The DG2MCM-15 ground station is located in Kempten, Bavaria, Germany (47.7245° N, 10.3279° E) at 686 m above sea level. The site is situated in a semi-urban environment at the northern edge of the Alps, characterized by moderate aerosol loading and frequent cloud cover typical of Central European mid-latitude locations.

The sensor array is mounted at 4.7 m above ground level in a radiation shield providing ventilation while preventing direct precipitation contact and solar heating artifacts. Sky-viewing sensors are oriented at 50° elevation angle toward the north to avoid direct solar contamination while maintaining atmospheric path sensitivity.

Sentinel-5P TROPOMI Level-2 products were obtained from the Copernicus Data Space Ecosystem (CDSE) for the period 11 January–8 February 2026, using both Near Real-Time (NRTI) and Offline (OFFL) processing streams depending on availability. The following products were downloaded: the aerosol index product (L2__AER_AI, processor version 02.09.01) providing UV aerosol index at 340/380 and 354/388 nm wavelength pairs; the nitrogen dioxide product (L2__NO2___, processor version 02.09.01) providing tropospheric vertical column densities; the ozone product (L2__O3____, processor version 02.08.00) providing total vertical column amounts; and the cloud product (L2__CLOUD_, processor version 02.08.00) providing cloud fraction, cloud top pressure, and cloud optical thickness derived using the Optical Cloud Recognition Algorithm/Retrieval of Cloud Information using Neural Networks (OCRA/ROCINN) algorithm . Where both NRTI and OFFL products were available for the same overpass, the OFFL product was preferred due to its more complete auxiliary data and refined processing. All products have native spatial resolution of approximately 5.5 km (along track) ×3.5 km (across track) at nadir, degrading to 5.5 km ×7 km at swath edges.

For each satellite overpass, the nearest pixel to the DG2MCM-15 station coordinates was identified using great-circle distance calculation. The quality assurance (QA) value from each product was used to filter retrievals: cloud products required QA ≥0.5 following the S5P Product Readme File for the cloud product (PRF-CL) and the recommendation of , trace gas products required QA ≥0.75 for high-quality retrievals, and aerosol products retained all QA values due to limited data availability. Pixel centre coordinates, observation time (UTC), and all geophysical parameters were extracted and stored with satellite-ground distance and spatial representativeness metadata.

Ground measurements were matched to satellite overpasses using a ±30 min time window. For each satellite observation time tsat, the nearest ground measurement tground was selected such that: 2Δt=|tsat-tground|<30min The 30 min window accounts for atmospheric temporal variability in clouds and aerosols, the ground station measurement frequency of 1 min, and variations in satellite overpass geometry.

The horizontal distance between ground station and satellite pixel centre was calculated for all matched pairs. Given the TROPOMI pixel footprint (3.5–7 km depending on viewing geometry), acceptable matches were defined as: 3dhorizontal<3.5km This threshold ensures the ground measurement falls within the nominal satellite observation area at nadir geometry.

Validation metrics employed in this study include the Pearson correlation coefficient (R), root mean square error (RMSE), mean bias (defined as satellite minus ground measurement), and standard deviation of differences. Sample sizes (N) are reported for all comparisons.

Over the four-week study period, 348 Sentinel-5P overpasses provided coverage over the Kempten region, distributed across four TROPOMI products (Fig. a). The ground station DG2MCM-15 acquired 3574 measurements during this period, providing ample temporal coverage for satellite-ground collocation.

Temporal matching with the ±30 min window yielded 276 satellite-ground observation pairs (79 % match rate). The mean temporal difference was 2.7 min with a standard deviation of 2.1 min (Fig. b). Ninety-five percent of matches occurred within 8 min of the satellite overpass, demonstrating excellent temporal collocation for atmospheric validation purposes.

Spatial analysis shows all matched observations fall within 3.5 km of the satellite pixel centre, with 85 % of cloud product matches within 2 km (Fig. c). The small spatial offsets confirm adequate representativeness for ground-based validation of TROPOMI observations.

