OpenSense II
Crowdsourcing High-Resolution Air Quality Sensing Presenter: Adrian Arfire, EPFL PI: Alcherio Martinoli, EPFL Co-PIs: Karl Aberer, Boi Faltings, EPFL Andreas Krause, Lothar Thiele, ETH Z端rich Lukas Emmenegger, EMPA Murielle Bochud, University Hospital Lausanne Michael Riediker, Institute for Work and Health, Lausanne opensense.epfl.ch OpenSense II
IMPORTANCE
OF
A IR Q UALITY
On March 25, 2014, the WHO reported: “… in 2012 around 7 million people died – one in eight of total deaths – as a result air pollution exposure. This finding more than doubles previous estimates and confirms that air pollution is now the world’s largest single environmental health risk. ” “The new estimates are not only based on more knowledge about the diseases caused by air pollution, but also upon better assessment of human exposure to air pollutants through the use of improved measurements and technology.”
http://www.who.int/mediacentre/news/releases/2014/air-pollution/en/
OpenSense II
URBAN A IR POLLUTION
Air pollution in urban areas is a global concern • •
affects quality of life and health urban population is increasing
Air pollution is highly location-dependent • •
•
traffic chokepoints urban canyons industrial installations
Air pollution is time-dependent •
• •
rush hours weather industrial activities OpenSense II
OBJECTIVES
IN
A IR POLLUTION MONITORING
Accurate location-dependent and real-time information on air pollution is needed Officials • •
•
public health studies environmental engineers: location of pollution sources municipalities: creating incentives to reduce environmental footprint
Citizens •
• •
advice for outside activities assessment of long-term exposure pollution maps
OpenSense II
OpenSense ultra fine particle levels map in Zürich during winter months
M ONITORING TODAY
Expensive mobile high fidelity equipment
Sparse sensor network (Nabel)
Stationary and expensive stations
Personal exposure with specialized punctual studies
Garage
Indoor
4E5
3
2E5
OpenSense II
17 :0 0
16 :0 0
15 :0 0
14 :0 0
13 :0 0
12 :0 0
11 :0 0
10 :0 0
09 :0 0
1E5
08 :0 0
Coarse models (mesoscale = 1km2)
particles/cm
3E5
Vehicle Road
S YSTEM VISION Measurement data
Model input
Crowd-Sensors, mobile sensors, monitoring stations
Terrain, Meteorology, Source strength, Background
Modeling Lagrangian dispersion models, Data-driven methods
High resolution urban atmospheric pollution maps OpenSense II
OVERVIEW
Health Impact Studies Air Pollution Mapping
Crowdsensing
Data Quality Mobile Sensor Networks
OpenSense II
Health Impact Studies Air Pollution Mapping
Crowdsensing
Data Quality Mobile Sensor Networks
OpenSense II
OPENS ENSE D EPLOYMENTS 10 streetcars in Zurich & 10 buses in Lausanne • CO, NO2, O3, CO2, UFP, temperature, humidity • Localization: GNSS for trams, GNSS fusioned with odometry and stop information for buses • Communication: GPRS
At NABEL stations in Dübendorf & Lausanne • Stations run by EMPA • Calibration and sensor drift evaluation • Testing new sensors
On top of C-Zero electric vehicle • 100% electric, flexible mobility • system test bed, targeted investigation tool On top of “LuftiBus” • Since March 2013, covers whole Switzerland
OpenSense II
GSN
FOR
OPENS ENSE II
BACKEND • Unified data acquisition process • Web data access/filter/download • Sensor data archiving • Sensor data search
GPRS • • • • • • • •
packet parser system logging database server GPS interpolation advanced filtering fault detection system health monitor automatic reporting
• Time series processing
OpenSense II
Health Impact Studies Air Pollution Mapping
Crowdsensing
Data Quality Mobile Sensor Networks
OpenSense II
C ROWDSENSING: I NITIAL Zurich prototype
EFFORTS AT THE END OF
Lausanne prototype
Smartphone connected to ozone sensor and various application software for Android [Hasenfratz et al., Mobile Sensing 2012]
[Predic et al., PerCom’13]
OpenSense II
OPENS ENSE
AirQualityEgg
Low-cost devices for home deployment (calibration tests at the NABEL station in Dübendorf)
C ROWDSENSING: NODE
DEVICES AS
POTENTIAL
