Integration and calibration of non-dispersive infrared (NDIR) CO 2 low-cost sensors and their operation in a sensor network covering Switzerland

Michael Müller, Peter Graf, Jonas Meyer, Anastasia Pentina, Dominik Brunner, Fernando Pérez‐Cruz, Christoph Hüglin, Lukas EmmeneggerAtmospheric 2020 · DOI

Abstract

More than 300 non-dispersive infrared (NDIR) $CO_2$ low-cost sensors labelled as LP8 were integrated into sensor units and evaluated for the purpose of long-term operation in the Carbosense $CO_2$ sensor network in Switzerland. Prior to deployment, all sensors were calibrated in a pressure and climate chamber and in ambient conditions co-located with a reference instrument. To investigate their long-term performance and to test different data processing strategies, 18 sensors were deployed at five locations equipped with a reference instrument after calibration. Their accuracy during 19 to 25 months deployment was between 8 and 12 ppm. This level of accuracy requires careful sensor calibration prior to deployment, continuous monitoring of the sensors, efficient data filtering, and a procedure to correct drifts and jumps in the sensor signal during operation. High relative humidity (> ~85 %) impairs the LP8 measurements, and corresponding data filtering results in a significant loss during humid conditions. The LP8 sensors are not suitable for the detection of small regional gradients and long-term trends. However, with careful data processing, the sensors are able to resolve $CO_2$ changes and differences with a magnitude larger than about 30 ppm. Thereby, the sensor can resolve the site-specific $CO_2$ signal at most locations in Switzerland. A low-power network (LPN) using LoRaWAN allowed for reliable data transmission with low energy consumption and proved to be a key element of the Carbosense low-cost sensor network.

Introduction

High-precision $CO_2$ monitoring typically relies on expensive cavity ring-down spectrometers that cost tens of thousands of dollars. While these are the gold standard, their cost limits the spatial density of a network. The Carbosense network addresses this by deploying over 300 low-cost non-dispersive infrared (NDIR) sensors across Switzerland's diverse topography, creating one of the largest dense sensor networks for $CO_2$ monitoring.

Map of the Carbosense network across Switzerland, showing the distribution of low-cost sensors (red/yellow dots) and reference sites (blue dots).

The fundamental challenge investigated is whether sophisticated data processing can transform "low-cost" data into scientifically valuable information. Low-cost sensors (LCS), like the Senseair LP8, are sensitive to temperature and pressure, prone to drift, and can exhibit sudden "jumps" in data due to environmental factors.

Sensor Integration and Methodology

The team integrated the Senseair LP8 sensor into a custom battery-powered unit designed for long-term field deployment. The hardware design was optimized for low power consumption and environmental protection.

Hardware design of the LP8 sensor node: (a) photograph of the open unit showing batteries and (b) schematic diagram illustrating internal components and dimensions.

The data quality strategy involved several key steps:

  1. Individual Calibration: Every sensor underwent rigorous calibration in pressure and climate chambers at Empa and METAS, followed by ambient co-location with a reference instrument.
  2. Physical Modeling: An extended model based on the Beer-Lambert law was derived to explicitly account for temperature ($T$), pressure ($p$), and detected light intensity ($I_1$), compensating for optical and thermal artifacts.
  3. Drift Correction: An automated "windy-period" algorithm was developed to identify well-mixed atmospheric conditions. During these windows, regional $CO_2$ gradients are small, allowing sensors to be "re-anchored" to the nearest high-precision reference station.
  4. LoRaWAN Infrastructure: The use of a Low Power Wide Area Network (LoRaWAN) enabled data transmission every 10 minutes for over five years on just two standard alkaline batteries.

Key Insights and Results

The study revealed that humidity is a critical factor; relative humidity (RH) exceeding ~85% causes signal degradation due to condensation on sensor mirrors, requiring strict data filtering. Furthermore, the research emphasized that each low-cost sensor is unique, requiring individual lab calibration rather than a global correction factor.

Post-calibration and drift correction, the sensors achieved an accuracy of 8–12 ppm, a significant improvement over the manufacturer's raw spec of ±50 ppm. While these sensors cannot detect subtle 1 ppm climate gradients, the network's density allows it to resolve site-specific signals (like urban plumes or valley accumulations) larger than 30 ppm.

Conclusion and Future Work

The Carbosense network demonstrates that with continuous monitoring and automated drift correction, low-cost sensors can remain useful for over two years in the field. Future directions include integrating atmospheric transport models to better identify air mass consistency between sensors and reference instruments, which would further refine drift correction in regions with complex local emissions.

Summary updated

Abstract

More than 300 non-dispersive infrared (NDIR) CO2 low-cost sensors labelled as LP8 were integrated into sensor units and evaluated for the purpose of long-term operation in the Carbosense CO2 sensor network in Switzerland. Prior to deployment, all sensors were calibrated in a pressure and climate chamber and in ambient conditions co-located with a reference instrument. To investigate their long-term performance and to test different data processing strategies, 18 sensors were deployed at five locations equipped with a reference instrument after calibration. Their accuracy during 19 to 25 months deployment was between 8 and 12 ppm. This level of accuracy requires careful sensor calibration prior to deployment, continuous monitoring of the sensors, efficient data filtering, and a procedure to correct drifts and jumps in the sensor signal during operation. High relative humidity (> ∼85 %) impairs the LP8 measurements, and corresponding data filtering results in a significant loss during humid conditions. The LP8 sensors are not suitable for the detection of small regional gradients and long-term trends. However, with careful data processing, the sensors are able to resolve CO2 changes and differences with a magnitude larger than about 30 ppm. Thereby, the sensor can resolve the site-specific CO2 signal at most locations in Switzerland. A low-power network (LPN) using LoRaWAN allowed for reliable data transmission with low energy consumption and proved to be a key element of the Carbosense low-cost sensor network.
BibTeX
@inproceedings{muller2020integration,
  author = {Michael Müller and Peter Graf and Jonas Meyer and Anastasia Pentina and Dominik Brunner and Fernando Pérez‐Cruz and Christoph Hüglin and Lukas Emmenegger},
  title = {{Integration and calibration of non-dispersive infrared (NDIR) CO <sub>2</sub> low-cost sensors and their operation in a sensor network covering Switzerland}},
  booktitle = {Atmospheric},
  year = {2020},
  doi = {10.5194/amt-13-3815-2020},
}
Figure