Adcon device status

Checked once a day, straight from Adcon's own admin console: which RTUs they have flagged "not in use" -- the one place that distinguishes a known-retired unit from a station that just stopped reporting.

📊 The Report button on the public pages

Every public page has a Report button: a graph from any period to any other, PNG and PDF, and the map’s picture. Signed-in users see everything. For the public, the stations’ readings (and every model against a station) and the CSV download stay off until ticked here. Soil, pan and battery sensors are never public.

Station data (detail)

One station, every sensor, one row per time step.

Latest values, all stations

Storms & thunder

Thunderstorm log: every place flagged as a possible, likely or severe thunderstorm (cold cloud top, heavy radar echo, raining gauge). Storm tracks: tracked storm cells with size, speed, heading and estimated arrival. Up to 400 days per download.

Rain sources vs gauges

MOWRAM radar, GK2A and JMA side by side at the same moment, replayable day by day, and scored against every rain gauge (/rain-compare). Signed-in users only, unless Allow public is ticked.

Station list & GPS registry

Admin only (contains source details).

Rain sources vs gauges

Source timing

For every source: how long after its time the source itself put its newest item out (nothing here can shorten that), and how long this server then took to store it. A source counts as in step when its newest item is stored; late when it has been out longer than a few checks and is still not stored. Checked every 2 to 15 minutes, the last 24 hours.

Sutron data

Check that Sutron data is arriving and import it by hand. The database feed (report app) is checked every 15 minutes. Sutron publishes new readings about once an hour, so readings already stored are skipped.

Report-app feed (database, near real time)

JMA 6-hour rain vs the model learned from history

For every 6-hour window the stations measured whether it rained. The same windows were forecast by JMA's ensemble, the model that learns each station's history, Open-Meteo and the combined model. Compared on identical windows: which is closer to reality, and are they close to each other?

Compare weather models

Pick two or more models, or all of them, and see which one stayed closest to the local stations' own readings in the database.

n     = number of hours scored
bias  = (1/n) · Σ (f − o)                       positive: the model reads high
MAE   = (1/n) · Σ |f − o|                       average size of the error
RMSE  = √( (1/n) · Σ (f − o)² )                 like MAE but large misses count more
r     = Σ(f−f̄)(o−ō) / √( Σ(f−f̄)² · Σ(o−ō)² )   1 = rises and falls together with the station
Mean of chosen = (1/k) · Σ f_model              the average of the k models picked

POD = H / (H + M)             share of real rain hours the model caught
FAR = FA / (H + FA)           share of the model's rain hours that stayed dry
CSI = H / (H + M + FA)        overall hit score, 1 is perfect   (the models are ranked on this)

5-minute rain nowcast formula ⓘ info

The weights behind the public "5-minute rain nowcast" card for the station currently picked above (). Public cards only name which formula produced a number; the actual maths lives here.

Satellite/radar "pattern reading" calibration ⓘ show

How a satellite cloud-top brightness or radar dBZ reading gets turned into an approximate mm/h: learned from every real station-hour where both that station's own gauge and its own satellite/radar reading exist, then applied to ANY point's current reading -- including a province/district with no gauge at all. See app/satellite.py:train_gray_mm and app/radar.py:train_dbz_mm.

Public feedback ("was this prediction right?") ↻ reload

The public 1-5 star rating anyone can leave on a station or a satellite/radar map point (app/feedback.py). Cambodia-only, rate-limited, and the visitor's own device location is checked against the target -- a record to compare by eye against what the prediction cards actually showed (each row's snapshot), never fed back into any model.

Ground gauge over radar / satellite: how "raining now" is decided

A rain gauge measures rain at its own spot, once a minute. Radar (every ~15 min) and the satellite image (every ~10–60 min) only estimate it from what they see aloft, and are older by the time they are drawn. So the public map trusts a fresh gauge over both, in the browser (app/static/app.js: overlayGauges / overlaySatGauges), re-run on every station refresh:

1. Which gauges count: a station reporting rain now (raining, rain in the last 30 min) whose latest report is no older than GAUGE_FRESH_MIN = 45 min.
2. Where: every radar / satellite point within GAUGE_RADIUS_KM = 6 km of that gauge.
3. Intensity: rate = mm in the last 30 min × 2 (mm/h, minimum 0.2); then dBZ = 10·log10(200 · rate1.6) (the same Marshall–Palmer relation the radar layer uses), floored at 14 dBZ ("light"). Bands: ≥13 light, ≥26 moderate, ≥42 heavy, ≥60 severe.
4. Radar points: raised to at least that dBZ / band, and the gauge's own spot is added to the rain image and the falling-rain animation.
5. Satellite points: label set to "likely rain", cloud shade raised to at least 200/255, next-hour chance shown as 100 %, 3-hour chance at least 90 %.
6. One direction only: a gauge can raise a reading, never lower it. A dry gauge does not erase an echo, because rain can be falling a few km away.

Not applied during playback (it replays archived frames as they were). The server-side archives and the nowcast are unchanged by this rule.

Cloud top temperature (CTT) in the rain nowcast

GK2A's Level-2 cloud-top temperature (app/satellite_ctt.py, actual °C, decoded from the product's colour legend) feeds the 5/10/15/30/60-minute nowcast as the "cloud" term: level = clamp((−10 − Ttop) / (−10 − −60), 0, 1), added to the log-odds as W_SAT[h] × level. It replaces the older IR-brightness level when fresh (≤ 50 min) and is never added to it — the two measure the same thing, so adding both would double-count. Each station's card/inputs record which one was used (cloud_proxy_source). CTT is a cloud-top reading, not ground air temperature.

Advection (cloud motion)

Built: the prediction simulation shifts rain chance downwind by the satellite's measured cloud motion (app/satellite.py:cloud_motion(); position(t+Δt) = position(t) + velocity × Δt, applied as the higher of a place's own forecast or the chance at the upwind place, for the next 2 h). The nowcast itself still treats a reading as staying where it was seen.

Data sources now (updated 19 October 2026)

