The measurement unit (Fig. A) houses all components needed for the operation of the GMP343 sensor. The main air intake is filtered by a 2 µm particulate filter (Fig. A2) to prevent contamination. For internal compensation, the GMP343 sensor receives relative humidity readings inflow from a Sensirion SHT45 sensor (Fig. A4) and pressure readings from a Bosch BME280 sensor (Fig. A4). Although the BME280 also provides relative humidity readings, the SHT45 offers better accuracy, particularly under very low relative humidity conditions. To enable operational resilience, the system incorporates an uninterruptible power supply (UPS) and a battery (Fig. A7), which provides one hour backup power during outages. A Raspberry Pi 4 (Fig. A8) serves as the central controller of the system, managing all sensors and actuators. For remote data transmission, the system is equipped with a cellular 4G module (Fig. A8). A custom mainboard (Fig. A9) integrates terminal block sockets, serves as the main power distribution hub for all components, and houses a second BME280 sensor.

The measurement unit (Fig. A) is integrated in an outdoor enclosure (Fig. A). The enclosure is modified with additional cartridge silicone sealing to protect the components from all local weather conditions. The inner walls are insulated with ArmaFlex to improve thermal stability. For calibration, two 10 L aluminum calibration cylinders (Fig. B11) are integrated. These cylinders are provided by Westfalen Gas. We used a gas mixture consisting of synthetic air (80 % Nitrogen Grade 6.0, 20 % Oxygen Grade 6.0), spiked with CO2 (400 and 520 ppm Carbon-dioxide Grade 3.0), with a precision of about ±2% and stability of 36 months. The exact CO2 concentration is determined using a Picarro instrument calibrated against the WMO CO2 X2019 calibration scale , thereby transferring the calibration scale to the calibration cylinders. The cylinders are connected to the system through demand flow regulators (Fig. B12), which release gas flow upon application of underpressure by an eccentric diaphragm pump (Fig. A5). The system is designed to have approximately 0.5 Lmin-1 of flow during both calibration and ambient measurements. An additional advantage of demand flow regulators is that, while they are not in use, only the regulator is pressurized, thereby minimizing the risk of leakage from downstream tubing or valves. Four 2/2 solenoid valves (Fig. A3) allow to switch between the main air intake and up to three additional lines. An external LTE antenna mounted on the aluminum outdoor enclosure provides 4G cellular reception. An industrial grade 350 W control cabinet heater (Fig. B14), a control cabinet fan (Fig. B13), and Arduino Nano (Fig. B15) with hysteresis control provide an internal temperature of 25±3°C. At low ambient temperatures, the heater is activated to maintain the target range, while at higher temperatures, which are primarily the result of solar-induced heat buildup, the ventilation system prevents overheating.

To allow flexible placement on structures such as poles in an elevated rooftop position, the system includes an external air intake (Fig. B). Accompanying the intake, a Vaisala WXT-532 wind sensor (Fig. B) records wind speed and direction, which are transmitted together with the CO2 data. It can be connected to the outdoor enclosure (Fig. A) with a maximum of 50 m of data cable and air tubing.

The choice of tubing diameter and flow rate is important to ensure representative air sampling, particularly when longer intake lines are used. In our configuration, ambient air is drawn through 50 m of 1/4′′ (Outside Diameter (OD): 6.35 mm, Internal Diameter (ID): 4.3 mm) SERTOflex tubing at a flow rate of 0.5 Lmin-1. To verify the suitability of this setup, we calculated the Reynolds number, residence time, and pressure drop (see Appendix ). The corresponding Reynolds number of approximately 160 confirms laminar flow. Under these conditions, the residence time is about 1.45 min, which is acceptable given that the network provides hourly averaged data rather than high-frequency observations. The calculated pressure drop of 0.09 hPa is negligible compared to the approximately 15 hPa pressure drop introduced by the 2 µm inlet filter.

In practice, the installed tubing lengths vary depending on local rooftop conditions and typically range between 10 and 20 m, resulting in residence times of only 17–34 s.

