Summary Description

Fourteen-month deployment of four Sensirion SCD41 low-cost CO₂ sensors at the SMEAR Estonia station, Järvselja, Estonia. Three sensors (A, B, C) operated at 30 m on the atmospheric sensing mast; the fourth (Sensor D; SCT1) at 2 m. One LGR reference instrument provided ground-truth measurements at 30 m. Sensors operated under adverse environmental conditions outside manufacturer specifications. The dataset includes 10-minute aggregated readings, per-sensor with merged outputs, quality-control flags, and validation receipts.

Intended as an adverse-conditions benchmark for evaluating robust calibration methods for low-cost environmental sensors.

Depicted preview Taylor diagram summarising SCD41 sensor performance relative to the LGR reference. Radial distance encodes standard deviation (ppm); angular position encodes Pearson correlation; dashed arcs show centred RMSD. Arrows indicate sensors whose standard deviation exceeds the plot radius.

Data Dictionary

Timestamps

All timestamps are in Coordinated Universal Time (UTC), formatted as ISO 8601 strings in CSV files (YYYY-MM-DD HH:MM:SS) and as datetime64[ns] in Parquet files. Each timestamp marks the start of a non-overlapping 10-minute aggregation window aligned to the Unix epoch.

Missing Data

Missing values appear as empty cells in CSV and as NaN in Parquet. No interpolation or gap-filling is applied. The merged file uses an outer join: if a sensor has no data at a given timestamp, all its columns are NaN for that row.

Per-Sensor Files: SCD41 (Sensors A, B, C, D)

File pattern: {SENSOR}_10min.csv, {SENSOR}_{YYYY}_{MM}.csv

Sensors: CO2_SCT1_2M (Sensor D, 2 m), CO2_30M_A, CO2_30M_B, CO2_30M_C (30 m)

13 columns per file. At the SCD41 native sampling rate of 0.2 Hz, a full 10-minute window contains up to 120 deduplicated samples.

Per-Sensor Files: LGR Reference

File pattern: LGR_30M_10min.csv, LGR_30M_{YYYY}_{MM}.csv

Each measured variable produces four columns following the pattern {variable}_mean, {variable}_std, {variable}_count, and (where applicable) {variable}_sd for pooled standard deviation of the instrument-reported uncertainty.

At the configured logging interval of approximately 120 s, each 10-minute window contains at most 5 LGR samples.

Merged File

File pattern: SMEAR_EE_CO2_merged.csv, SMEAR_EE_CO2_{YYYY}_{MM}_merged.csv

100 columns. All sensors are aligned to a common 10-minute timestamp index via outer join.

Column Naming Convention

Merged columns are prefixed with the lowercase sensor name:

{sensor}_{variable}_{statistic}

Examples:

• co2_sct1_2m_co2_mean → Sensor D (2 m), CO₂, mean
• co2_30m_a_temp_std → Sensor A (30 m), temperature, standard deviation
• lgr_30m_CO2_dry_mean → LGR reference, dry CO₂, mean

SCD41 Sensor Columns (×4 sensors, 12 columns each)

Each SCD41 sensor contributes 12 columns: co2_mean, co2_std, co2_count, temp_mean, temp_std, temp_count, humidity_mean, humidity_std, humidity_count, co2_qc, temp_qc, humidity_qc — all prefixed with the sensor name.

LGR Columns (47 columns)

The LGR contributes 48 columns. Ten measurement variables (CO₂, CH₄, H₂O, pressure, temperatures, ring-down times) each produce four columns (mean, std, count, and pooled SD of the instrument-reported uncertainty); quality and Fit_Flag produce three columns each (mean, std, count — no instrument-reported SD exists for these). Two QC flag columns (co2_qc, quality_qc) complete the set. Pooled SD columns carry the suffix _sd__pooled_sd (double underscore).

Cluster Statistics (3 columns)

Computed from the three co-located 30 m SCD41 sensors (A, B, C):

Cluster statistics use only sensors with valid data at each timestamp. If one sensor reports, std is NaN.

Quality-Control Flags

We recommend filtering on qc_flag = 'OK' for initial analyses.

In this deployment, only OK, BELOW_RANGE, ABOVE_RANGE, and NO_DATA appear in the output. The remaining flags (SPIKE, LOW_QUALITY, BAD_FIT) are defined by the pipeline but were not triggered by any record in the current dataset.

File Formats

CSV and Parquet files for the same time period contain identical data. Parquet files are typically 5–10× smaller.

File Naming