Variable Inc. NODE device -
Under 500 CHF(base module + two gas sensors) Works via Bluetooth communication with a smartphone Lightweight; can be mounted on bike/car or carried in pocket
Data quality testing of NODE @EMPA -
Collected CO and CO2 recordings over two days Reference measurements provided by EMPA Promising quality for CO2 while issues with CO sensors
OpenSense II
PLATFORM
PRIVACY
AND
INCENTIVES
IN SENSOR SELECTION Singla et al., poster
Incentives for data gathering [Singla & Krause, HCOMP’13] How to valuate and negotiate access to private information of strategic agents? Main contribution: Privacy-aware, adaptive, truthful mechanism with monetary incentives to compensate for information shared
-
Case study of air quality monitoring [Singla & Krause, HCOMP’13] Simulation studies, with data collected from survey
Bid ($)
400
80
Bid Sensi vity
300
100
60
200
40
100
20
0
Country
State
City
Zip
Address
0
% Change in Budget with increasing obfusca on
Budget required for specified u lity 40 36 32 28 24 20 16 12 8 4 0
140
Random Const Truthful Greedy Seq Truthful Greedy
Seq Greedy 0
20
40
60
80
100
120
140
U lity specified
OpenSense II
160
% Change in Budget (abs)
Bids ($) and Sensi vity for sharing loca on
Budget required
500
Sensi vity
-
% Loss from Truthfulness
120 100
% Loss from Privacy
80 60 40 20 0
% Gain from Adap vity 0
10
20
30
40
50
60
70
80
Obfusca on Level (in miles radius)
90
100
TINYGS N : M OBILE A CTIVITY S ENSING
•
TinyGSN Android app –
•
Tracking user location in the city
GSN principles (wrapper, virtual sensors, streamElements)
Human-as-Sensor –
•
Eberle et al., poster
activity recognition, correlation with air quality levels
Field Experiments –
Goal: 24h of recording with one battery charge
Detecting activity levels: stationary, running, walking, etc
OpenSense II
Health Impact Studies Air Pollution Mapping
Crowdsensing
Data Quality Mobile Sensor Networks
OpenSense II
S ENSOR C ALIBRATION
Hop-by-hop GMR algorithm • Calibrates network of noisy, unstable sensors • Leverages meeting points between any pair of sensors • Based on Geometric Mean Regression (GMR) – No regression dilution problem (no calibration bias) – Low error-accumulation over multiple hops – Resistant against sensor noise
Simulated evaluation and comparison against standard methods
Evaluation on OpenSense data set obtained with Zurich deployment [O. Saukh, D. Hasenfratz, L. Thiele, Reducing MultiHop Calibration Errors in Mobile Sensor Networks, IPSN’15]
OpenSense II
IMPACT
OF
MOBILITY
ON
passive open sampling system
M EASUREMENTS active hybrid sampling system
• Electric car node used for studying impact of mobility on measurements
• Testing in parallel of novel sampling systems
high traffic NABEL
• Flexible platform for controlled experiments
park low traffic
• On-line monitoring of system parameters
OpenSense II
IMPACT
OF
MOBILITY
ON
NETWORK C OVERAGE
hourly
daily
weekly
monthly
• Uncontrolled mobility leads to dynamically changing coverage • Data-driven probabilistic coverage of street segments of Lausanne • Modeling framework needed for filling in the gaps in space and time [Arfire et al., SenSys’15, submitted]
Probability of coverage of street segments in Lausanne
OpenSense II
PEER INCENTIVE M ECHANISMS
FOR
H IGH-QUALITY C ROWDSENSING Radanovic et al., poster
How to ensure data quality from the Crowd? data
Peer Truth Serum • •
Sensor
rewards more ‘surprisingly common’ reports discourages colluding and random reporting
Payoff
data
data Aggregator
Participants contribute their data for a reward
Collude on Random Collude on Truthful 1 value reporting 2 values reporting
[Radanovic & Faltings, AAAI’14], [Radanovic & Faltings, AAAI’15], [Radanovic & Faltings, AAMAS’15]
OpenSense II
Health Impact Studies Air Pollution Mapping
Crowdsensing
Data Quality Mobile Sensor Networks
OpenSense II
D ATA-DRIVEN POLLUTION M APPING
Arfire et al., poster
