Add Thailand earthquake dataset (earthquake_events, earthquake_reports_raw)

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+earthquake_reports_raw.csv filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,114 @@
+---
+license: cc-by-4.0
+language:
+- th
+- en
+tags:
+- earthquake
+- seismology
+- thailand
+- geospatial
+pretty_name: Thailand Earthquake Dataset
+size_categories:
+- 1K<n<10K
+configs:
+- config_name: earthquake_events
+  data_files:
+  - split: train
+    path: earthquake_events.csv
+- config_name: earthquake_reports_raw
+  data_files:
+  - split: train
+    path: earthquake_reports_raw.csv
+---
+
+# Thailand Earthquake Dataset
+
+Seismic event data from the [Thai Meteorological Department (TMD)](https://data.tmd.go.th/)
+`DailySeismicEvent` API, collected by an Airflow pipeline (the earthquake side
+of the same `Air_Pipeline_2` project as
+[Datchthana/WeatherDatasetThailand](https://huggingface.co/datasets/Datchthana/WeatherDatasetThailand))
+that polls roughly hourly.
+
+This is a separate repo from the weather dataset on purpose — it only covers
+earthquakes, no air-quality/weather data.
+
+## Files
+
+| File | Rows (as of 2026-08-08) | Description |
+|---|---|---|
+| `earthquake_events.csv` | 923 | Cleaned, deduplicated event list — one row per unique quake, deduped on `(datetime_utc, lat, lon)`. **Recommended table for most uses.** |
+| `earthquake_reports_raw.csv` | 525 | Full raw archive: one row per API poll, each containing the complete JSON `payload` TMD returned at that moment (`DailyEarthquakes` list + response header). Kept for audit/reproducibility. |
+
+`earthquake_reports_raw.csv` is much bigger (~180 MB vs ~0.2 MB) than its row
+count suggests: TMD's API returns a rolling window of recent events on every
+call, so consecutive polls re-embed mostly the same events verbatim inside
+the JSON payload. `earthquake_events.csv` is the deduplicated result of
+parsing all of those payloads over time — use that unless you specifically
+need the raw API responses.
+
+## Schema — `earthquake_events.csv`
+
+- `datetime_utc` / `datetime_thai` — event time in UTC and Thailand local time
+- `magnitude` — Richter magnitude
+- `depth_km` — focal depth in km
+- `lat` / `lon` — epicenter coordinates
+- `title_th` — original Thai-language title string from TMD
+- `tambon_th` / `amphoe_th` / `province_th` — Thai administrative location
+  parsed out of `title_th` (null when the quake is outside Thailand)
+- `location_en` — English location name
+- `is_domestic` — `true` if the event was located within a Thai province
+- `created_at` — when the row was first written to the database
+
+## Schema — `earthquake_reports_raw.csv`
+
+- `fetched_at` — when this API poll happened
+- `payload` — the full TMD API JSON response as a string (header + list of
+  events as returned, before parsing/cleaning)
+
+## ⚠️ Known data gaps
+
+Same pipeline host as the weather dataset, same downtime pattern — mostly
+Windows Update restarts. Checked against `fetched_at` in
+`earthquake_reports_raw.csv` (2026-07-11 to 2026-08-08, ~28 days): 14 gaps
+longer than 1.5h, totaling roughly **6 days** of missed polls. The largest
+gaps line up almost exactly with the ones documented in
+[WeatherDatasetThailand](https://huggingface.co/datasets/Datchthana/WeatherDatasetThailand)
+(same host, same outage):
+
+| Gap start | Gap end | Duration | Likely cause |
+|---|---|---|---|
+| 2026-07-15 23:00 | 2026-07-19 11:13 | ~3d 12h | Windows Update restart — pipeline wasn't manually restarted afterward |
+| 2026-07-27 10:00 | 2026-07-29 14:00 | ~2d 4h | Windows Update restart |
+| 2026-08-06 00:00 | 2026-08-06 04:24 | ~4.5h | Windows Update restart |
+| 2026-08-06 14:01 | 2026-08-06 18:00 | ~4h | Windows Update restart |
+
+Plus several smaller ~2h gaps (single missed hourly poll) scattered through
+late July/early August, also mostly Windows Update restarts.
+
+Unlike the weather pipeline, a short gap here usually **doesn't** lose data —
+TMD's rolling-window report tends to still contain an event on the next
+successful poll. But the 3.5-day gap is long enough that any event whose
+window had already rolled off the API response by the time the pipeline
+came back up is permanently missing from this dataset.
+
+## Source & license
+
+Data is sourced from the public TMD `DailySeismicEvent` API. TMD's own data
+catalog ([catalog.tmd.go.th](https://catalog.tmd.go.th/)) lists an "Open Data
+Common" license on its datasets, which is why this repo defaults to
+**CC BY 4.0** (attribution) rather than a software license like Apache 2.0 —
+but I could not pull TMD's exact terms text to confirm the precise clauses,
+so verify against their published terms before relying on this for anything
+commercial.
+
+One more caveat worth knowing: the ingestion code calls the API with the
+literal demo credentials (`uid=demo`, `ukey=demokey`) that ship in TMD's own
+docs, not a registered production key. It has worked so far, but there's no
+confirmation these credentials are meant for sustained/redistributed use —
+treat the ongoing availability of this data source as unverified.
+
+## Update cadence
+
+The upstream pipeline polls roughly hourly; this snapshot reflects a manual
+export and is not kept in continuous sync.

earthquake_reports_raw.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac3fcedc2328c7efe1399882e1e9443398dc1d2707b0045d2e0a79e0f810514b
+size 188066363