Objectives

 

Carbon dioxide (CO2) is the most important greenhouse gas (GHG) contributing to climate change. In order to better predict future CO2 levels and the corresponding climate forcing, it is thus vital to better understand the global carbon cycle. The combination of a dense measurement network, atmospheric transport modelling and statistical modelling has the potential to improve our knowledge of CO2 fluxes. Eventually, it could provide a tool that independently assesses the effect of greenhouse gas reduction initiatives of cities and countries.

Carbon dioxide concentration can be computed using atmospheric transport models.  These models are similar to models employed for operational weather prediction and simulate the transport of CO2 in the atmosphere taking into account spatio-temporal information about emissions and biosphere fluxes. However, the results might be biased due to limited quality of the input data such as emission inventories and model’s inability to represent subgrid-scale CO2  variability. Statistical methods that are only based on atmospheric observations and input data such as topography, land-use, vegetation or traffic are a complementary approach to obtain the spatial distribution of CO2 concentration. In principle, this approach can provide a map of the instantaneous CO2 concentration in a region and, thus, link observations and results from atmospheric models. However, most of the current operational CO2 measurement networks have a much sparser spatial coverage then is required for this modelling approach. The operation of a large number of low-cost CO2 sensor units that are battery powered and equipped with wireless data transmission capabilities in a dense sensor network is appealing. However, in general, the accuracy of such sensors is significantly below that of traditional instruments.

The goal of CarboSense4D is to produce an accurate map of the evolution of carbon dioxide (CO2) over Switzerland at high temporal and spatial resolution by applying machine learning methods to observational data from a network of low-cost sensors and combining it with atmospheric simulations and atmospheric transport modelling. The project is led by the Swiss Federal Laboratories for Materials Science and Technology (Empa) and involves the partners Swisscom, Decentlab and the Swiss Data Science Center (SDSC)

Sensor network

 

The Carbosense CO2 sensor network consists of more than 250 measurement nodes and covers the whole of Switzerland (Figure MAP). It has a special focus on the city of Zurich where about 50 nodes are deployed. The network is operational since July 2017. It consists of three types of sensors:

  1. seven high-precision laser spectrometers (Picarro G1301/G2302/G2401)
  2. 15 temperature stabilized, mains powered NDIR low-cost instruments with reference gas supply (SenseAir HPP)
  3. 250 nodes of battery-powered NDIR low-cost sensors (SenseAir LP8).

 

The low-cost sensor units were engineered by Decentlab GmbH. Each unit is a relatively small box (Fig 1.) containing a SenseAir LP8 sensor, a Sensirion SHT21 sensor, a LoRaWAN communication module, a microprocessor, and two batteries for power supply. The LP8 sensor reports the infrared measurement, a CO2 molar fraction based on factory calibration, temperature, and its status. The SHT21 sensor measures temperature and relative humidity. The measurement frequency was set to 1 minute for all the sensors. The measurements are transmitted as 10 minute averages over Swisscom’s Low Power Network (LPN; based on LoRaWAN).

Figure 1: CO2 low-cost sensor unit and LP8 sensor (in the front).

The sensor network provides measurements from many different locations, including urban, suburban and rural areas, forests, and mountain areas.   Thus we can observe different dynamics of daily variations in CO2 concentration depending on the location and the prevailing meteorology. For example, on Fig 3 one can see measurements from 5 locations in Zurich area in one week of June 2019. Location ALBS is quite elevated  (around 810m) and demonstrates low daily variation of CO2 concentration. In contrast, RECK is located at much lower altitude (around 445m) in suburbs of Zurich and one can see much more pronounced daily variation in carbon dioxide concentration. Also, density of the network in the city of Zurich allows us to capture effects of traffic emissions. In particular, ZUE and ZSCH are two very close locations (distance about 850 m). While site ZUE is located in a courtyard situation, site ZSCH is located next to a busy road. The impact of the traffic is clearly visible in the time series.
Figure 2: Carbosense sensor network (as of 2 October 2019). Red dots depict LP8 sensor locations, yellow dots depict HPP sensor locations and dark blue dots depict locations of Picarro instruments.

Figure 3: Measurements of CO2 concentration  from 5 locations, averaged per hour, in June 2019.

Sensor network provides temporally resolved information about carbon dioxide concentrations, but only at a limited number of locations. In order to be able to predict CO2 in other locations we build a statistical model based on Gaussian processes [1]. It is based on the idea that CO2 concentration changes smoothly in space (but possibly differently in different directions). In addition it accounts for the effects elevation, traffic emissions and level of vegetation in the area. Focusing on the area of the city of Zurich we were able to produce a model of hourly CO2 concentrations with 100m spatial resolution.

 

Reinhard Bischoff,Decentlab

Dominik Brunner, EMPA

Lukas Emmenegger, EMPA

Michael Jaehn, EMPA

Jonas Meyer, Decentlab

Michael Muller, EMPA

Anastasia Pentina,SDSC

Fernando Perez-Cruz, SDSC

 

[1] Carl Edward Rasmussen and Christopher K. I. Williams, “Gaussian Processes for Machine Learning”, 2005

 

* picture for the thumbnail : https://www.empa.ch/web/s604/carbosense4d