Bangkok, live, flood edition Bangkok

Who

Dr Non — Bangkok, 1981 →

That’s me. Days old. Siam Square — the most expensive ground in Bangkok today. Back then it was just where our house was. I grew up in the heart of this city and learned it from the inside: heat, floods, skytrain openings, AQI at 200, tax schedules frozen while the market doubled.

I am not going to let anyone fuck up this city. Not the current governors. Not the deputies. Not the committees writing ten-year “smart city” PDFs while the public gets nothing they can use. No apologies. I’m building something — a working twin of the city I was born into. My city. Not theirs.

Thesis

One laptop. Zero politicians who care.

The BKKx Atlas is a thesis statement: a working digital twin of a real city can be built in an afternoon, on one laptop, for the cost of a domain. The data, APIs, and maps are here. What’s missing is a city government with the will to use them.

BMA has been “planning a smart city” since 2016. Ten years. Ten digital strategies. Zero public APIs worth using, zero digital twin for residents. Buildings get taller. Floods get deeper. The public still gets a 1990s PDF. We shipped the whole stack on a one-laptop budget. The hard part is keeping every layer honest and stackable.

My city, not theirs.

How it’s built

Basic infrastructure — still one machine

• Front. Static MapLibre GL atlas in public/ — no SPA framework. Layers, presets, inspect, and this About drawer are plain HTML / CSS / JS.
• Edge. Cloudflare Worker (src/index.ts) on atlas.nonarkara.org: proxies live feeds, serves risk / civic APIs, Range-passthrough for PMTiles from R2.
• Tiles & store. 432,077 OSM buildings as PMTiles on R2; civic amenities in D1 (CIVIC_DB) ingested from BMA / data.go.th.
• Live inputs. iTIC/Longdo traffic + CCTV, RainViewer radar, Open-Meteo wind/heat/AQI, water gauges judged by FloodDash — honest about what is live vs schematic.
• Reuse. GeoJSON downloads and /api for Chula, UNL, or anyone building on Bangkok without a consulting fee.

What’s on the map

Shipped — stackable, honest

1. Multi-layer city stack. Buildings under translucent rain, wind, heat, AQI, and flood. Presets: City · Live pulse · Rain now · Flood watch · Satellite air.
2. Inspect + risk API. Click a building → OSM height/type + schematic flood × live PM2.5 via /api/risk.
3. Live traffic + CCTV. iTIC/Longdo expressways and public road cameras (HLS). Not the closed BMA security net. /api/cctv
4. Rail + water transit. Nine rail lines plus four dotted boat corridors and 16 passenger piers. Dots mean scheduled orientation, not pretend-live boats. Ask the multimodal API.
5. Civic open data. Amenities + density by district from BMA / data.go.th via D1 — /api/civic-pois.
6. Land price by district. Real Treasury road-level analysis as a choropleth — not a fake heat paint.
7. Waterways, gauges + rain radar. Every canal sized by reach and flowing downstream, coloured where a FloodDash-judged gauge is worried; animated RainViewer frames over the silhouette. Method on the Research tab.
8. Current wind field. A 6 × 6 Open-Meteo model grid turns 10 m speed and direction into clickable arrow shafts and heads. It explains where rain and haze may move without pretending to be a street-level sensor network.
9. Night lights / blackout. Desaturated satellite + emissive buildings — pure client paint.

Roadmap

Still one laptop

1. Official flood extents. Swap the schematic envelope for DDPM/BMA GeoJSON when it’s open.
2. Property Tax Verifier (ภาษีที่ดิน) v2. Deed / parcel lookup over Treasury zones vs a typed tax bill.
3. BTS/MRT isochrones. 5 / 10 / 15-minute walksheds from every rail station.
4. Risk API v3. Fold rain intensity + building height into the score.
5. Offline atlas pack. Pin MapLibre + pmtiles + samples onto a USB stick.

Horizon

6. BKKx Digital Twin — handover to BMA. Open-source, no lock-in, no ฿5.4M consulting fee.
7. Property-tech API (full). 432K buildings + cadastral polygons — institutional, not code, is the blocker.
8. Tourist AR · planning sandbox · command center. Same Worker, denser chrome — after the civic layers are boringly reliable.