Based on the experience gained during the first year of deployment (Sect. ), we built a prototype for a second generation system (Fig. C). The main focus was on improving the temperature control and maintainability in the field.

We removed the enclosure of the measurement unit (Fig. B10) and directly mounted individual components on DIN rails (Fig. C2, 3, 8, and 9) to enable quickly changing components in the field. By reducing the number of brass connections and integrating Swagelok tube connectors, we simplified the process of creating an airtight flow through the system.

The main improvement is the addition of a dedicated PID-controlled temperature-stabilized sensor chamber with a temperature sensor (Fig. D17), a 10 W positive temperature coefficient (PTC) heater (Fig. D18), and a small ventilator (Fig. D19). This fully enclosed and small-volume aluminum box maintains the sensor environment at a configurable target temperature ±0.1°C. A high temperature target eliminates the effects of solar radiation during the hot summer months. The cabinet heater (Fig. C14) and the ventilator (Fig. C13) are unchanged to prevent condensation or freezing, and to ensure a stable environment for the calibration cylinders (Fig. C11). We refer to the updated sensor enclosure as second-generation hardware (also termed v2).

The bill of material of the system for generation 1 is EUR 7800, with approximately 50 % attributed to the Vaisala GMP343 sensor and the WXT-532 wind sensor. This excludes consumables, labor, and operational costs. Generation 2 reduced the total system cost to EUR 7300.

Each field station runs the Python-based ACROPOLIS edge software on a Raspberry Pi, consisting of a continuously running gateway process and a containerized version controlled controller software managed by the edge gateway.

The edge gateway is a standalone process that ensures the availability of the station 24/7. It handles MQTT communication with the backend, serves as the endpoint for remote procedure calls, and manages the deployment and versioning of the edge controller container. After successful transmission to the MQTT broker, a copy of the measurement data is archived locally via SQLite and can be re-uploaded in case of transmission failure or data loss. In addition, selected log messages and health checks are also forwarded to Thingsboard for real-time monitoring and diagnostics. The gateway process monitors the software version assigned to it via ThingsBoard and upon receiving a new software version assignment, fetches the specified version from GitHub and redeploys the controller software using Docker.

The edge controller manages all connected devices through hardware-specific interfaces, organized into reusable and configurable modules. A key feature of the controller is its ability to perform on-device processing for dilution and calibration correction. This allows the system to transmit corrected measurement data in real time, making it immediately available to users and operators through the ThingsBoard dashboard, without the need for additional downstream processing. Measurement data are written to a persistent queue based on a SQLite database that is continuously monitored and processed by the gateway.

For cloud-side device management, data collection, and visualization, the project uses ThingsBoard , a scalable open-source IoT platform. It acts as the central backend for telemetry exchange, data storage, and real-time visualization through customizable dashboards. ThingsBoard supports remote procedure calls, rule-based alerting, and lifecycle management of connected devices, among other features. Programmatic access is enabled through a REST API. Data transmission between field stations and the backend is secured using TLS.

Measurement data from all stations are transmitted with a frequency of 10 s via MQTT. Internally, Apache Kafka handles asynchronous message queueing and buffering. Measurement data are stored in a PostgreSQL database. ThingsBoard was selected based on its maturity and comprehensive feature set after earlier prototyping with Hermes and Tenta .

To support local data validation, correction, and export, the post-processing pipeline is composed of two dedicated tools: the ThingsBoard Downloader and the ACROPOLIS-data-processing module. Both tools support a uniform and reproducible processing workflow, ensuring that all sensor outputs undergo the same calibration corrections, quality checks, and formatting procedures. This standardization allows reproducible comparison between systems.

The ThingsBoard Downloader enables authenticated access to the backend via the ThingsBoard REST API using JWT tokens for authentication. Users can select specific stations and telemetry keys and download compressed measurement datasets for offline analysis. This local copy of the data simplifies data handling and decouples local processing from backend operations.

The ACROPOLIS-data-processing module uses the Polars library for tabular data operations, enabling efficient handling of large datasets. Additional Jupyter Notebooks support calibration tank characterization and performance evaluation with deployed reference instrument. These tools complement the automated pipeline by enabling manual inspection, exploratory analysis, and quality assurance across the entire measurement network.