The matched observation dataset includes TROPOMI nitrogen dioxide (N=63) and ozone column (N=71) retrievals (Table ), co-located primarily to characterise cloud contamination effects on trace gas retrievals. Direct ground-based validation of column trace gas products is beyond the scope of the current study and would require dedicated instrumentation such as Pandora or Brewer spectrometers.

To further assess the effect of sample size, we extended the observation period from 28 d (Dataset A) to 96 d (Dataset B, 21 609 DG15 records). Applying the original QA ≥ 0.5 threshold yields N=27 matched pairs in both datasets – identical despite a 6× increase in ground data. This reveals that the S5P QA filter is the primary limiting factor in the Pre-Alpine region, rejecting the majority of overpasses due to broken cloud fields and reduced retrieval confidence in complex terrain. Relaxing the QA threshold to 0.35 increases the sample to N=31–32 (Fig. ). Results for both thresholds are summarised in Table . The corresponding confusion matrices for all four scenarios are presented in Fig. .

Table  summarizes the matched observation pairs for each TROPOMI product. Cloud and ozone products achieved the highest number of matched pairs (71 each), while nitrogen dioxide showed the highest match rate (82 %) due to a smaller total number of satellite observations with more successful QA filtering (63 matched pairs from 77 total).

The temporal distribution (Fig. d) shows regular coverage throughout the study period with near-daily observations for most products. Gaps in coverage correspond to satellite orbit geometry (no overpass) or failed QA criteria for specific products.

The dual AS7341 sensor configuration (2.7 m baseline) enables instrumental uncertainty quantification through direct sensor intercomparison. For clear-sky conditions (N=45 matched pairs), the standard deviation of inter-sensor channel differences ranges from 3.2 % (red channels) to 8.7 % (violet channels), with higher relative uncertainty at shorter wavelengths due to reduced atmospheric transmission and increased Rayleigh scattering sensitivity.

This sensor reproducibility provides a lower bound for measurement uncertainty in future aerosol optical depth validation studies, where retrieval uncertainties typically exceed 15 %–20 % for ground-based sun photometry.

Concurrent measurements of aerosol optical depth proxy and trace gas co-locations (NO₂, O₃) are available from the DG2MCM-15 sensor suite but are not further analysed in this study. The TROPOMI aerosol index is derived from UV channel pairs at 340/380 and 354/388 nm, outside the spectral range of the current AS7341 sensor (415–680 nm), precluding a physically motivated comparison. Preliminary results including detection of a Saharan dust intrusion event (AOD proxy +10.5σ above baseline) demonstrate the network's sensitivity to episodic aerosol events and will be the subject of a dedicated follow-on study. Integration of the AS7331 UV sensor (280–390 nm, operational since March 2026) will enable direct spectral overlap with TROPOMI UVAI retrieval wavelengths in future work.

The reported correlation (R=0.879) between ground-based infrared sky temperature and Sentinel-5P cloud fraction indicates the potential of low-cost thermal infrared sensors for satellite cloud product comparison. However, as noted in Sect. 3.3.1, this correlation is driven substantially by two high-cloud-fraction observations and should be interpreted with caution. While the correlation coefficient is comparable to or exceeds those reported for dedicated cloud validation instruments such as ceilometers (R=0.85–0.92) and whole-sky imagers (R=0.78–0.88) , the higher RMSE (29.1 %) and positive bias (+20.8%) indicate that the method is better suited for cloud screening (distinguishing clear, cloudy, and overcast conditions) than for quantitative cloud fraction retrieval.

The discrete nature of the ground-based cloud fraction values – constrained to exactly four levels (0 %, 25 %, 50 %, 75 %) by firmware-level integer quantisation in the RAK4631 microcontroller (see Sect. 2.1.2) – means the correlation coefficient should be interpreted with caution. The R=0.879 is driven substantially by the separation between these four discrete levels rather than by a continuous linear relationship, and the effective degrees of freedom are lower than the sample size (N=27) would suggest. This hardware-imposed quantisation fundamentally limits the utility of the current sensor generation for point-to-point quantitative cloud fraction validation.