Data & APIs

For Chula, UNL, anyone building on Bangkok

• Live API catalog JSON · traffic, cctv, risk, flood-sensors, civic-pois
• 16-area curated tour JSON · districts + city-scale “Bangkok from above”
• 126 BTS/MRT/ARL/SRT segments GeoJSON · 9 lines, official colors
• 4,063 canals & rivers GeoJSON · Chao Phraya + khlongs
• 9.3K Rattanakosin buildings GeoJSON · height + base_height
• 2,273 shophouse row candidates GeoJSON · Overture morphology screen · unverified candidates, not a heritage survey
• 769 renderable shophouse rows GeoJSON · the subset carrying a real Overture storey count
• 1.9K expressway traffic segments GeoJSON · iTIC-compatible sample
• 5 schematic flood envelopes GeoJSON · civic demo · NOT official

Full city extract: python3 scripts/osm-buildings-to-geojson.py with a Bangkok bbox or local PBF. Scripts are the canonical extractors.

Research

Papers that sit on the map

Working papers live under /labs — a research tab, not a press kit. Paper 01 is the system. Paper 02 is the shophouse bible: the ตึกแถว as the actual unit of the inner city. The shelf is the recovered thesis catalog that sits under it.

Contact

Use the data. Write him.

Building something on this atlas — research, civic tooling, a classroom demo — email Dr Non. Same for press, partnership, or calling bullshit on a layer that should be more honest.

• nonsmartcity@gmail.com Direct
• nonarkara.org Home
• GitHub · bkk-3d-atlas Source

Dao

Two editions of the Dao De Jing, kept side by side

The atlas is a Bangkok object. The Dao is the other side of the same author — a quiet one, written for the people who ask him for it in person. Eighty-one chapters, three languages (Wang Bi's Chinese, clear English, Thai), dedicated to the Upatising family and in memory of Mother Lilly Upatising (1948–2026).

Both editions are developed in parallel. D2 is the default for people who want something simple and easy to read; D1 is the classic archive for people who study it. The atlas links to both — pick the one that fits your day.

Credits

Sources that stay cited

Base map & buildings. © OpenStreetMap contributors (ODbL 1.0) via OpenFreeMap; BKKx PMTiles extract of Bangkok metro footprints. Imagery © Esri, Maxar, Earthstar Geographics.

Live & open civic. iTIC/Longdo (traffic, public CCTV), กรมทางหลวง · Department of Highways (traffic CCTV carried on the iTIC/Longdo platform), Maholan (cctv.maholan.net, community CCTV aggregator), RainViewer, Open-Meteo (10 m wind / heat / PM2.5), HII/ThaiWater (water levels), BMA / data.go.th / data.bangkok.go.th (schools, sports grounds, car parks, museums, waterway landmarks). Land-price choropleth from Treasury road-level analysis (Home Buyers).

Passenger boats. Maintained service references from Chao Phraya Express Boat, Chao Phraya Tourist Boat and the Thai Government Contact Center’s Khlong Saen Saep record. Corridors and piers are not live vessel positions.

Host. Cloudflare Workers + R2 + D1. Built by Dr Non · BKKx · live at atlas.nonarkara.org. Installable as a web app — Android (Chrome): menu → Add to Home screen; iPhone (Safari): Share → Add to Home Screen. Live layers need a connection.

The question

A diagram nobody could read

The BMA Drainage and Sewerage Department publishes a water-management schematic of the city — แผนผังบริหารจัดการน้ำ กรุงเทพมหานคร — at weather.bangkok.go.th/KlongMap. It is the picture that gets pointed at in flood briefings. It is also a diagram: canals as straight strokes, no ground under them, no scale, nothing that says which way the water drains or where it is high right now.

The atlas already had the city in 3D over satellite imagery and 4,063 waterway segments from OpenStreetMap. The question for the water layer was plain: can one map show where Bangkok's water runs, which way it drains, and where a gauge says it is high — without claiming more than the data knows?

FloodDash

What it is, what it is not

FloodDash is the atlas's twin for water: a real-time, openly sourced flood watch for every Thai province, run under the Smart City Thailand Office (depa). It streams about 17 public pipelines — HII, RID, PCD, NOAA, Copernicus, Open-Meteo, NASA GPM, GISTDA, TMD, iTIC, CheckFlood CCTV, news — judges every gauge into a band, and every province into safe, watch, prepare or danger. Watch-score weights: water 40 · rain 25 · forecast 15 · wetness 10 · rise 10.

Its own definition, word for word:

Useful as a 24-hour-ahead watch and a municipal action board. Not a flood clock.

We cannot say “this village floods at 14:32.” No DEM/HAND. No street inundation. No village ETA. Silence is not all-clear. No nationwide EVACUATE from a heuristic. Orders stay with ปภ. / อปท. 1784.

When the data is thin, the useful sentence is the refusal. Do not inflate the claim.