• Satellite pictures (real clouds on the public map, the 17 products on /jma-satellite, the satellite dots): Himawari-9, Japan Meteorological Agency (JMA). A new picture every 10 minutes, about 12 minutes after the scan; 2 km over Cambodia and its borders, 5 km further out.
• Rain layer (Rain Image on the public map, its replay and prediction, rain on /clouds, the storm inference): the rain rate measured by GK2A, Korea Meteorological Administration (KMA), every 10 minutes, 2 km. Chosen over JMA after scoring both against our rain gauges.
• MOWRAM radar: in use again from 18 October 2026 — on the public map’s satellite & radar layer, in the rain comparison and on /check (below); since that afternoon from MOWRAM’s own radar server on the 2 km squares, the square popup giving its dBZ and rain rate. MOWRAM satellite (its GK2A receiver and its website’s satellite picture), added the same morning, was removed that afternoon at the user’s request with all its data.
• Detail pages: /gk2a-satellite (GK2A satellite) and /jma-satellite (JMA satellite), every 10-minute picture on one shared 0.02° grid (about 2 km, the smallest pixel either satellite publishes here) with the storm cells, kept in the database; tap a cell to read it.
• /clouds (Detail Satellite Images) removed on 17 October 2026: it was built on the old Cambodia GK2A mirror, which is no longer used; old links open /jma-satellite.
• Public map menu (17 October 2026): Satellite images = GK2A satellite / JMA satellite (switch the map's satellite layer between the two to compare them: every 10 minutes, each 2 km square cloudy / likely rain, tap a square to read it), Rain Image, Thunderstorm; JMA prediction and JMA pins moved to More. Removed, with the history only they used (backup in /backups/removed-components-*): Imagery, Detail Satellite Images, Satellite cloud image (its old, coarser squares), Cloud top temp.
• Stations: MOWRAM automatic weather station networks, read every minute; measured on 18 October 2026, 27 send a reading every minute, 23 every 15 minutes and 32 once an hour (sent up to an hour late). Forecasts: Open-Meteo, every 8 hours (below). All sources are named on the public map (attribution line and About); the line under the map’s title shows these intervals as measured (/api/timing).
• Every new picture is looked for every minute (both satellites publish one every 10 minutes; neither has a 2-minute picture over Cambodia).
• How often, how fine (checked 17 October 2026 on the operators’ own file lists): GK2A and Himawari-9 both picture Cambodia every 10 minutes. GK2A’s 2-minute scans (rain rate, visible) cover only the Korea area (28 pictures an hour there, 6 for the full disk), its East Asia area stops at about 20°N, and Himawari’s 2.5-minute scans cover Japan and typhoon target areas — so 10 minutes is the shortest interval over Cambodia. The server looks for a new picture every minute and grids it as soon as it lands. Rain rate and cloud-top temperature are made from the infrared channels: 2 km (0.02°) is their smallest pixel. GK2A’s visible channel is 0.5 km, but over Cambodia only in the full-disk file of 260 MB every 10 minutes (37 GB a day) — not fetched.
• Thunderstorm grid (app/thunder10.py, table thunder10_grid): every 10 minutes, every 2 km cell possible / likely / severe from GK2A’s cloud tops and rain rate and JMA’s heavy-rain areas — the storm clouds the satellites capture (neither carries a lightning sensor; FengYun-4’s lightning data is not public over Cambodia). On the public map’s Thunderstorm layer, on both satellite pages (Thunder) and in Reports.
• History: GK2A rain rate and cloud-top grids every 10 minutes from 1 January 2020 (app/history_import.py, container weathermap-history, ~6 GB downloaded and ~35 MB stored per day of history); its storm cells, storm grid and thunderstorm grid are made for every slot. JMA keeps only one day on its own site: its history starts 16 October 2026. Kept from 2020 on (app/partition10.py KEEP_FROM).
• Prediction (/prediction, app/outlook.py): GK2A + JMA only, no radar; the satellite part and the model part both calibrated on our rain gauges, a gauge raining now lifts the next hours, refreshed every 10 minutes, 2 km cells. The live proof is in the Prediction demo below.
• Rain comparison (/rain-compare, app/rain_compare.py + app/mowram_radar.py, collected by the history container; admin tab Rain vs gauges): MOWRAM radar — from MOWRAM’s own radar server (its login only in the server’s environment): CAPPI 480 km, the reflectivity 2 km above the ground every 15 minutes at 1.2 km a pixel, read by its 15-colour dBZ legend and put on the shared 2 km squares (the strongest echo in each; 13 dBZ, the weakest band, is about 0.25 mm/h), with its whole history from 15 May 2026 (app/mowram_radar.py, table mowram_radar; the pictures already stored at the radar’s own pixels were converted, the rest fetched). Until the afternoon of 18 October it was the website’s SRI pictures (app/mowram_web.py: no longer collected, its table kept). With GK2A and JMA it is scored hour by hour against every rain gauge (a source sees rain when any of its pictures in the hour shows it within ~2 km of the gauge), on the hours all three cover, every pair and per day, with the best source and the average day; refreshed every hour (each verdict stored in rain_seen). MOWRAM’s CAPPI radar server is no longer collected; what it stored is kept.
• Public map, 2 km squares only (17 October 2026, late): rain (GK2A rain rate), cloud (GK2A or JMA) and thunder are drawn only as their 0.02° squares — the blurred washes, falling drops, band icons and storm outlines are gone; the province / district dots and the storm markers come back only when ticked in each layer’s legend. More card: a Colour slider (10–100%, 80% to start) for all three layers, so the map and the province names show through, Names on top (province names, district names from zoom 9 and the border above the squares) and Cambodia only for all three. Tap any square (or point at it with a mouse): its province, district and nearest village, GK2A’s rain rate as a mm/h range with its 1 / 3 / 24 h estimate, the cloud top, JMA’s reading, the thunder level, the storm cell and the nearest rain gauge (/api/grid2km/cell, app/sat2km.py cell_info). Admin: the Rain vs gauges tab shows /rain-compare inside these pages, with its Allow public switch.
• 18 October 2026. One satellite & radar layer on the public map: choose the source (GK2A, JMA, MOWRAM radar — the chosen one bold on blue), then one of its pictures (GK2A: cloud & rain squares, rain rate, cloud-top temperature, infrared, storm; JMA: its squares and its 17 pictures; MOWRAM radar: its 2 km squares). Rain Image is GK2A’s rain rate, still the first view. Live and in playback from the stored grids; a tapped square lists every source there. Source timing (app/source_clock.py, Data Collection tab): for every source, its own delay and ours — measured KMA GK2A ~17–18 min after the slot, JMA ~12, MOWRAM radar ~2.5; ours ~0.5 min for GK2A, and JMA’s storm analysis now runs on its own thread, so it no longer holds the pictures up. Storm alarm: a storm heading into Cambodia shows one line on the map and on /prediction, blinking faster the sooner it arrives, with an alarm under 2 hours; a tap opens the storm view (GK2A infrared, the thunderstorm grid and the tracks, /?storm=1). Travel & transport (/travel, app/sectors.py): tourism sites, the two railway lines, the national roads 1–7, 48, 62, 76, 78 and boats (Tonle Sap, Mekong, coast) for every part of the day with warning levels, from the prediction — like MOWRAM’s road, railway and tourism products, but every 10 minutes. Sutron rain: RAIND (yesterday’s total) is left out and the 15-minute loggers’ running total is turned into rain per reading; 13–18 October repaired (backup in /backups). Cambodia only now blurs everything outside the border on every map.
• 18 October 2026, later. Storm alarm on the public map too: the same line and alarm as /prediction, from the same per-place arrivals (/api/storm/alarm, app/outlook.py storm_arrival), fetched every 10 minutes and counted down every minute; under 2 hours a popup (given once a storm, level and hour, across both pages) whose See the storm opens the storm view. Storm tracks on time: tracked every minute from the newest infrared picture (app/satellite_ch.py tick, app/stormtrack.py), ~1 minute after the picture instead of 30–35; live tracks count for an hour (40 minutes before, which blanked the alarm). JMA on time (app/satellite_jma.py): the infrared and the heavy-rain picture are fetched each on its own until both areas are in (an infrared with one area was never completed, and the heavy-rain picture waited for the infrared), straight from JMA’s server past its CDN’s 10-minute copy, a pass starting every minute — stored 1–3 minutes after JMA writes them instead of 7–9. The line under the title says what is measured (/api/timing: our clock’s 6-hour medians, each station’s own interval, the forecast’s real schedule). Open-Meteo every 8 hours: a cycle now costs ~2,500 of the 9,000 calls a day, so the two-hourly cycles stopped at 13:00 and left the forecast up to 18 hours old every night (15 October went past Open-Meteo’s 10,000); the cycles are now spread over the day by their measured cost (app/forecast.py cycle_sec, 85 % of the budget). Khmer district and commune names from Cambodia’s official gazetteer (app/data/kh_gazetteer.csv, from the SBMS seed data), matched by their codes in app/places.py: all 199 districts and 1,610 of 1,621 communes (GeoNames’ own Khmer names had lost their subscripts and vowel signs: ‘អរយាដាវ’ for អូរយ៉ាដាវ). Forecast check (/check): a new card scores MOWRAM radar, GK2A and JMA against the rain gauges, with our forecast made at the start of each hour, on the same station-hours (/api/check/sources, app/rain_compare.py check). Each source’s verdict at each gauge and hour is stored once (table rain_seen: every hour, the last 15 days again every day) and the gauges are joined when scoring, so any period scores in about a second. First week, 1,105 station-hours: GK2A 0.28 (CSI), JMA 0.19, our 1-hour forecast 0.15, MOWRAM radar 0.10 (it catches few of the gauges’ rain hours); MOWRAM satellite scored 0.35 before it was removed.