Sensor settings include activated optics heating, linearization, and 10 s averaging of output data. Supplier-provided sensor internal compensation for temperature, relative humidity, pressure, and oxygen is enabled. Relative humidity and pressure readings are updated live by external sensors in the gas flow at a sampling frequency of 0.1 Hz. A moving average is used over 15 consecutive measurements to smooth short-term fluctuations in the low-cost sensor output. During the injection of dry calibration gas, the humidity of the GMP343 sensor is set to 0 %, while the offset of the SHT45 sensor is determined. The offset is then corrected for until the next calibration event.

The dry air mole fraction of CO2 is calculated in three steps. First, by approximating the saturation water vapor pressure using the Wagner equation (Eq. ). 1pσ=pc⋅exp⁡TcT⋅(a1θ+a2θ1.5+a3θ3+a4θ3.5+a5θ4+a6θ7.5) with θ=(1-T/Tc), Tc=647.096K, pc=22.064×106Pa, a1=-7.85951783, a2=-1.84408259, a3=-11.7866497, a4=22.6807411, a5=-15.9618719, a6=1.8012250 .

Then the mole fraction of water vapor is calculated (Eq. ) according to Dalton's law of partial pressures, by dividing the partial pressure of the water vapor by the total ambient pressure. The ambient pressure (Ptotal) is supplied by the BME280 sensor, while the relative humidity (RH) is read from the Sensirion SHT45. Finally, CO2dry is calculated (Eq. ) using the water vapor mole fraction. 2xH2O=PH2OPtotal=pσ⋅RH100Ptotal3CO2dry=CO2wet1-xH2O

Uncertainties in the humidity measurements can influence the correction from wet to dry CO2 mole fractions. To quantify this effect, we evaluated the impact of the SHT45 humidity sensor's specified accuracy (±1% RH) under representative environmental conditions. Using the Wagner equation, the resulting uncertainty in the dry CO2 mole fraction is approximately 0.04 ppm under cool and dry conditions (5 °C, 10 % RH) and about 0.20 ppm under warm and humid conditions (30 °C, 80 % RH). Across all relevant conditions, this effect remains negligible compared to the overall target precision of the system.

Each system is calibrated using an automated 2-point calibration procedure with two synthetic gas calibration bottles filled with approximately 400 and 520 ppm CO2. The calibration procedure is performed every other day at 03:00 LT using an alternating bottle sequence. At this time, the ambient temperatures are consistently below the target (25±3°C for v1, 40±0.1°C for v2), allowing the control cabinet heater to maintain stable internal conditions. We use the first bottle in the sequence to dry the air channel before the calibration, and alternating cylinders balance the usage of cylinder gas for drying over time. Each calibration bottle is sampled for 10 min, preceded by a 10 min pre-drying phase and a 5 min flushing period at 2 Lmin-1 before switching back to measurements. The total sequence takes 35 min.

The calibration parameters for slope (m) and intercept (b) are calculated by taking the median of both cylinder runs and are automatically validated after each calibration event to be within ±10% of the last valid run or are ignored otherwise. The state of all calibrations can be assessed through a ThingsBoard Dashboard for Network Operation, which also offers an option for automated alarms. 4CO2cal=m⋅CO2dry+b

The collected data is downloaded using the ThingsBoard Downloader and aggregated to 1 min data and further processed by the post-processing pipeline (Sect. ). Outliers are detected and flagged using a Hampel filter configured at a 3σ threshold in a sliding window of 120 min. It uses a Median Absolute Deviation (MAD) to detect values that deviate significantly from the surrounding data.

Although careful site selection reduces the influence of nearby emission sources, some level of local contamination is unavoidable in dense urban environments. The Hampel filtering therefore provides a first automated diagnostic of short-lived disturbances. Following this step, the Level 2 (L2) data set is produced through manual operator validation, during which persistent or recurrent anomalies attributable to identifiable local sources are flagged. The resulting one-minute and hourly averaged L2 data supplement the automated filtering L1 data. Both L1 and L2 datasets are uploaded to the ICOS Cities Portal, with L1 available continuously and L2 released once per year.