FloodDash · About · What this is, what it is not

The measured record

Three floods in four, about half a day ahead

Recall

74.2%

Precision

3.8%

Median lead

19.7 h

Recall, observed rain removed

73%

FloodDash's 40-day prediction ledger · 76 truth gauges · window ending 9 Sep 2026 · lead q25 12 h, q75 35.7 h.

89 genuine over-bank events, 89 warned ahead, 31 missed — 74.2% recall but only 3.8% precision, because 2,241 alerts did not flood. […] A watch board that catches three floods in four about half a day ahead — not an alarm that is almost always right.

Forecast rain is biased +5.3 mm (MAE 7.7 mm across 3,744 pairs) — rain forecast ranks risk, never triggers a warning. […] Silent or flatlined gauges are the real blind spots: silence is not an all-clear.

FloodDash · About · What the collected data says

The atlas inherits that character. A red canal means a gauge on it is worried — often rightly, sometimes not. It never means the street beside it is under water.

How the atlas reads it

A relay, not a second opinion

• Pull. The Worker's cron reads FloodDash's keyless /api/twin, /api/snapshot and /api/dds/briefing about every 10 minutes. Visitors never touch FloodDash; they read the copy in D1.
• Keep. Each of the three reads fails on its own and keeps its last good copy. A failed refresh never blanks the map, and the age of what you see is always printed.
• Pass through. Band names, actions and the Bangkok verdict are FloodDash's own words, in English and Thai. The atlas adds placement and age. It never adds a verdict.
• Why. On 25 Sep 2026, with Bangkok at danger, ThaiWater answered the atlas's shared Cloudflare egress with HTTP 429 and the old gauge layer went blank. FloodDash already ingests that feed once, politely, from one machine.
• The rule it keeps. A fresh FloodDash reading, or the last good one with its age. Never blank while a copy exists, never an unlabelled age, never a verdict FloodDash did not give. scripts/check-flooddash.ts asserts this against live FloodDash.
• Open. /api/flood/bkk · /api/flood-sensors · the live block in /api/risk.

Method

Width, motion, colour

1. Width is reach. OSM's width tag here is a type default, not a measurement. So every segment takes the rank of its named waterway by total length across the metro: the Chao Phraya (102 segments, 108 km); 20 km and longer (124 segments — Saen Saep 47 km, Prem Prachakon 40, Samrong 40); 7–20 km (387); other named khlongs (1,672); unnamed ditches (1,778, drawn only close in). 819 named waterways, about 4,000 km in all.
2. Motion is drawing direction. OSM draws waterways in the direction they flow, and the dashes slide that way. 95 of 102 Chao Phraya segments run southward end to end; the other seven sit on meanders. Tidal reaches and pumped canals can run backwards, so the dashes show nominal drainage, not a measured current. Still, for reduced motion.
3. Colour is the nearest judged gauge. Each gauge snaps to its own named waterway within 1 km, or to any waterway within 300 m. Segments of that waterway within 1.5 km, straight-line, take its band; the most severe gauge wins; judged stretches swell. Silent and untrusted gauges colour nothing. Blue means no gauge concern — or no gauge at all.
4. Flooded roads: every signal, on its street. Residents' Traffy Fondue tickets, drivers' iTIC reports and BMA DDS's flooded-road list are placed on the OpenStreetMap road network (267,856 roads). A report snaps to the nearest road within 90 m (60 m for iTIC) and lights about 100 m of it each way; reports on one road merge into a stretch. DDS gives only a road name and a khet, so the named road inside that khet is lit and marked as such. Gauges and geolocated news add to the heat, never to a road. Depth is the reporter's own word — ankle, knee, waist — ported from FloodDash's streetReports.js; no ticket text, name or photo is kept, and heat is summed onto a ~220 m grid.

Waiting for the first FloodDash reading…

Flood preparedness

Do not infer safe travel from rainfall

Electrical hazards, missing drain covers, currents and vehicle clearance are not measured by this panel. Rain totals alone cannot determine them. Do not enter floodwater; follow DDPM and BMA instructions. For emergency help call 1784, 1555 or 1669.

Weather models and future research

What is connected, and what is not

Open-Meteo supplies modelled current weather and hourly precipitation forecasts with explicit UTC valid times. It is not a street sensor or an inundation model.

TimesFM is not connected. No calibrated river crest, tide timing, road depth, vehicle clearance or shophouse lead time is supplied. The cinematic tour is an illustrative sequence, not model inference.