The residual scatter and bias can be attributed to three principal factors. The most significant is the spatial scale mismatch between instruments: the ground-based infrared pyrometer's effective FOV (estimated 1–2 km at cloud base height) is substantially smaller than the TROPOMI pixel (3.5–7 km, corresponding to 12–39 km²). The satellite pixel consistently includes more cloud elements than the ground sensor observes, explaining the positive bias. Additionally, cloud type sensitivity introduces systematic differences, as thin cirrus clouds and high-altitude ice clouds emit less thermal radiation than low-altitude liquid water clouds despite similar optical thickness. Finally, even with a mean temporal offset of only 2.7 min, rapidly evolving convective clouds can produce significant differences between satellite and ground observations.

Traditional satellite validation networks such as AERONET and NDACC provide high-accuracy reference measurements but are constrained by limited spatial coverage with approximately 500 sites globally, and the Pandonia Global Network for UV–VIS trace gas columns . The DG2MCM approach demonstrates that citizen science networks using calibration-informed protocols can complement these established networks, particularly in data-sparse regions. We emphasise that this study does not constitute a validation of the Sentinel-5P cloud fraction product. Rather, it demonstrates the feasibility of the COTS-based estimation method by showing statistical agreement with an established reference dataset. Future work involving comparison with independent ground truth sources (e.g. AERONET, CLOUDNET, visual observer reports) is needed to fully establish the method's validity as a ground reference.

The low-cost sensor approach offers several practical advantages over reference instrumentation. Station hardware costs remain below USD 500 compared to USD 50 000–100 000 for reference sun photometers, enabling deployment at scales that would be prohibitively expensive with professional instruments. The stations operate autonomously with wireless telemetry, making them suitable for remote locations. Each station provides simultaneous cloud, aerosol proxy, and photometric observations, and the 1 min sampling interval exceeds the 15 min cadence typical of reference networks. The primary trade-off is absolute accuracy (15 %–20 % for low-cost sensors versus 1 %–2 % for reference instruments), which remains acceptable for satellite product validation where retrieval uncertainties typically exceed sensor limitations.

Throughout this paper, the term calibration-informed refers to measurement protocols that apply quality assurance principles drawn from formal spectroradiometric metrology practice – including systematic uncertainty analysis, dark current monitoring, and cross-sensor consistency checks – without claiming equivalence to formally accredited calibration in the sense of ISO/IEC 17025. A distinctive aspect of this study is the application of calibration-informed measurement principles to citizen science atmospheric monitoring. The methodology incorporates quality assurance procedures and uncertainty awareness drawn from spectroradiometric measurement practice, including systematic consideration of sensor drift, dark current, and cross-sensor consistency.

However, it is essential to distinguish between calibration-informed methodology and formally calibrated measurements. The sensors used in this study rely on factory calibration and cross-sensor consistency checks rather than traceable calibration against reference standards. The outdoor deployment conditions (temperature range -20 to +35 °C, varying humidity, solar exposure) fundamentally differ from the controlled laboratory environment required for formal calibration.

The measurement system design reflects systematic consideration of error sources, signal saturation avoidance, and automatic gain control – principles relevant to any rigorous low-cost sensor deployment. Regular monitoring of sensor stability, dark current behaviour, and cross-sensor agreement between co-located stations DG2MCM-15 and DG2MCM-16 provides confidence in data consistency. The distinction between sensor precision (reproducibility) and accuracy (closeness to true value) is maintained throughout, enabling honest uncertainty assessment.

Future work should pursue formal calibration of low-cost sensors using custom adapters for integrating sphere and blackbody reference sources. This would require addressing the practical challenge of the small physical dimensions of commercial sensor packages (e.g., AS7341: 3.1 × 2.0 mm) and the development of standardized calibration protocols suitable for citizen science networks. Organizations with existing spectroradiometric infrastructure – testing laboratories, universities, and national metrology institutes – are well-positioned to develop and disseminate such protocols.