In our implementation, no data points are removed. Potential contamination events are identified by comparing the Hampel filter output with the original signal and flagging deviations directly on the original time series. This approach allows users with different scientific objectives, for example those focusing on short-term plumes, to apply their own filtering or thresholding strategies using the published dataset.

Following the deployment strategy, 20 stations were installed across 17 sites in urban, suburban, and rural locations in and around Munich, covering all major districts. The stations were placed indoors (e.g. elevator shafts, technical rooms) or directly outdoors on the roof. Schools and hospitals offered ideal conditions, as they are typically free-standing, elevated structures distributed throughout the city and are overseen by a centralized authority.

The present deployment (Fig. ) includes six school sites (MAIR, PASR, SCHR, BALR, SENR, TAUR), and four hospitals (RDIR, HARR, BOGR, NPLR). Other deployed sites are two city halls (FINR, FELR), private institutions (SWMR, BLUT), and research buildings (DLRR, TUMR, GROR). Although most of the sites rely on district heating or detached block heating, two locations (MAIR, BALR) have close-by gas heating infrastructure, which can influence the measurement signal under specific conditions, such as low wind speeds. We placed inlets in a way to support uncontaminated measurements from the dominant wind directions. A particularly notable site is the Blutenburg Tower site, which has two systems with inlets at 48 and 85 m above ground level, allowing for vertical profile measurements.

On the TUM campus, the generation 1 and the generation 2 ACROPOLIS system are deployed along with a reference Picarro G2401 instrument to evaluate the long-term performance and temporal stability of the updated temperature stabilization.

Detailed information on deployment sites can be found in Table  and on site profiles on the ICOS Cities Portal .

Some degree of local influence is unavoidable in dense urban environments, even with careful site selection. The availability of suitable installation locations is limited, and the deployed stations therefore represent the best feasible choices within these constraints. Remaining local effects are handled in the post-processing workflow described in Sect. , which includes automated Hampel filtering and manual validation in the Level 2 data products. The percentage of one-minute values flagged by the Hampel filter varies between sites and provides a first indication of short-lived local disturbances. An overview of the percentage of flagged data for each site can be found in Table  in Appendix .

All stations are connected via LTE and support remote management, including software updates, configuration changes, and system diagnostics. Our modular network architecture allows for easy future expansion and quick changes of selected sites. Routine site visits are conducted annually to replace calibration gas tanks, verify physical integrity, and inspect inlet and wind sensor positioning. Operational costs are kept low, with approximately EUR 10 per year per station for mobile communication, EUR 600 per year per station for calibration gas (assuming two tanks per station), and a total of EUR 400 per year for the shared cloud infrastructure supporting backend services.

Appendix  shows that R2 for all systems is ≥0.99, which indicates a good agreement for the sensor measurements and the reference instrument within the measured range. The results of the sensor performance (MAE, RMSE) are shown in Table  and in Fig. . The figure shows that the sensors achieve an accuracy of at least 1.1 ppm RMSE, with the exception of systems with a strong temperature sensitivity (ID 4, 8, 9, 10, 11, 15, 17, 19) indicated by the red dots. These sensors show a range of 1.1 to 2.6 ppm RMSE. System 17 is the system with the worst performance with an RMSE of 2.6 ppm, which is partially related to the damaged SHT45 sensor. The difference in performance based on temperature sensitivity is magnified in summer conditions, when solar radiation heats up the outdoor enclosure.

Based on these results, we developed a second-generation prototype featuring an additional temperature-controlled enclosure around the GMP343 sensor. This allows for more precise regulation and a higher temperature target, helping to eliminate heat build-up effects during summer and reducing the sensor's exposure to ambient temperature variability. We kept the 350 W cabinet heater for the outdoor enclosure and added a dedicated PID-controlled temperature-stabilized sensor box (Sect. ) that maintains the sensor environment at the target temperature ±0.1°C, while only requiring 10 W additional energy input. After assembling the prototype, we performed a second side-by-side comparison for the second-generation system.