IDEA: Use sensor measurements in conjunction with other explanatory data to interpolate/extrapolate pollution data Street segment-based space discretization better suited than grid-based Types of models considered so far: • Log-linear regression – NABEL station & meteorological explanatory variables only • Network-based log-linear regression – explanatory variables + measurements on other segments • Probabilistic Graphical Model
[Marjovi et al., DCOSS’15, to appear]
OpenSense II
DISPERSION MODELING: T HE GRAMM/GRAL M ODELING SYSTEM Berchet et al., poster
Setup •
•
GRAMM (Graz Mesoscale Model): mesoscale flow simulations for city region (100 m resolution) accounting for topography & land use GRAL (Graz Lagrangian Model): flow and air pollutant dispersion simulations at building resolving scale (5 m) for city area forced by GRAMM meteorological data
Achievements • •
Successful setup for Lausanne, including preparation of emissions and other inputs Hourly maps for NO2 pollution generated, good match with observations
Next steps • •
Extension to other pollutants (e.g., PM) Setup for Zürich [Berchet et al., EGU 2015]
OpenSense II
Health Impact Studies Air Pollution Mapping
Crowdsensing
Data Quality Mobile Sensor Networks
OpenSense II
INTEGRATION
WITH
PARALLEL H EALTH STUDIES
Analyze association between health and pollution exposure Health data (N=6184, 2003-2009)
Exposure data Link data by GIS
Air pollution map based on OpenSense II dispersion model
(N=1100, 2009-2012)
Preliminary work has shown association with blood pressure and renal function: • Air pollution and blood pressure (CoLaus & Bus Santé) •
Positive associations of pulse pressure and systolic blood pressure with short-term PM10 exposure.
[Tsai et al., Journal of Hypertension, 2015] • Air pollution and renal function and related phenotypes (CoLaus & SKIPOGH) •
Association of increased PM10 levels with increased levels of selected urinary protein among women
OpenSense II
PILOT S TUDY E XPOSURE
ON
PHYSICAL A CTIVITY
VS.
A IR POLLUTION Tsai et al., poster
Design of a pilot study about physical activity on exposure to air pollution
Report on recommendation Once the pilot study is complete, we will send the volunteers recommendation reports
OpenSense II
“HROUTING” - H EALTH- OPTIMAL ROUTE PLANNER Saukh et al., poster
Uses UFP pollution maps developed in OpenSense to compute healthy routes for pedestrians and cyclists in Zurich
• • •
Open source, free on Android/iOS >350 downloads so far Received much attention in local media
[D. Hasenfratz, T. Arn, I. de Concini, O. Saukh, L. Thiele, Demo-Abstract: Health-Optimal Routing in Urban Areas, IPSN’15]
OpenSense II
C ONCLUSIONS •
OpenSense II seeks to integrate different air quality sensing platforms, including vehicular sensing networks, traditional monitoring stations, and novel crowdsensing platforms
•
A solution for a viable crowdsensing platform still needs to be found.
•
Simulation-based efforts for developing techniques for managing crowdsensing have seen significant progress.
•
Data-quality is critical when considering mobile measurements; more so if the measurements are crowdsourced.
•
Deriving accurate high resolution maps is critical for assessing correlation between population health and pollutant exposure.
OpenSense II
OPENS ENSE II T EAM Alcherio Martinoli, EPFL-DISAL, PI • Adrian Arfire, PhD student • Emmanuel Droz, engineer • Ali Marjovi, postdoc Karl Aberer, EPFL-LSIR, Co-PI • Berker Agir, PhD student • Jean-Paul Calbimonte, postdoc • Julien Eberle, PhD student • Tian Guo, PhD student • Mehdi Riahi, PhD student Murielle Bochud, CHUV-IUMSP, Co-PI • Dai-Hua Tsai, postdoc Lukas Emmenegger, EMPA, Co-PI • Antoine Berchet, postdoc • Dominik Brunner, senior researcher • Christoph Hüglin , senior researcher • Michael Müller, postdoc • Katrin Zink, postdoc Boi Faltings, EPFL-LIA, Co-PI • Goran Radanovic, PhD student Andreas Krause, ETHZ-LAS, Co-PI • Adish Singla, PhD student
Michael Riediker, IST, Co-PI • Nancy Hopf, senior researcher • Guillaume Suarez, postdoc • Nicole Charrière, technical staff Lothar Thiele, ETHZ-TIK, Co-PI • David Hasenfratz, PhD student • Balz Maag, PhD student • Olga Saukh, postdoc
OpenSense II