AIR QUALITY · DATASET
Where the hourly gas-sensor readings come from, what the file contains, and how missing readings decide which 357 days enter the analysis.
The UCI record describes hourly measurements at a roadside site in an Italian city from five metal-oxide gas sensors, whose electrical response changes with the surrounding air, and from reference analysers, instruments that report gas concentrations for comparison.
Attribution: Saverio De Vito. Dataset DOI: 10.24432/C59K5F.
2004-03-10 18:00 → 2005-04-04 14:00
9,357 usable hourly records, 15 columns and no duplicate timestamps. One raw row is an hourly observation; one analysis row is a complete day for a single sensor.
The stored hourly clock has no missing hours.
Path of the data: UCI Air Quality · dataset 360 → supplied AirQuality package → original CSV and saved analysis → these pages.
The source report cites De Vito et al., Sensors and Actuators B 129(2), 2008, 750–757, and refers to a Kaggle redistribution of the data.
| Property | Source description / original report | What this file or record supports |
|---|---|---|
| Instances | Original card comparison: 9,357; UCI record: 9,358 | 9,357 timestamped rows after removing 114 blank rows. |
| File format | CSV | Semicolon separators; decimal commas; 2 empty unnamed columns. |
| Attributes | 15 columns / features | Date, Time and 13 numeric fields after cleanup. |
| Recording period | March 2004–February 2005 | 10 March 2004 18:00 to 4 April 2005 14:00; approximately 12.8 months. |
| Sampling | Hourly averages | 9,357 expected hourly slots, 9,356 one-hour gaps, no duplicate timestamps. |
| Missing-value convention | −200 marks missing observations | The sentinel occurs in every numeric field; it must not be treated as a measurement. |
| NMHC coverage | Listed as a reference-labelled variable | 90.2% missing; 914 observed hours, all within the first March–May window. |
| Benzene reference | C6H6(GT), described as a concentration | Strongly coupled to S2 (r = 0.982). |
| Licence | Research-only wording appears in the description | The UCI record also displays a CC BY 4.0 licence label; both statements appear on the UCI page itself. |
| Absolute humidity | AH is listed alongside T and RH | r = 0.999908 against a Magnus transform of T and RH. |
| Drift and cross-sensitivity | Described as characteristics of the deployment | S4 PC1 / retained-day order r = −0.652. |
| Field | Recorded metadata |
|---|---|
| Creator | Saverio De Vito (ENEA) |
| Donated to UCI | 22 March 2016 |
| Dataset DOI | 10.24432/C59K5F |
| Licence label | CC BY 4.0; see the Licence row above |
| Instances / features | 9,358 / 15 as listed in the UCI record |
Local package imported 30 September 2026; UCI metadata checked 1 October 2026.
| Fields | Meaning | Units / format |
|---|---|---|
| Date / Time | Recorded date and hour | DD/MM/YYYY · HH.MM.SS |
| PT08.S1–S5 | Hourly responses of the five metal-oxide sensors | Raw sensor scale, not a gas concentration |
| CO(GT) | Reference CO concentration | mg/m³ |
| NMHC(GT) / C6H6(GT) | Reference-labelled hydrocarbons / benzene | µg/m³ |
| NOx(GT) | Reference total nitrogen oxides | ppb |
| NO2(GT) | Reference nitrogen dioxide | µg/m³ |
| T / RH | Temperature / relative humidity | °C / % |
| AH | Absolute humidity | Unit not recorded in this CSV |
Each sensor’s oxide and the gas it was designed to target (its nominal target): S1 tin oxide / CO; S2 titania / NMHC; S3 tungsten oxide / NOx; S4 tungsten oxide / NO₂; S5 indium oxide / O₃.
The file uses semicolons as separators and commas as decimal points, and marks a missing reading with −200. The preview below shows the first six rows after parsing.
| Date | Time | PT08.S1(CO) | PT08.S2(NMHC) | T | RH | CO(GT) |
|---|---|---|---|---|---|---|
| 10/03/2004 | 18.00.00 | 1360.0 | 1046.0 | 13.6 | 48.9 | 2.6 |
| 10/03/2004 | 19.00.00 | 1292.0 | 955.0 | 13.3 | 47.7 | 2.0 |
| 10/03/2004 | 20.00.00 | 1402.0 | 939.0 | 11.9 | 54.0 | 2.2 |
| 10/03/2004 | 21.00.00 | 1376.0 | 948.0 | 11.0 | 60.0 | 2.2 |
| 10/03/2004 | 22.00.00 | 1272.0 | 836.0 | 11.2 | 59.6 | 1.6 |
| 10/03/2004 | 23.00.00 | 1197.0 | 750.0 | 11.2 | 59.2 | 1.2 |
Hourly rows → replace −200 with missing → group by recorded date → keep 24 complete sensor hours → five aligned 357 × 24 matrices
The full matrix viewer is on the Workflow page.