Sensor 3 was one of the first systems deployed during the project, installed at the FINR site in late February 2024. Following a modem hardware failure, it was returned to the laboratory. Being among the remaining sensors at TUM in November 2024, Sensor 3 showed the strongest temperature sensitivity and was therefore selected for reassembly and integration into the generation 2 system.

Due to its early deployment, Sensor 3 did not participate in the April 2024 side-by-side comparison during warmer conditions. As a result, the v1 data set for this sensor is limited to January and February 2024, when the internal sensor temperatures remained below 31 °C.

Figure  shows the sensor temperatures for the side-by-side comparison period 2025 for the first-generation system (v1, ID 6) and the second-generation system (v2, ID 3). Both systems are located on the TUM rooftop next to each other and should experience comparable conditions. Furthermore, the outside temperature, measured by the LMU Meteo Station MUM01 (450 m direct distance) at 30 m above ground, is shown. The figure shows that the temperature control for generation 2 works as expected, while the temperature for generation 1 is affected by solar radiation in the months April and May. The data gap is a result of a power outage.

Figure  shows the performance of Sensor 3 in both the first-generation (v1, 2024) and the second-generation (v2, 2025) system versions. The sensor response to pressure and absolute humidity in generation 2 shows trends similar to those in generation 1, but with a visibly reduced spread and lower standard deviation. The stabilized temperature range of 42.5 to 42.7 °C in generation 2 minimizes the influence of temperature on sensor performance, and the temperature sensitivity observed in generation 1 is no longer evident. For this comparison period, the target temperature was set at 40 °C. The observed temperature of 42.6 °C is attributed to activated internal optical heating.

With these hardware upgrades, we could successfully achieve an improvement in sensor accuracy even in the presence of a broader range of environmental conditions for the second-generation system. Sensor 3 in the second generation system performed with an RMSE of 0.60 ppm, an MAE of 0.49 ppm, and a standard deviation of 0.52 ppm, indicating a clear improvement in precision compared to its performance in the first-generation system (RMSE 0.91 ppm, MAE 0.68 ppm, standard deviation 0.91 ppm). The generation 2 data set contains more than 2.5 times the number of data points and was exposed to a wider range of environmental conditions compared to the v1 data set, highlighting the observed performance improvements.

Although the target temperature of 40 °C worked well for the first half of 2025 the target temperature should be chosen to match local conditions and temperatures. We are currently experimenting with different temperature points for the hot summer months and first results are shown in Appendix .

For a robust assessment of intra-urban CO2 variability and urban-rural gradients, we restricted our analysis (Fig. ) to four sites with elevated inlet positions at approximately 30 m above ground level (a.g.l.). These inlets are less affected by immediate surface-level emissions and are more likely to capture air masses representative of the surrounding environment. This approach helps to ensure that observed differences reflect true spatial patterns rather than differences in measurement height or local source proximity.

The selected sites span different urban contexts across Munich: GROR is located at the southwestern edge of the city and frequently upwind of Munich with respect to the prevailing wind direction, making it a useful reference for incoming background air. TUMR, located in the urban core, serves as a representative high-density inner-city site. SWMR is located in the transition between urban and suburban environments, located between the main inner city highway (Mittlerer Ring) and the busy Dachauer Straße, and is surrounded by mixed land use, including residential areas and green spaces. NPLR, at the southern edge of the city, is adjacent to forested and agricultural land, offering a semi-rural footprint. For additional context and to approximate regional background concentrations, we also include data from the ICOS site Hohenpeißenberg (HPB), located about 50 km southeast of Munich in a rural mountain top location.

In spring and summer, TUMR, SWMR, and NPLR show similar afternoon minima, with overall concentrations lower in summer. We believe that the lower CO2 levels during this period are primarily driven by reduced emissions, and that enhanced photosynthetic uptake together with higher boundary-layer heights further contribute to the observed seasonal decrease. Nighttime peaks differ by site: In spring, SWMR shows strong early-morning accumulation, likely due to anthropogenic emissions trapped in a shallow boundary layer. In summer, NPLR records the highest nighttime values, likely due to nearby forest and agricultural respiration, while SWMR also shows elevated levels, consistent with its proximity to Olympia Park and Westfriedhof, which are large vegetated areas. TUMR, in contrast, has the lowest nighttime concentrations, likely due to its location in the urban core with less vegetation.