• 18 October 2026, MOWRAM satellite removed (the user): nothing fetches from MOWRAM’s GK2A receiver or from its website’s satellite picture any more, and nothing stores, shows or scores them. app/satellite_prod.py keeps only the JMA picture menu and the rain layer’s grid cells; app/satellite.py, satellite_ch.py and satellite_ctt.py read only JMA and KMA; app/satellite_ref.py and the admin panels Satellite & radar data sources and the picture evidence are gone. Deleted: the receiver’s 33 products (grids and hourly pictures), the website’s infrared, the storm backlog made from the receiver’s infrared, its clock rows and its comparison verdicts — a copy in /backups/removed-mowram-satellite-20261018. Kept: the 256-value grey mapping that puts JMA’s infrared on the scale the satellite dots and the storm thresholds use (fitted once, on both satellites).
• 18 October 2026, evening. Station chance of rain with GK2A, JMA and MOWRAM radar (the drawer’s Chance of rain · station forecast card, app/rainchance.py + app/srcnow.py): the station’s own calibrated chance, moved by what the three saw over the station in the latest 60 minutes on the 2 km squares (the rain comparison’s rule: GK2A 1 mm/h or more, the radar any echo, JMA a heavy-rain area or its likely-rain band, within ~2 km). The weights are learned from our gauges: every stored station-hour (rain_seen, 75,908 with the next hour measured) against the rain in the next hour and the next two, one logistic regression per set of sources present, the gauge’s own rain as an input so it is not counted twice. First fit: each source seeing rain raises the odds (all three, next hour: GK2A ×4.8, MOWRAM radar ×3.0, JMA ×2.2); over a dry gauge with all three seeing rain, Ankor Chey went from 34% to 61%. When none of them sees rain the chance goes down. Refitted daily and after the radar’s history is redone (kv srcnow_model); the card names what each source saw and how many points it moved the next hour; rain_track.sat_lift keeps those points with every tracked prediction. Replaces the fixed infrared lift (at most +12 points, from the older infrared picture) and the radar lift, which only ever went up. Method: /predict-method section 12.
• 18 October 2026, evening. MOWRAM radar decoded properly: a pixel counts when it has a legend colour and differs from the picture’s own map (the old 5×5 one-colour rule dropped about 70 % of the echo); the whole history decoded again; its rain-comparison score (CSI) went from 0.10 to 0.27. Every grid 2 km: the rain layer, the satellite dots, the channel pictures and the storm tracking read the shared 0.02° grid directly. Each satellite where it is best (checked on our data): GK2A first for rain and clouds on the map (JMA satellite under More, not in the legend switch); JMA first for the storm tracking and alarm (21 of its 117 alarms right against 10 of GK2A’s 69, and about 5 minutes sooner); JMA’s bands matched to GK2A (cloudy from grey 145, likely rain from 245; its rain CSI 0.34); the station lift weighs every source by what the gauges say. Map: every legend lists the picture’s classes in its own colours with the 2 km squares of Cambodia in each — a tap on a row makes them blink; falling rain and lightning over the rain squares (lightning over Cambodia only, the flash around the strike); the raining squares outlined from zoom 10; Cambodia only on by default; storms outside Cambodia recorded only when heading in; buttons Satellite images & history and Radar images & history; the new page /rada-images-mowram plays every radar picture since May; the history pages swap between GK2A, the radar and JMA on the same time. Phone page /mobile: only what is useful around the visitor — the stations near them, the rain around (the nearest rain, its strength), the nearest station’s card, storms coming within 60 km. Faster: GK2A’s files in 8 resumable parts at once (KMA served 12–15 KB/s a connection), the last 6 hours caught up after a restart; the map’s first answers kept warm.
• 18 October 2026, night. District rainfall /districts: the 197 districts on the 2 km grid with GK2A’s and the radar’s rainfall over any period (GK2A back to 2020 as its history comes in), the forecast per district, a replay hour by hour and a sortable table; each picture counts until the next one (at most an hour). Phone page: opens at zoom 11 on the visitor; zoomed in, Cambodia is no longer blurred (the Cambodia only mask is cut to the screen); the names no longer pile up; the minute’s refresh no longer pulls the map back. One grid, one set of formulas (checked across every module): every stored grid is the shared 750×800 grid of 0.02°; one rain scale for every picture and page — light under 2.5 mm/h, moderate to 10, heavy to 50, very heavy above (a radar’s dBZ through Marshall–Palmer: 29.4 / 39.0 / 50.2 dBZ) — where seven sets of numbers were in use; GK2A’s colour boxes read at their middle everywhere (90 mm/h no longer reads 86.5); GK2A’s storm cells are the cores of the thunderstorm grid, with its levels; one distance rule (111.32 km a degree) and one square area (4.84 km²); every grid read checks its size. Against the rain gauges (30 days): GK2A’s rain rate adds up to about 2.6–2.9× what the gauges caught; not corrected yet (a decision for the owner). MOWRAM’s radar in the prediction: its echo is one more key of the gauge-calibrated table of the next hours; held-out score 0.0954 against 0.0981 without it (Brier, lower is better). Fresh data: every source checked every minute (the radar too, before every 2 minutes); a new GK2A, JMA or radar picture rebuilds the next hours at once (before, every 10 minutes); the rain layer and the district sums every minute. GK2A history from 2020: only the part of each file down to the grid is fetched (44–48 %, the same grids); every hour back to 2020 first, then the half hours, then the rest; never while the live pictures are being fetched. KMA has served 12–40 KB/s a connection since 18 October 02:00 UTC, after this import took ~120 GB in half a day: at that speed the hourly pass takes about two months. Map: GK2A opens on its rain rate; ▶ in Now plays the last hour; a 2 km square of Cambodia has Accurate / Not accurate, accepted only from inside that square and from an address in Cambodia; the admin feedback page shows the address, the verdict and the square. Stations (admin): every station, the broken and never-fetched ones too, with its status, its last reading, whether it is on the public map and why not, and a show / hide switch in its row. Fixed: a slow query kept the station list past its time-out after a restart (now read by its index); the web workers could deadlock while creating their tables at start (now one at a time). Then: /districts sums up the history and the prediction with Cambodia’s average (each district by its area); /prediction says its next hours use the GK2A and JMA satellites, MOWRAM’s radar and the ground stations, and shows the radar’s reading; /check names what its 2 km prediction grid is made of; the radar’s legend gives the radar’s own timing. Measured delays (minutes after a picture’s time): GK2A rain rate out +14, cloud top +17; JMA +14 to +15 (its file says +12, its server gives it out 1.5–2.6 minutes later); MOWRAM’s radar +2.5 — each stored here within a minute, so on the map a satellite picture is 14–27 minutes old and a radar picture 3–19; the satellites’ own delay cannot be shortened here. The GK2A history now comes at ~7 s a file (KMA’s slowing eased): every hour back to 2020 in about ten days. Front page / (the map moved to /map; an old map link goes on to it): the visitor’s city as the browser finds it (else Phnom Penh), the main cities, the radar, the week, the next 24 hours, and tiles that say where each number comes from; its look follows the time of day at that city (morning, afternoon, evening, night from 10 minutes after sunset). Site title and text: Settings → Site title and text, English and Khmer (the subtitle may carry {stations} and {every}). Domain: the proxy’s domains come from SITE_DOMAINS in /opt/weather/.env (comma-separated, so the old and the new answer together during a move); every link is relative; Google sign-in comes back to the domain in use (each domain’s callback must be listed with Google).