The comparison approach presented here is directly applicable to current and future satellite missions. The upcoming Sentinel-5 operational mission (expected 2025+) will carry a similar instrument to TROPOMI, ensuring continuity of the validation methodology. Geostationary missions including the Tropospheric Emissions: Monitoring of Pollution (TEMPO) instrument over North America and the Geostationary Environment Monitoring Spectrometer (GEMS) over Asia require continuous ground-based validation that autonomous low-cost stations can provide. The recently launched Sentinel-4 mission provides continuous daytime monitoring of atmospheric composition over Europe from geostationary orbit, offering particularly high temporal resolution for ground-based validation; the DG2MCM-15 Central European site is well positioned to contribute to this effort. The next-generation Meteorological Operational Satellite – Second Generation (MetOp-SG) polar-orbiting platform will likewise benefit from dense ground-validation networks. The multi-spectral ground station design accommodates varying satellite wavelength ranges (UV to near-IR) and temporal sampling requirements for both polar-orbiting and geostationary platforms.

Several limitations of the current study should be addressed in future work. The quasi-discrete clustering of infrared-derived cloud fraction values demonstrates that the temperature-ratio method (Eq. ) is insufficient for continuous quantitative cloud fraction retrieval, particularly during winter conditions when the large clear-sky temperature depression compresses derived values. Alternative approaches, including multi-threshold classification methods, machine learning-based cloud detection using the full multi-spectral dataset, or narrower FOV infrared sensors that better resolve individual cloud elements, may improve cloud fraction resolution. The positive bias (+20.8%) points to systematic spatial mismatch effects that could be partially addressed by comparison with higher-resolution satellite cloud products from MODIS, VIIRS, and SLSTR (1 km or better).

The current sensor configuration lacks UV spectral coverage, preventing meaningful validation of TROPOMI aerosol index products. Integration of the AS7331 UV sensor covering 280–390 nm will enable spectral overlap with TROPOMI UVAI retrieval wavelengths. Similarly, column trace gas validation (NO₂, O₃) requires dedicated instrumentation such as Pandora spectrometers that is beyond the scope of the current low-cost station design, though the multi-spectral data may support indirect proxy approaches in future studies.

Regarding spatial representativeness, the current micro-array configuration with 2.7 m sensor separation provides instrumental uncertainty quantification but limited spatial sampling relative to satellite pixel dimensions. Future networks should incorporate multiple ground stations separated by 3–5 km within a single satellite pixel footprint, supplemented by mobile validation campaigns and integration with broader citizen science networks. The four-week observation period of this study, while sufficient for demonstrating the comparison methodology, is insufficient for characterizing seasonal variability and long-term sensor stability. Multi-year datasets, systematic sensor aging and drift characterization, and regular recalibration protocols will be essential for establishing the long-term reliability of low-cost sensor comparison. Extended observation periods should also capture episodic aerosol events (Saharan dust, biomass burning) to test aerosol detection capabilities once UV sensors are deployed.

An important finding from the extended dataset analysis (Dataset B: 96 d, 21 609 ground records) is that extending the observation period six-fold does not increase the number of QA-filtered matchups with the strict threshold (QA ≥ 0.5): both the 28 and 96 d datasets yield N=27 matched pairs. This reveals that the S5P QA filter is the primary limiting factor in the Pre-Alpine region, rejecting the majority of overpasses due to broken cloud fields and reduced retrieval confidence in complex terrain. Relaxing the QA threshold to 0.35 (following RC2 recommendation) increases the sample to N=31–32, confirming this interpretation. The full QA threshold sensitivity analysis is presented in Fig. . We report results for both thresholds and note that relaxed-QA matchups include overpasses with lower retrieval confidence, which may introduce additional uncertainty in the TROPOMI cloud fraction values.

The comparison results presented here have implications for both Sentinel-5P data users and for the design of future low-cost validation networks. The strong ordinal agreement (R=0.879) supports the use of ground-based infrared thermometry for cloud screening – identifying cloud-contaminated satellite retrievals and distinguishing between predominantly clear and cloudy conditions. However, the quantitative limitations (RMSE =29.1%, quasi-discrete cloud fraction values) indicate that this approach should be considered a complement to, rather than a replacement for, dedicated cloud validation instruments. The Central European mid-latitude comparison site contributes to geographic diversity in validation coverage, and the high-temporal-resolution ground data (1 min sampling) enables detailed studies of cloud evolution within satellite overpass windows.