In fall and winter, the diurnal cycle flattens and overall concentrations rise. Reduced biogenic uptake, increased heating emissions, and shallower mixing layers lead to persistently higher CO2. The contrast between urban sites in Munich and the regional background (HPB) is most pronounced in winter. For the urban sites, SWMR exceeds TUMR and NPLR in afternoon CO2, with a distinct late-day peak likely driven by the higher traffic volume in close proximity. GROR exhibits the lowest daytime concentrations across all seasons, reflecting its upwind position and correspondingly reduced exposure to urban emissions from Munich.

To better understand seasonal patterns at all sites, we calculated seasonal diurnal variation. The diurnal variation is the difference between the maximum and minimum diurnal CO2 concentrations. Figure  displays the variation of diurnal CO2 variation for each site and season. The stations are sorted by their summer variation from low to high.

In summer, the largest diurnal variation is observed in rural sites. MAIR and TAUR, both located on the fringe of the city and adjacent to open land, show the strongest variation. In comparison, FELR and FINR are located atop municipal buildings in rural town centers and exhibit lower amplitudes. Suburban stations generally show greater summer variation than central urban sites, consistent with a stronger local biospheric influence. Exceptions are SWMR and SENR, which both show a high summer variation despite their urban classification. This is likely due to their proximity to large parks and green areas, where vegetation can drive both daytime drawdown and nighttime respiration. The lowest summer variation is found at the urban core stations TUMR and BLUT, where limited vegetation leads to a flatter diurnal profile.

In winter, overall diurnal variations are reduced, especially at city edge (e.g., HARR, GROR, NPLR, BOGR) and rural sites, reflecting the absence of daytime photosynthetic uptake. An exception is MAIR, which exhibits relatively high winter variation. This is likely due to intermittent influence from a nearby gas heating exhaust located 5 m east of the sensor inlet, contributing to elevated nighttime or early-morning CO2 peaks. This suggests that Hampel filtering alone may be insufficient to fully remove localized contamination effects from the data.

To further assess seasonal differences and spatial variability, we analyzed afternoon (10:00–16:00 LT) CO2 concentrations across the network. This period typically reflects the most well-mixed atmospheric conditions and removes the influence of nocturnal accumulation. Figure  shows the seasonal mean for each site in the afternoon, with vertical bars representing the quantiles 2.5 % and 97.5 %. The sites are sorted by winter averages from low to high. The summer and spring averages are consistently lower than in the fall and winter. This seasonal progression is in line with the typical CO2 cycle in urban environments, where biospheric uptake dominates during the growing season, while anthropogenic emissions and limited mixing contribute to winter accumulation.

In winter, rural stations (FINR, FELR, TAUR, MAIR) show lower average concentrations compared to urban and suburban sites. An exception is FELR, which exhibits elevated winter CO2 likely due to its location downwind of the city under prevailing wind conditions. Importantly, rural stations still separate clearly from urban sites despite their lower inlet heights (see Table ), demonstrating the robustness of the spatial signal. Urban sites show a wider range and particularly high upper quantile values.

In the fall, some sites stand out with unexpectedly high values. PASR and RDIR may be influenced by building ventilation and HARR by ongoing construction activity. We moved the inlet of the RDIR site on late October 2024 to be upwind of the building ventilation outlet, which should reduce local contamination. Optimizing sensor placement is an ongoing process, and further site-specific analysis will enhance network performance.

Taken together, the diurnal variations and afternoon averages reflect the network's ability to resolve spatial differences in CO2 concentrations throughout the city, especially for rural to urban gradients. The seasonal patterns are consistent with expectations based on urban biospheric activity and anthropogenic emissions, while the spatial differences highlight the influence of local sources and sinks. The findings in Munich are consistent with patterns reported in other urban mid-latitude networks .