• 18 October 2026, night. Every rain gauge counts (the user: “most stations have a gauge, why only 24?”): the checks and comparisons read each station’s rain per hour, which the 27 KMA stations never had — their gauge sends only its running total of the day (0.5 mm tips; it starts again at 00:01). That total is now turned into the rain of every 5 minutes (app/collectors/kma.py, common.rain_steps: the rise above the day’s highest reading, one rule with the Sutron 15-minute loggers), and their past is filled in: exact 5-minute steps from the stored readings since 6 October, every hour from the hourly totals back to October 2021 (10,190 of 10,190 complete days add up to the day’s own total). Rain-scored stations on /check: 82 of 82 (was 55); gauges in the rain comparison: 85 (was 58), its whole history scored again. KMA’s API sends 2–3-minute averages when asked for more than a day, so it is now asked one day at a time. The Sutron hourly stations get a Rain today (the sum of their hours since midnight), and the sunshine recorders (HRSSUN, 31 stations) are read: Sunshine today, in hours. Fixed: the hourly and daily roll-up rewrote the hour and the day at the start of its 2-day window from their last minutes only, so every hour leaving the window kept 1 of its 4 readings (an Adcon hour stored 7.2 mm of 15.0); the window now starts on a whole hour and day, and 65,067 hours and 3,030 days were rebuilt from the raw readings (the old rows kept in obs_hourly_cut_backup / obs_daily_cut_backup).
• 18 October 2026, night. Front page “My location” from the visitor’s own 2 km square (/api/outlook/spot: the prediction grid’s square, satellites + radar + gauges; /api/grid2km/cell: what GK2A, JMA, the radar and the thunder grid see there, in words): rain, sky and the hours; the temperature, humidity and wind still from the nearest station. The exact position is kept in the browser only (GPS with its accuracy, a dragged pin or a tap on the map; a fix rougher than 1 km, as from an internet address, says so). Falling drops while it rains there. The phone icon is gone; Travel sits at the foot of the side bar. /map?open=gk2a|radar|thunder|near|stations|jmapins|device|about|fullscreen and ?show= open the map on one view.
• Reports from the visitor’s square (front page, My location): what they saw in the 15 minutes before or after now (rain / thunder / no rain) and 1–5 stars for the forecast shown; kept as go:<row>:<col> in pred_feedback (columns seen, device) with what the card said. Accepted only from inside that 2 km square (a fresh GPS fix, 0.5 km margin), an internet address in Cambodia, and at most 2 an hour on one square from one device (a random id the browser keeps) or one address. Shown with the address in the admin feedback feed.
• Front page, raining now (the owner’s rule): the card says it rains and its drops fall when either satellite sees rain on the visitor’s square (GK2A’s rain rate, or JMA’s likely-rain band or heavy-rain area), the radar has an echo there, the square’s own hour is likely rain or a gauge within 3 km is raining; thunder on the square (any thunder-grid level, or a likely-thunder hour) switches it to lightning (the bolt flickers, the card flashes). The square is read again every minute.
• Data alerts (app/alerts.py, Settings → Alerts): every minute each source’s newest stored item against how late it may be (MOWRAM radar, GK2A, JMA 45 min; KMA stations 30; Adcon 60; Sutron 180); one down shows in red at the top of Settings (with the radar fetch’s last error, its login masked) and goes to Telegram once, every 6 hours while down and once when back. The bot token and chat id are set on that tab, kept in the config table and never shown again. Domains: camweather.info and www.camweather.info serve the same site as weather.vitsolutions.asia, each with its own Let’s Encrypt certificate (LETSENCRYPT_SINGLE_DOMAIN_CERTS).
• /jma-vs-prediction (app/jma_vs.py): JMA’s ensemble (/jma) and our prediction (/prediction) on the same 6-hour windows at every rain gauge within 40 km of JMA’s 25 points, each made before the window began, scored against the gauges (Brier, right, caught, false alarms, CSI, amounts), with ours also made as early as JMA’s run, the model alone and every JMA point. First 355 windows (16–18 October): ours 0.332 against JMA 0.613 — JMA’s driest member is usually above 0.2 mm, so it says rain in almost every window.
• Rain timeline (app/raintime.py, /api/rain-timeline, front page My location): the clouds’ motion measured at the spot by matching the rain pattern within ~40 km in GK2A’s two newest pictures (else the storm tracks), the newest picture moved on by its own age (~15–20 min) and then 5 minutes at a time for 2 hours with the motion held steady: when the rain should stop, start or get heavier, and whether the rain upwind is growing or weakening.
• Rain timeline checked and calibrated (app/raintime_check.py, daily in the history container, /check card, /api/rain-timeline/check): replayed every hour of the last 10 days at the 50 gauges reporting every 5 or 15 minutes, as made when the picture arrived (~15 min), each 15-minute block of the next 2 hours scored (CSI) against the gauge and against “as it is now”; the best rain threshold, reading width and speed scale are kept (kv raintime_cal) and used live. First run (10,156 forecasts): learned settings 0.221 in the first hour (first settings 0.162), as-it-is-now 0.233 — moving the rain pays from about 90 minutes ahead; the best settings were at the edge of the grid, which is now wider. Storm wording (NOAA): “storm” only from a tropical storm’s 62 km/h sustained wind; the satellite systems are rain clouds / thunder clouds; one wind scale (strong wind from 39 km/h, tropical storm force from 62, typhoon categories 1–5 from 119).
• Following the visitor (19 October 2026): the front page takes a fresh position when it is opened, every time the visitor comes back to it (from minimised, another app or another tab) and once a minute while it is shown — only where location is already allowed, never a new prompt. A sure fix (a phone’s GPS, within 150 m) moves “My location” when the visitor has moved more than 300 m, even over a place set on the map, and the card says “You moved: My location now follows you to …”. A rough fix (a laptop or computer, from Wi-Fi or the internet address) moves it only when the move is larger than the fix’s own error; over a place set on the map, or after “Forget my location”, it only suggests the new place (“Use this place” / “Keep my place”), and a suggestion turned down is not repeated until the visitor moves on.
• Speed for 500–1000 visitors at once (19 October 2026): nginx keeps every public /api/ and /static/ answer for a short time (10 s for the data, up to a year for versioned files) and one request refreshes it while the others get the copy — about 3,500 requests a second through the proxy in a 1000-connection test. Never cached: the pages (each view is counted), admin, feedback, anything with the admin cookie. Pages are kept in memory; page views are written by one queue per process. On the map, the rain and lightning animation no longer freezes zooming (draws ready-made drops, keeps its shaders; long pauses 4 s → 0.3 s). Nginx settings: /opt/proxy/tune.conf and vhost.d/wx-cache.inc (all three domains).
• Legend switches apply at once (19 October 2026): falling rain and lightning follow each check and uncheck immediately; two quick clicks used to load the animation twice and leave one copy that no switch could reach.
• Picture delay (19 October 2026): KMA puts each GK2A picture out ~15 min after its time (JMA ~14, Cambodia Meteo radar ~2.5); that part cannot be shortened here. Our part was 2–7 min; now the map’s two GK2A pictures have their own loop that looks every 20 s while one is due, fetches both at once and holds the history import back, and the map checks for a new picture every 20 s (was 60). The collector log gives each picture’s download and gridding seconds.
• Rain “estimated now” (19 October 2026): the newest GK2A rain picture (~15–25 min old on the map) is moved on to the present with GK2A’s own cloud motion — a motion field measured every ~100 km (rain pattern matched 20 min apart, else the cloud tops, else from the points around), the calibrated speed factor, a smooth warp. Labelled as an estimate in the rain legend, with a switch back to the picture as GK2A took it; never in playback; display only (the card’s rain decision is still GK2A’s own picture). Checked with every new picture: the estimate from 20 min before against what GK2A then saw in Cambodia, and the old picture kept as it was (CSI, last 24 h in the legend). First backtest: estimate 0.053, kept 0.022. Code: app/gk2a_now.py, /api/sat10/now.
• Site name and icons (19 October 2026): admin → Settings → Site name and icons: the name (English, Khmer), the site icon (browser tab, phone home screen) and the front page logo, on every page (app/branding.py; admins only, audited; images up to 256 KB, SVGs with scripts refused).
• Rain-cloud alarm colours (19 October 2026): by the cloud’s speed — yellow under 40 km/h, orange 40–60, red from 60; a red one pops up at once and blinks fast (the others pop up 2 h before arrival, as before). Map and prediction page.
