportional (rank-one, $\mathbf{w} \mathbf{w}^\top$)
  • Toy example: Three sensors. One sensor swaps location. Element-wise product (not Kronecker product)
  • Hyperparameters $\mathbf{w}$ only relevant where 'proximity' matrix is non-zero. So need to have sensors co-located to estimate these.
  • Need to use CCD or similar to incorporate uncertainty in $\mathbf{w}$.

PART 3: BODA BODAS

Boda bodas

video!

Boda bodas

  • Boda-bodas provide spatial detail about pollution
  • Might help with calibrating fixed sensors?
  • Appear to be providing reasonably 'realistic' data

Boda boda data

$w$ might change?

  • Need very regular calibration (problems with clogging)
  • Estimates of $\mathbf{w}$ will need to vary. Model with a GP too?

Future ideas

  • Use active learning to select location for bodaboda to visit
  • Look at sources of pollution to try to start assigning causality
  • Apply differential privacy to ensure boda location remains private
  • Correlate with hospital admissions?

New! The Sheffield Urban Observatory is lending three MCERTS TurnKey OSIRIS OPCs for deployment in Kamapala (and funding 15 low-cost sensors). This will allow us to calibrate the low-cost units.

Actions/Effects

  • Awareness raising - leads to civic response?
  • If we get causation - we'll know what to target.
  • Baseline necessary for mitigation to be assessed.
  • KCCA planning non-motorised area in Old Kampala!

Thanks

DIL Berkeley award, ARM, the Urban Observatory and the EPSRC.

Engineer Bainomugisha; Ssematimba Joel

Pius Kavuma, AI Lab (3d printer!); Irene Michalaki, ARM;
Mauricio Alvarez and Neil Lawrence

References

Schwander, Stephan, et al. "Ambient particulate matter air pollution in Mpererwe District, Kampala, Uganda: a pilot study." Journal of environmental and public health 2014 (2014).

Kirenga, Bruce J., et al. "The state of ambient air quality in Two Ugandan cities: a pilot cross-sectional spatial assessment." International journal of environmental research and public health 12.7 (2015): 8075-8091.

Manchester air pollution averages: http://www.manchester.gov.uk/download/downloads/id/24305/greater_manchester_air_quality_annual_status_report_asr_2015.pdf

Diurnal peak time in sub-tropical/tropical locations
Latha, K. Madhavi, and K. V. S. Badarinath. "Seasonal variations of PM10 and PM2. 5 particles loading over tropical urban environment." International journal of environmental health research 15.1 (2005): 63-68.
Zhang, Yan-Lin, and Fang Cao. "Fine particulate matter (PM 2.5) in China at a city level." Scientific Reports 5 (2015): 14884.