Weather source and units ↗

Limits

What this map cannot tell you

• Not inundation. A coloured canal says a gauge on it is near or over its bank. It does not say which streets are wet.
• Not hydraulic. The 1.5 km reach is a display rule, measured in a straight line — not a flow model.
• Thin coverage. A few dozen gauges watch about 4,000 km of waterway. Most canals have none; blue there means unknown, not safe.
• Roads without places. BMA DDS road reports carry no coordinates, so they stay a list by district. “Clearance not recorded” is not “still flooded”.
• May, never will. No ETA the engine does not trust. No data is not an all-clear.
• Not a warning. DDPM, BMA and TMD issue warnings. DDPM hotline 1784.

Build log

Nine weeks, one laptop, in the open

1. 4 AugDay one. A Worker on atlas.nonarkara.org; 432,077 OSM buildings as PMTiles streamed from R2 with Range passthrough; live iTIC traffic; rail and 4,063 waterways.
2. 5 AugCCTV, animated rain radar, the first gauge layer straight from ThaiWater, civic open data in D1. A land-price layer shipped and came out the same day — there was no real geometry behind it — and returned only as a Treasury road-level choropleth.
3. 6–8 AugAbout drawer, deploy-on-main CI, a clean console, the /api/atlas mirror for agents.
4. 9–11 AugLive incidents, geolocated news, Old Town heights from monument silhouettes.
5. 17–18 AugNASA aerosol, the wind field, boats and piers; the map stopped painting into a corner.
6. 2–5 Sep100 curated landmarks. The unverified icons were quarantined, not shipped.
7. 16 SepBangkok's 133 provincial rain gauges and the live strip.
8. 25 SepThe FloodDash relay and this waterway network. The same day: the live strip had been squeezing the phone map to zero height — fixed.
9. 26–28 SepCross-referenced road flood ground truth: 2,300+ Traffy Fondue citizen reports merged with BMA DDS bulletins and verified against FloodDash gauges within 250 m. Maholan community CCTV snapshot proxying with honest SVGs.
10. 1–2 OctCrisis flood heatmap ramp with dynamic pulse shader, honest version indicator (#atlas-version), and the 72-hour 30-second simulation tour.
11. 4 OctWeather and preparedness panel prototype. Current integration uses real weather model data; road safety remains unknown and TimesFM inference is not connected.
12. 14 OctTableau Public Bangkok survey (3,263 workbooks audited; sub-provincial data exhausted; Inside Airbnb adopted; HII 17-year satellite flood climatology followed to data.go.th root). 3D building height precision upgraded via building:levels * 3.15m and typology matrices without payload increase. Chao Phraya Basin Hydraulic Profile HUD (Pak Nam to Bang Sai C.29A) with real-time Gulf tidal lock calculation.

The pattern that repeats: ship the layer, then take out whatever cannot be sourced. The two removals — land price and the unverified icons — did as much for the map as any addition. Full history on GitHub.

Still building

Open questions

1. Hydraulic reach. Replace the straight-line 1.5 km with distance along the connected network, weighted downstream.
2. Gates and pumps. DDS publishes gate levels as a bulletin with no coordinates. Placing them needs a source, not a guess.
3. Water on land. GISTDA's satellite flood extent as the first layer that shows water outside the channels.
4. Check the colours against the ground. Compare judged stretches with DDS road reports, district by district, across the season.
5. Official extents. Retire the schematic envelope the day DDPM or BMA publish theirs as open GeoJSON.
6. Edge AI on WebGPU. Run TimesFM decoder weights directly in the client browser via WebGPU to compute street-by-street recession curves offline without Cloudflare round-trips.
7. Citizen sensor validation network. Pair street CCTV snapshot feeds with phone camera depth calibration to convert user Traffy photos into millimeter ground-truth water marks.

Credits

Where the water comes from

Flooded roads. Method stolen from FloodDash: streetReports.js (Traffy depth words) and streetFlood.js (residents → DDS → gauge → risk spot, in that order). Residents' reports: Traffy Fondue public API (NECTEC × BMA). Drivers' reports: iTIC / Longdo. Road network: OpenStreetMap via Overpass, ODbL. Judgement. FloodDash (flood.nonarkara.org), from HII/ThaiWater, RID and MRC telemetry. Gates and road reports. BMA Drainage and Sewerage Department (DDS), relayed by FloodDash. Waterways. © OpenStreetMap contributors, ODbL 1.0. Imagery. © Esri, Maxar, Earthstar Geographics. Weather model. Open-Meteo current conditions and hourly precipitation forecasts. TimesFM is a research direction, not an active source in this atlas.

FloodDash thanks Dr Preesan Rakwatin (depa) for guidance and for the remote-sensing flood work behind its connected-waterways and satellite-extent roadmap, and GISTDA for access to its data. That roadmap is this layer's roadmap too.