• Estimate vs real (/estimate-check, 19 October 2026): every “estimated now” is checked against the picture GK2A really took 10, 20 and 30 min later (15 min lies between: GK2A comes every 10), and against keeping the old picture as it was, on Cambodia’s 38,214 squares: CSI exact and within one square, rain found (POD), false rain (FAR), area ratio; 24 h / 7 / 30 days, day by day, every picture, and one check side by side (start, estimate, real, hit/miss/false maps). The month back was measured afterwards (gk2a_now.backtest, two processes in the history container, ~9 h); new pictures are checked as they come. Kept 45 days (tables gk2a_now_field, gk2a_now_score). Linked from the rain legend and the map’s Prediction menu.
• Wind from GK2A (19 October 2026): over the GK2A and JMA pictures and GK2A’s rain (not the radar), arrows coloured by Beaufort force or streamlines over a speed wash (switch in the legend), from GK2A’s cloud motion (the field the estimated-now uses, ~100 km points, yellow dots where measured); flowing streaks; the stations’ own measured wind as black arrows; hover any arrow or station for direction, km/h, m/s, Beaufort and the probable ground wind (learned); the replay moves the arrows to each time, blending between pictures. Wind rose at a place: 24 h / 7 / 30 days, in colours that stand out from the map under it, with the up to 3 nearest stations (60 km, distance shown) compared at their own place: direction within 45/90°, typical difference, turn, speed ratio. All from numbers stored per picture (gk2a_wind, 45 days), no pictures stored. Learning (app/gk2a_learn.py, daily in the history container): the estimate’s speed factor (0.6–1.4) against GK2A’s next pictures, and the ground wind’s turn and slowing by hour against the stations, each checked on a day it did not learn from; shown on /estimate-check.
• Storm watch (/storm-watch, 19 October 2026): every GK2A cloud-top picture is checked for a tropical storm (app/tc_detect.py): a cold cloud shield (tops −40 °C or colder, 40,000 km² or more, round), turning in GK2A’s own cloud motion (cyclonic), and an eye (a warm spot walled by very cold tops in at least 7 of 8 directions) → organised storm cloud / possible tropical storm / tropical storm with an eye; tracked from picture to picture; for each: position, size, coldest top, turning, movement (its track, else the steering wind 300–700 km around), strongest cloud-level wind within 400 km, the chance to reach Cambodia in 72 h (15 ways: heading ±30°, speed ±30%, within 50 km), earliest arrival and closest approach to Phnom Penh. All kept (table tc_detect), history from April 2026. The animation is drawn from the stored grid (white cloud by coldness over imagery of the ground), not from stored pictures. Not an official warning; area 97–112°E, 5–21°N. Also: Facebook links (fbclid) open the front page; pages carry their build and reload once if a browser shows an old copy.
• Flood watch (/flood, 19 October 2026): four views, never mixed — Now and Next 2 h from GK2A (each district’s 1/3/6/24 h rain against its own GK2A history: watch at its 95th percentile, warning 99th, severe 99.9th, +1 on soaked ground; next 2 h adds GK2A’s moved picture; fast-growing storms flagged from cloud tops 10 °C colder in 20 min), Model 12 h (the prediction’s mm, plain thresholds) and NASA measured (LANCE VIIRS flood from NOAA-20 + NOAA-21, 250 m, 3-day window: km² of flood, recurring flood, water and cloud per district, daily since 24 June 2025, kept as district numbers and a 2 km grid for the replay — no image stored). NASA’s own tiles (VIIRS/MODIS flood, NOAA-21 true colour, GPM IMERG) for any day, NASA Worldview link. Code: app/flood.py, app/nasa_flood.py. Flash floods only: the Mekong’s river floods need river levels (MRC).
• Flood page, more (19 October 2026): the levels explained (watch = the district’s wettest 5% of hours, warning 1%, severe 0.1%, one step up on soaked ground) with typical impacts and indicative water depths (5–10 / 10–30 / 30+ cm — guidance: the satellite measures rain, not depth) and each district’s own thresholds in mm; NASA flood history also from MODIS (Terra + Aqua) back to January 2021 (NASA has no flood map for 2020), replay by a chosen date range; accuracy: the GK2A watch checked against NEW floods NASA measured (none in the 5 days before; cloud under 80%; river floods left out) — first result POD 16%, FAR 84%, CSI 0.09: not yet proven, and NASA sees only clear days; seasons: rainy (May–Oct) and dry (Nov–Apr) seasons compared, this season so far against the others, month by month over every year (what past years did, not a seasonal forecast). Storm watch history since April: 4,008 organised storm clouds, 36 possible tropical-storm detections (12 tracks), 5 with an eye.
• Compare sources on /check (19 October 2026): drag or tap 2, 3 or all of stations, GK2A, JMA, Cambodia Meteo radar and Open-Meteo; each pair on its own, never combined. With a station (or the model): every source read on the station’s own 2 km square, hour by hour, on the same station-hours (table cmp_seen, back to 14 May), the gauge as the truth, each station listed. Satellites and radar alone: the whole pictures over Cambodia’s 38,214 squares at every shared time, a where-they-differ map and any one time side by side. Freshness rule: “rain now” everywhere is the one rule (below) on each source’s newest picture of the square, never a moved or calculated value; nginx never serves an old copy while refreshing.
• Raining now — one rule (19 October 2026, chosen on the gauges): rain on a 2 km square when GK2A sees 5 mm/h or more, or when the Cambodia Meteo radar’s echo confirms GK2A’s rain or JMA’s; without a radar picture GK2A alone from 2.5 mm/h. Tested on 62,190 moments at 50 gauges (22 days): score 0.268 and right 93 % against 0.200 and 86 % for GK2A alone (which shows rain 3× too often); the plain 2-of-3 vote 0.234. JMA adds almost nothing to GK2A; the radar is the real check. The rule is computed for every square at once (app/rainnow.py): the front page’s map, its card, the city buttons, /travel and the map’s popups read the same grid, so they never disagree (0 differences over Cambodia’s 38,214 squares). Checked again every day (app/rainnow_check.py) and on /check (“Raining now”). The front page shows no accuracy figures: they are on /check.
• When the rain stops: the time rain this strong usually goes on at our gauges (half of the rains stop within 30 minutes) — 24 min off on average, against 40 for GK2A’s moved picture and 66 for the old “more than 2 hours”. How strong it stays: what most often followed each call (rain usually gets lighter within the hour).
• /check periods: every card has its own period (today … 12 months, the rainy season May–October, the dry season November–April, any dates) and one period sets them all; which satellite is best against the gauges, and how sure (500 re-draws of the days).
• /travel: now and the next 2 hours by the same rule on each place’s own square; places to stay, eat and visit from OpenStreetMap (hotels, guesthouses, restaurants, cafés, sights, ATMs, fuel, health, transport, bars), refreshed weekly; Directions from our own router (OSRM on OpenStreetMap’s roads of Cambodia, car / motorbike and walking) with the rain on the way at the time you get there, places on the way by the time of day (breakfast, lunch, the evening’s hotels, motels, guesthouses and dinner, the night’s hotels and bars), fuel on long drives, live tracking and spoken turn-by-turn directions (female / male voice); Nearby now; search the start and the destination, Go here on every tourism site, suggestions on the way on / off with tap-to-go and nearby now during the trip, and the trip kept in the browser across a refresh or a minimised window (20 October); Travel now without a destination: the map follows you, your speed, the rain here and ahead in your direction (within 30 km, with the time at your speed), places ahead by the time of day. Privacy: positions stay on the phone and are never stored; only the number of trips per day, district to district, and devices (a one-way hash) are counted (admin: /admin/api/route-usage).
• Raining now counts the rain gauges (21 October 2026): a gauge whose last 15 minutes measured 0.2 mm or more (its reading at most 30 minutes old) makes its own 2 km square and the 8 around it “raining”, even when the satellites and the radar see nothing — that morning a gauge measured rain all three missed. A gauge speaks only for where it stands (about 90 gauges, one or a few a province), so the satellite and radar grid stays the answer everywhere else. On /check: at the gauges the rule now catches 93 % of the rain; from the pictures alone 50 %. The station dots on the map say “raining” by the same test.
