Datchthana/WeatherDatasetThailand

Thailand Air Quality & Weather Dataset

Dataset Description

Hourly air-quality readings from AIR4Thai (Thailand Pollution Control Department) enriched with matching weather/air-pollution data from the OpenWeather API, collected by an Airflow pipeline (Air_Pipeline_2) that ingests every hour and builds a small star-schema mart on top:

• AIR4Thai (getNewAQI_JSON.php) — the PCD's public real-time station network: overall AQI, PM2.5, PM10, O3, CO, NO2, SO2, per-pollutant sub-AQI, station metadata/coordinates.
• OpenWeather Air Pollution + Weather APIs — pollutant concentrations plus temperature, humidity, pressure, wind, cloud cover and weather description, fetched per station's lat/lon as a cross-check alongside the ground-truth AIR4Thai reading.

Update cadence: AIR4Thai only ever serves the current reading per station (no history API), so ingestion runs hourly (45 * * * * Asia/Bangkok) and this dataset is the historical archive — there is no way to backfill air-quality history from the source beyond what this pipeline has already captured.

Geographic scope: Thailand — all stations reported by AIR4Thai (up to 180 station IDs observed).

Not yet measured: total distinct provinces/regions covered, exact AIR4Thai polling uptime/SLA — not independently verified, don't assume.

Dataset Structure

Four configs, matching the pipeline's raw and mart layers:

air_stations (raw, one row per station-hour as ingested)

40 columns: station identity/location (station_id, area_th, area_en, station_type, lat, lon), AIR4Thai readings (aqi, aqi_param, pm25_value, pm25_aqi, pm10_value, pm10_aqi, o3_value, o3_aqi, co_value, co_aqi, no2_value, no2_aqi, so2_value, so2_aqi), OpenWeather fields (ow_aqi, ow_co, ow_no, ow_no2, ow_o3, ow_so2, ow_pm25, ow_pm10, ow_nh3, ow_temp, ow_feels_like, ow_humidity, ow_pressure, ow_wind_speed, ow_wind_deg, ow_clouds, ow_weather), and recorded_at/created_at timestamps. 182.6K rows.

dim_station (mart dimension)

station_key (surrogate PK), station_id (natural key), area_th, area_en, location (derived), station_type, lat, lon, updated_at. 177 rows.

dim_date (mart dimension)

Standard date dimension: date_key (YYYYMMDD), full_date, year, quarter, month, month_name, day, day_of_week, day_name, week_of_year, is_weekend. 45 rows.

fact_air_quality (mart fact table)

36 columns: fact_key (PK), station_key/date_key (FKs), station_id, recorded_at, created_at, the same AIR4Thai + OpenWeather measurement columns as air_stations except ow_co, which is captured upstream but not currently carried into this table. Grain and upsert key: one row per (station_id, recorded_at). 141.3K rows.

AQI computation

ow_aqi (computed from the OpenWeather PM2.5 reading, for cross-checking against AIR4Thai's own aqi) uses linear interpolation over this PM2.5 breakpoint table, which matches Thailand's PCD PM2.5 AQI scale (not the differently-bucketed US EPA table):

PM2.5 (µg/m³) low–high	AQI low–high
0.0 – 15.0	0 – 25
15.1 – 25.0	26 – 50
25.1 – 37.5	51 – 100
37.6 – 75.0	101 – 200
75.1 – 150.0	201 – 300
150.1 – 250.0	301 – 400
250.1 – 500.0	401 – 500

PM2.5 above 500.0 clamps to AQI 500. All other aqi/*_aqi columns are AIR4Thai's own reported sub-indices, not recomputed by this pipeline.

Data Collection & Processing

Three-stage Airflow pipeline:

1. Ingest (hourly) — calls AIR4Thai for all stations, then OpenWeather per station by lat/lon, upserts into air_stations on (station_id, recorded_at). A malformed payload or flaky OpenWeather call on one station is skipped and logged, not allowed to fail the whole batch — but if every station fails in one run, the task fails loudly instead of silently publishing zero rows.
2. Per-station transform — casts raw text fields to numeric with range clamping (e.g. PM2.5/PM10/O3/CO/NO2 to 0–900, SO2 to 0–600, AQI sub-indices to 0–500); anything unparseable or out of range is nulled rather than propagated.
3. Mart build — unions per-station tables and upserts into dim_station/dim_date/ fact_air_quality.

Design intent: upstream data-quality problems (a bad sensor reading, a station momentarily offline, a renamed API field) degrade gracefully to nulls on that one row/column rather than blocking processing for every other station — so a non-null value means "parseable and in a plausible range," not "independently verified correct."

Known Limitations / Considerations

From a live verification pass against the source database (2026-08-18):

• Ingestion and per-station transform layers were current to within ~1.5 hours and free of nulls/duplicates at check time.
• The mart (dim_station/dim_date/fact_air_quality) can lag behind the raw layer when the mart-build stage doesn't run on schedule — at one point it was ~63 hours behind. If you need guaranteed-current data, check fact_air_quality's max created_at against air_stations' before relying on it as "now."
• At least one row (station_id = 33t) has an AQI value of 501, outside the valid 0–500 range, predating a range-clamp fix — not yet backfilled at the time of the check. Exclude or clip outlier AQI values above 500 before computing aggregates.
• A small number of stations (observed: 32t, o24, o28, o29, o38, o67, o72) can go dark at the AIR4Thai source for extended periods — this is an upstream gap, not a pipeline defect, but it reduces effective spatial coverage below the nominal station count; check for gaps per station_id before assuming uniform coverage.
• ow_co is captured during ingestion but not currently included in fact_air_quality — use air_stations if you need OpenWeather CO readings.

Licensing / Attribution

Released under Apache 2.0. Underlying source data originates from:

• AIR4Thai (Thailand Pollution Control Department) — a public government API.
• OpenWeather — a commercial API; if redistributing raw ow_* values at scale, check OpenWeather's current terms of service.