| Field | Missing hourly readings | Missing share |
|---|---|---|
| CO(GT) | 1,683 | 18.0% |
| PT08.S1(CO) | 366 | 3.9% |
| NMHC(GT) | 8,443 | 90.2% |
| C6H6(GT) | 366 | 3.9% |
| PT08.S2(NMHC) | 366 | 3.9% |
| NOx(GT) | 1,639 | 17.5% |
| PT08.S3(NOx) | 366 | 3.9% |
| NO2(GT) | 1,642 | 17.5% |
| PT08.S4(NO2) | 366 | 3.9% |
| PT08.S5(O3) | 366 | 3.9% |
| T | 366 | 3.9% |
| RH | 366 | 3.9% |
| AH | 366 | 3.9% |
The first and last calendar dates are partial days. All five sensors are missing together for 366 hours, in 16 episodes (unbroken stretches of missing hours); the longest lasts 76 hours. Missing sensor values are not filled in.
| Daily covariate | Available days | Entirely absent days |
|---|---|---|
| Temperature, RH and AH | 357 | 0 |
| CO reference | 321 | 36 |
| NOx reference | 323 | 34 |
| NO₂ reference | 323 | 34 |
| Date order and calendar labels | 357 | 0 |
Reference-gas daily means use the hours with readings on each day, so an available daily mean can rest on fewer than 24 readings.
All five sensors are absent together for 366 hours. CO, NOx and NO₂ are all absent together for 1,226 hours; only 31 hours overlap. Here “reference absence” means all three fields are missing at once.
Complete sensor days can still contain missing reference readings.
The 366 missing hours form 16 episodes, with median length 9 hours and maximum 76 hours. Five episodes exceed 24 hours.
| Episode starts (stored clock) | Missing hours |
|---|---|
| 2004-06-19 14:00:00 | 38 |
| 2004-08-26 06:00:00 | 45 |
| 2004-12-14 17:00:00 | 75 |
| 2005-01-02 21:00:00 | 52 |
| 2005-02-08 17:00:00 | 76 |
The analysis excludes the 32 otherwise full dates that contain these gaps and does not interpolate sensor readings.
All 914 nonmissing NMHC observations lie between 10 March 2004 18:00 and 1 May 2004 00:00. They are not 914 consecutive hours: this window itself contains missing readings.
A model that needs NMHC readings could use only early-2004 data.
The supplied check compares AH with water-vapour partial pressure e calculated from T and RH using the Magnus formula, over jointly observed hourly rows.
T, RH and AH contain strongly overlapping information.
1,078 of 1,245 distinct observed S2 values correspond to exactly one benzene value. The largest within-S2 spread is 0.1 µg/m³, one reported rounding step; group-mean benzene never decreases by more than that step as S2 increases.
Knowing S2 almost fixes the benzene value.
The three checked dates, 28 March 2004, 31 October 2004 and 27 March 2005, each contain 24 hourly rows. There are no duplicate timestamps or missing hourly slots anywhere in the cleaned file.
The stored clock therefore gives a consistent hourly index for the daily tables.
| Observed pattern in supplied analysis | Why it matters |
|---|---|
| Benzene-labelled C6H6(GT) is strongly coupled to PT08.S2 | Knowing S2 almost fixes the benzene value |
| AH closely tracks a transform of T and RH (r ≈ 0.9999) | The three humidity and temperature covariates carry overlapping information |
| Complete days are separated by excluded dates | Neighbouring rows of the daily tables are not always consecutive days |
| Reference gas means omit missing hours | A reference daily mean can rest on fewer than 24 hourly readings |
Retained file: AirQuality.csv · 785,065 bytes.
SHA-256
13277ae5d8581e80b7be09d47c7d3d06fe9b8e957078f2cf6e859f955e62f996
Source record: UCI Air Quality · dataset 360. Citation: Vito, S. (2008). Air Quality [Dataset]. UCI Machine Learning Repository. 10.24432/C59K5F.
Licence: the UCI page shows a CC BY 4.0 licence label, and its description also contains research-only wording.
The original data file is kept together with the supplied analysis script, saved results, page template and two HTML pages.