• The learning agent (its own container on the server; Settings → Learning agent shows it running, what it is doing and its results; /check shows them to everyone): every night at 02:00 it studies every quarter hour at every fast rain gauge — what GK2A, the radar, JMA, the other gauges and five weather models (ECMWF IFS and AIFS, GFS, ICON, Google WeatherNext 2) showed just before — and learns what really fell there, now and 15, 30, 60 and 120 minutes later. How smart it is: it sits an exam it has never seen — each of the last 7 days and each group of gauges is predicted by a model that learned only from earlier days and from the other gauges, as a square of the map without a gauge — and it replaces the site’s answer only where it wins by a clear margin on most days. It checks itself: on its first night it scored high by memorising where the gauges are and drew a ring of rain around Phnom Penh; the stricter exam caught it, the place clues were removed and nothing went to the public map. For “raining now” it does not beat the rule yet (0.277 against 0.289), so the rule stays; ahead it already beats the site’s timeline at 30 and 60 minutes (0.254 against 0.231; 0.237 against 0.210, better on 5 of 7 days) and at 1–2 hours, where the timeline foresees nothing (0.212), but its maps for those windows still show patterns that are not weather, so they stay off the public map until they look right. It finds faulty gauges by itself (Vihear Loung and Koh Thom report rain nothing else sees) and leaves them out. Every night with new data it learns again; a manager can hold the map on the rule at any time.
• Radar quality (21 October 2026): over 14 days the radar’s picture showed hills echoing 60 % of the time, its own map labels, borders and thin beams never echoing, and wedges its beam does not reach. Every night the agent marks those squares (about a third of the radar’s reach) and there the satellites decide, as outside its range. At the gauges the score stays the same (0.29) and the rain caught rises from 50 % to 55 %.
• The usage agent (Settings → Usage agent): once a day it reads only anonymous counts of which features are used — never a position, a route or a search — and writes suggestions for the team. On its first day (424 visitors, 907 visits in 30 days, once the server’s own load tests were taken out): 85 % of visits only look at a page, so the first screen has to answer the question by itself; after “near me” 45 % open a station; the busiest hours are 20:00–22:00. For visitors: on the map one “people often open … next” suggestion after an action (switched off with one tap), and on the front page a link to the rain along the road (/travel) when rain is coming.
• The motion agent (every night): is the cloud and rain motion right between two pictures? Each picture of GK2A, JMA and the radar is moved on and compared with that source’s next pictures; it learns how much faster and which way the motion should be turned, the wind at the ground under it, how fast the storms really go, and whether a storm rains. First night (scored on the last 2 days only): moving the pictures beats keeping them still for every source (JMA 0.69 → 0.77, the radar 0.55 → 0.59, GK2A 0.62 → 0.65); JMA’s arrows now go 1.4× faster and turned 15°; storms move slower than their measured speed (0.6× for the satellites’ storm cells, 0.8× for the border watch), so their arrival times are later and closer to what happened; over Cambodia, strong storms rained on the gauges under them 66 % of the time within 30 minutes (growing 66 %, steady 70 %, weakening 57 %), the map’s storm popups say so.
• The agents apply what proved better — and say so: every value an agent learns replaces the one in use only where it was better on days it never learned from; the rest stays. Settings → Suggestions lists what a person should do next (a gauge to visit, a key to add, a design to decide), most important first, then every change the agents applied and kept, with the evidence.
• Visitors by province and district (Settings → Visitors): zoom to Cambodia — by GPS (only where visitors let a page use their position; the server keeps the district, never the coordinates) and by internet address. Test traffic and robots are no longer counted (14,592 of 19,436 rows were the server’s own load tests).
• The site follows the best source, re-decided every night: GK2A and JMA are checked against our rain gauges on the same moments of the last 7 days; the leader is shown first — the map opens on it, its button has a ★, the front page names it under its map — and it changes only when the other is clearly better. On 21 October: JMA (0.193 against 0.179, best on 5 of the last 8 days; GK2A was best on the other 3), so neither is always the best. Where the radar does not decide, which satellite decides “raining now” is learned the same way over 30 days (GK2A, JMA, either, or both): GK2A stays there for now. The pictures stay the same; what is shown as real comes from what the gauges confirm.
• Faster for many visitors at once (21 October 2026): the database keeps 12 GB of its own cache on the 49 GB server (4 GB before) and the cache hit in Settings → System is the last 15 minutes’; 10 web workers (5 before); “the newest picture” is looked for in the last month only (the picture history now reaches back to 2020 and is kept for ever); pictures read again and again stay in memory for 10 minutes. Load test straight on the app, without the proxy’s cache: 1,000 requests at once answered at 1,200–2,400 a second, 3,000 at once at 1,400–2,000 a second, no errors.
• The map’s player picks its picture: GK2A, JMA or the Cambodia Meteo radar (or none) right in the player, beside the period; every picture legend says what is playing back, from which source, and the time of the picture on screen — folded or not — and its counts follow the replayed picture. The map opens on the whole picture (Cambodia only is one tick away) and with no station network ticked (Stations: Sutron, Adcon, KMA).
• The agents learn what makes rain (21 October 2026): besides the pictures, every square now carries the land under it (height, hills, the sea, forest, fields, the city’s built-up ground, water and wetland, from open elevation and land-cover maps) and the air over it (temperature, humidity and how close it is to its dew point, pressure and its fall, low and middle cloud, the energy for storms, the wind and how hard it pushes up the hills). The stations’ own sensors (pressure falling, a sudden cooling, gusts, sunshine dropping) are learned for the stations themselves, never drawn on the map. Every night each forecast window is tested with the pictures alone, with the land and the air, and with the sensors, on days and gauges it never saw, and keeps only what wins.
• The satellites at each gauge: the picture 10–20 minutes before matches the gauge best (the satellites see the rain first); GK2A shows 67 % of the gauges’ rain on their own square, 79 % within ±20 minutes; the land changes its meaning (best over cities, poorest over forest and the floodplain). Moving the rain back for the satellites’ viewing angle (parallax) was tested and did not help the rule, so the pictures stay as drawn.
• Why the ground wind is not the clouds’: KMA and Adcon stations only (not Sutron, not Phnom Penh’s buildings); they agree 38 % of the time — 49 % with no rain near, 31 % with rain within 25 km (a storm’s cold air spreading out), 64 % on the hills, 31 % in the city; the ground wind is about half the clouds’ speed and turned 20–30°. The arrows stay as the clouds move.
• Sounds and a moving sky: on the map, thunder with each lightning strike and a wind sound by the wind speed, each with its switch and a speaker button in the player (never before a click); on the front page the sky moves with the weather — clouds drifting, rain falling, a storm’s lightning — behind the page and the weather card.
• Seven weather-service pages (21 October 2026), each on its own and linked from the front page’s menu: /warnings (yellow / orange / red for every district: heavy rain, flash flood, thunderstorm, strong wind, heat, poor air, high river, tropical storm, rough sea; an alert for your district; a warning line on the front page), /rivers (the Mekong and the Tonle Sap: flow, the next 30 days, the usual for the date — GloFAS), /typhoon (tracks and 5-day forecasts from JMA’s typhoon centre, how close each comes to Cambodia; GDACS alerts), /farm (water balance, spraying and drying hours, days since rain, this month against the usual, soil moisture), /air (air quality and haze, CAMS), /heat (feels-like temperature and UV), /sea (waves and wind for boats off the coast and on the Tonle Sap). All open data fetched on its own clock, kept for ever; the front page also reads the weather aloud.
• History from 2020 on every weather-service page (21 October 2026): a History button under each map — pick the day and hour, step, play, share a link — and the map shows that time as squares: feels-like temperature, UV, air quality (from 31 July 2022), rain, ET0 and the water balance, soil moisture, waves and gusts, each day’s warnings, every typhoon’s track over the gusts or pressure, and each river station’s flow. From Open-Meteo’s own archive of ERA5, ERA5-Ocean and CAMS on AWS (only Cambodia’s squares are read; no API calls), IBTrACS for the storms and GloFAS for the rivers; kept for ever.
• Block visitors by place (Settings › Access): continents, countries, regions, cities or IP addresses — blocked, or the only places allowed; nginx applies it within a minute (403), from the GeoLite2 database; the admin who saves and the server itself are never blocked; any address can be checked.
• Travel on Google Maps (23 October 2026): with the two Google keys in the server's .env, the travel page uses Google's map, its directions (with traffic) and its place search; without them, OpenStreetMap and our own router as before.
• System as a slide (23 October 2026): the What runs behind it drawing has a Play button: each part lights in turn with a live caption.
