Modeling fine-grained spatio-temporal pollution maps with low-cost sensors

The use of air quality monitoring networks to inform urban policies is critical especially where urban populations are exposed to unprecedented levels of air pollution. High costs, however, limit city governments’ ability to deploy reference grade air quality monitors at scale; for instance, only 33 reference grade monitors are available for the entire territory of Delhi, India, spanning 1500 sq km with 15 million residents. In this paper, we describe a high-precision spatio-temporal prediction model that can be used to derive fine-grained pollution maps. We utilize two years of data from a low-cost monitoring network of 28 custom-designed low-cost portable air quality sensors covering a dense region of Delhi. The model uses a combination of message-passing recurrent neural networks combined with conventional spatio-temporal geostatistics models to achieve high predictive accuracy in the face of high data variability and intermittent data availability from low-cost sensors (due to sensor faults, network, and power issues). Using data from reference grade monitors for validation, our spatio-temporal pollution model can make predictions within 1-hour time-windows at 9.4, 10.5, and 9.6% Mean Absolute Percentage Error (MAPE) over our low-cost monitors, reference grade monitors, and the combined monitoring network respectively. These accurate fine-grained pollution sensing maps provide a way forward to build citizen-driven low-cost monitoring systems that detect hazardous urban air quality at fine-grained granularities.

calibration process. Kaiterra products, however, use over 6,000 individual measurements during the calibration and verification process. From testing chamber design to performance certification, we created a process that verifies that every sensor is properly calibrated before and after assembly. We have designed a testing environment that goes well beyond the required industry standards. Below, we outline the 3 components of our testing environment.
1. Testing chamber: An air-tight testing chamber is used to simultaneously test up to 360 monitors over a set amount of time and in varying conditions. This custom-made testing chamber is 90 cubic meters. Monitors are mounted on angled shelving that has been specifically designed to allow high airflow between the devices, preventing pollutant buildups from forming within the test chamber -the uniformity of the air within this chamber is extremely important. To ensure that each device is tested under consistent conditions, ventilation equipment controls the temperature and amount of the pollutant released.
2. Careful calibration: Certain sensors require calibration during the manufacturing process.
For these parameters, specific concentrations of pollutants are introduced into the test chamber. This process is carried out remotely outside the test chamber so that no human interference can influence the concentration of pollutants inside. Once the concentration of pollutant has stabilized within the chamber and is uniform throughout, all devices within the chamber will be remotely calibrated to this concentration.
3. Comprehensive measurement validation: Every single device that comes off of the production line must be fully tested for particulate matter over several hours, and all readings must fall within specification. While most air quality monitors on the market are tested over two or three different levels of concentration and only a handful of data points are used, the device takes over 6,000 individual measurements during the test process. Each of these mea-2 surements are processed onboard and sent to the cloud in real-time. Software in the cloud will process these readings and compare them to measured reference values within the test chamber. The divergence between the monitor's measured readings and the reference readings are calculated for every single data point and verified against the specifications of the device.
Only devices that fulfill the published technical specifications will leave the testing chamber.
Any failed devices will enter a separate process to determine where the failure occurred, and once identified, to resolve it, before returning to the production line to be fully re-tested. Professional-grade devices will usually allow users to enter in these details manually, but knowing the density of the particles being measured is no easy feat, and usually requires sending a sample of particles to a lab for analysis. In real life, the density and optical properties of particles will be heavily affected by geography. The particles are extremely different in various geographical areas, and using the wrong factor in calculating mass concentrations will lead to results that can be way off.

Localized Cloud Calibration
Simply using a default factor for all devices means that results will be wrong more often than they are right. Seeing that each city has a different pollution profile, the location will affect the accuracy of the device. Most devices need to be calibrated by hand using data sheets from each city or sent to a professional calibrating service, our device can get all of this data from a central server and make the necessary tweaks automatically. The key to long-term accurate readings is having an algorithm that can take these changing factors into account in real-time.We set up numerous colocations around the globe with reference grade monitors that provide particulate readings as mass 3 concentrations (e.g. BAMs and TEOMs). We then analyze the readings of our optical monitors against these devices in real-time. From this data, we are able to identify the key properties necessary to update the algorithm inside our devices. Based on the location selected for a device, the device will be provided with an updated calibration algorithm every five minutes. This means that as the wind blows from one side of a town to the other, and as the makeup of particles in the air changes, the algorithm inside every device based in that town will change accordingly, in realtime.The correlation between the outdoor unit and the reference grade E-BAM was also analyzed as shown in Supplementary Figure 6.
Consistency between sensors: Since the readings from devices may drift apart due to various factors, the sensors were randomly tested for consistency. They were taken from the site at random and the tests were conducted for consistency between them. Two sensors were compared by locating them in the same testing environment and making measurements for 7 days. The 2 sensors produced very similar readings in the comparison. Data was collected every 60 sec for 7 days, resulting in 10,080 readings. The full dataset was plotted for PM2.5 and correlation coefficient was checked.
The regression equations showed very good agreement between the sensors for both PM2.5 along with the R 2 in Supplementary Figure 7  The impact of sensor network size on prediction error. The blue line shows the errors for our low-cost sensors, and the black for the government monitors. We see that the more sensors we use in our model, the better the performance of the model in terms of the prediction error. The error flattens out about 30 sensors, which is approximately the number of sensors of each type that we have in our experiment. We infer that having an even denser deployment likely adds little value to the predictive performance.  Figure 4: Temporal Cross-correlation heatmap between our sensors grouped and ordered by the corresponding nearest government sensor in Delhi. This further shows that on average, correlation in nearby sensors (closer to diagonal) are higher than ones that are further away (away from diagonal), thereby showing spatial variability in pollution in Delhi.