• The formulas (23 October 2026): How we measure names every method (learning, scores, calibrations) and shows live this week's winning satellite, GK2A or JMA.
• What runs behind it (23 October 2026): on the How we measure page, the whole system drawn live: the sources, the services (12, each with a running dot), the 7 agents that learn every night with their last run and decisions applied, the gateway, the visitors and the Claude Code agent; pictures for slides in docs/system-poc-en.png and -km.png.
• How we measure (23 October 2026): a new page, the last icon of the menu, and a card on About: an animation of how the site learns to tell rain on every 2 km square — the satellite's picture, the station's rain gauge and sensors, the learning samples, the answer for every square; a Play button, English and Khmer.
• Languages and phones (23 October 2026): every page changes between English and Khmer at once, without reloading, and follows a choice made on another page or tab; www.camweather.info goes to camweather.info; on a phone a tap on a map square opens a small card at the foot of the screen (✕ or a tap on the map closes it, More details for the whole table); the dashboard's controls are one tidy row like the radar page's.
• Protection (22 October 2026): only camweather.info (weather.vitsolutions.asia retired; the Excel export at camweather.info/excel/); no answer by the server's IP; an “Updating” page during restarts; Fail2Ban blocks floods, scanners and password guessing, with its limits in Admin > Settings > Protection; the feedback form asks a sum or an animal; a man's English voice from the server; the travel page asks for the location when a destination is chosen.
• Phones, round three (22 October 2026): on the satellite, radar and district pages the days, options and legend behind small round buttons and a slim player; the map without the extra buttons and tabs (the legend behind one round button); the prediction page's storm line one line (a tap for the rest) and its rain-chance panel a round button; the travel drive card small under the same menu; English spoken in US English.
• Every page the same menu (22 October 2026): all 27 pages in coloured icons, the most useful first, fixed at the top of a phone; an About us page with the purpose, the disclaimer (follow MOWRAM and the local authorities for official information), contact info@camweather.info and a feedback form checked for robots and kept for Admin > Feedback; pages move only up and down on a phone; a swipe left changes light/dark, a swipe right Auto/Day/Night.
• Phones first (22 October 2026): one menu on every page — small icons across the top on a phone, the page names down the left on a computer; the home page's header in two rows on a phone, only the big weather card, less padding; no Report button on a phone (the dashboard is in the menu).
• Forecaster desk (22 October 2026): for meteorologists — GFS and ECMWF charts over Southeast Asia (sea-level pressure, 850 hPa wind and humidity, 500 hPa height with the 5880 m ridge line, 200 hPa jet, CAPE, lifted index, inhibition, K index, precipitable water, freezing level) to +72 h, model soundings (Skew-T) with their indices for 8 towns, and meteograms of four models with the ensemble spread for any point of Cambodia.
• Rain chances by province (22 October 2026): the new “Rain chances” page shows, for every province and each of the next 7 days, the chance of any rain, 10 mm, heavy (20 mm) and very heavy (50 mm) rain from the 51 runs of ECMWF’s ensemble forecast, how much rain is likely, and how far the runs agree. The Telegram morning message adds the province’s chance of 20 mm or more.
• Destinations: when to go (22 October 2026): a page for the 16 tourist places — the weather now and the best 3 hours outdoors, the next 7 days, sunrise and sunset with the sky expected then (Angkor Wat’s at the top), and each month coloured best / good / fair / poor from 30 years of rain and six years of daily heat, waves and haze, with the best months to go (Angkor: November–March).
• Lightning and thunderstorms, 0–3 hours (22 October 2026): the new “⚡ Lightning” page shows every district with a storm now or within about 10 km, where storm clouds are growing fast (likely within the hour) and where the air could make storms in the next 3 hours, with clear lightning-safety advice and how often its calls came true. The Telegram bot sends “go indoors” when a storm is near a followed district.
• Telegram alerts by district (22 October 2026): the site’s own Telegram bot — anyone chooses their district (or taps “Alerts on Telegram” on the warnings page) and gets its warnings when they reach orange or red, the all clear, and every morning at 6:00 the day’s forecast, in Khmer or English. Turn it on in Settings → Alerts → Public Telegram bot with a bot made by @BotFather. Only the chat and its districts are kept.
• Satellite rain corrected to our gauges (22 October 2026): GK2A’s rain amounts are now multiplied by a factor learned every month from our rain gauges (October: × 0.49), so district totals, the dashboard, reports and the front page’s “rain today” agree with what the gauges measure; “raining now” and the rain rates stay GK2A’s own picture. The districts page can still show GK2A as it is.
• Move the Report button and the legends (22 October 2026): drag the Report button anywhere (drop it back in the corner to let it place itself) and drag a map legend by its grip ⠿ or any empty spot of it (double-click to put it back); each page remembers where. The map’s title no longer has the station-count line under it.
• A clean map, the station card at the side, the management demo (22 October 2026): the update times moved from the map’s header to a small “ⓘ Update times” link beside the data credit; the nearest station’s card waits at the right edge (tap or swipe left to open it, swipe right or › to put it back, remembered); the dashboard’s sources in Khmer; a 15-minute management demo in English and Khmer with pictures of the live site (docs/MANAGEMENT-DEMO.md).
• Live dashboards (22 October 2026): /dashboard and the Report panel’s first tab: rain, heat, air, sea, rivers, warnings, storms and El Niño, the period (7 days to a year) against the one before or last year — headline cards with their change and trend, the values now marked live, the period day by day, a gauge, the top five, a share and a province table; refreshed every minute, light or dark. Rain shows the satellite’s estimate beside what the gauges measured.
• JMA satellite when JMA’s real-time page is down (22 October 2026): since 03:00 Japan time on 22 October JMA’s real-time image page answers “not found”; the infrared now comes from NOAA’s archive of Himawari’s raw scans (as fast as JMA’s page) and 12 colour products from JMA’s bosai map tiles, the missed hours filled in; only JMA’s heavy-rain areas pause until the page is back. The rivers’ usual level is now made from the daily flow the site keeps.
• Travel: when to leave, heat, the dark, breaks, floods, boats, speed limits, the arrival time shared, offline (22 October 2026): the trip card compares leaving now and in 1, 2, 3 hours (rain on the way, the hottest hour, an arrival after dark); heat and UV advice for a motorbike or a walk; a warning when you arrive after dark; a 15-minute break every 2 hours with the café or fuel stop near it, and a spoken reminder while driving; flood watch, NASA’s flood water and the rivers along the road; trips to the islands go to the pier and show the boat and the sea at the crossing’s hour; OpenStreetMap’s speed limits while driving; a link that shares only your arrival time (never your position); and the page and your trip work without the network.
• Reports on every page (21 October 2026): a 📊 Report button on every page: choose the data, the place and any period to any other, by the hour, the day or the month; line, area, bar, stacked bar, step and points graphs; every weather model against a station; downloads as PNG and PDF, each with its period and the time it was made; the map’s picture, Cambodia only or the whole map, as PNG or PDF. Every change redraws the report; each unit has its own axis. For the public, the stations’ readings and CSV stay off until Settings › Reports allows them (“Allow public: station readings and models”, “Allow public: download CSV”).
• El Niño and the climate outlook (/climate), volcanoes (/volcano), the storm watch from 2020, warnings on your way (21 October 2026): El Niño now with NOAA’s and JMA’s official outlooks and ECMWF’s forecast, Cambodia’s rain and heat for the coming months and every month since 1940, the long records back to 1866; the region’s volcanoes with the Himawari Ash RGB picture every 10 minutes and the Tokyo ash advisories in plain words; the storm watch reading Himawari’s infrared back to 2020; on /travel the warnings of the districts a route crosses and the emergency numbers.
• Pages: / (the front page), /map (map: GK2A satellite / JMA satellite switch its satellite layer, 2 km grid), /gk2a-satellite and /jma-satellite (every product, storm cells, thunder grid, From–to days, Cambodia only), /prediction, /check, /mobile (phones), /rada-images-mowram (the radar’s pictures and history), /districts (district rainfall). Old addresses: /satellites-10min → /gk2a-satellite, /clouds → /jma-satellite.

Demo: the front page — where every number comes from open the front page ↗

The front page (/; the map is at /map) shows the city the visitor lives in (found by the browser, else Phnom Penh). Its numbers come from three places: the nearest MOWRAM station (measured), our prediction (calculated) and astronomy. Each tile says its own source. Files: app/static/home.html; data /api/stations, /api/outlook/point, /api/mowram10.

Demo: the “real clouds” layer — how it is made and how accurate it is open the public map ↗

On the public map: the satellite panel → Show real clouds and Rain and thunder. The clouds are drawn from the real Himawari-9 (JMA) picture and the rain from the rain rate measured by GK2A (KMA); only the soft, billowy edges are decoration. Files: app/static/clouds.js, app/static/storm3d.js.

Demo: the Prediction page (next 24 hours, 7 days, month) — the formulas and how far to trust them open the page ↗

On /prediction every province, district and station gets a prediction of rain (chance and amount), thunder, heat and storms coming from outside Cambodia. Code: app/outlook.py. Refreshed automatically: the next 24 hours every 30 minutes, the 7 days every 6 hours, the month every 12 hours.

Demo: the satellite picture menu (rain, cloud, sea, colour pictures) — how it is made and how reliable it is open the page ↗

The Picture menu of the former /clouds page (removed 17 October 2026, old links open /jma-satellite) lives on as the product menu of /jma-satellite, with every product of the Himawari-9 (JMA) real-time imagery: the heavy-rainfall-potential areas, 4 bands (infrared, visible, water vapour, shortwave infrared) and 12 colour (RGB) pictures and combinations — plus the satellite rain estimate.

Demo: how the code reads a satellite picture, cell by cell ↻ reload frame

This is the newest real infrared picture the system stored. The code never "looks" at the picture like a person: it divides it into small squares, averages the grey level in each, and applies a few fixed threshold rules. Switch the view to see each component, and click any square to see its numbers and which rule fired.

LABEL_BANDS = ((0, "clear"), (145, "cloudy"), (245, "likely_rain"))
IR_WARM, IR_COLD = 40, 230

def label_for(gray):
    lab = LABEL_BANDS[0][1]
    for edge, name in LABEL_BANDS:
        if gray >= edge:
            lab = name
    return lab

def cold_level(gray):            # used by the rain nowcast
    return max(0.0, min(1.0, (gray - IR_WARM) / (IR_COLD - IR_WARM)))

def storm_level(ctt, dbz, gauge):
    if (ctt is not None and ctt <= -70) or (dbz is not None and dbz >= 50): return 3
    if (ctt is not None and ctt <= -62) or (dbz is not None and dbz >= 42): return 2
    if ctt is not None and ctt <= -55 and ((dbz is not None and dbz >= 30) or gauge): return 1
    return 0

CAPPI_DBZ_COLORS = ((60,(0,0,0)), (55,(190,0,0)), (50,(255,64,0)), (45,(255,128,0)),
                    (42,(255,192,0)), (38,(255,255,0)), (35,(0,57,0)), ... (13,(165,255,255)))
COLOR_MATCH_MAX = 18
LABEL_BANDS = ((0,"clear"), (13,"light"), (26,"moderate"), (42,"heavy"), (60,"severe"))

def _dbz_for(rgb):                      # nearest legend colour
    best, best_d = None, None
    for dbz, c in CAPPI_DBZ_COLORS:
        d = abs(rgb[0]-c[0]) + abs(rgb[1]-c[1]) + abs(rgb[2]-c[2])
        if best_d is None or d < best_d: best, best_d = dbz, d
    return best if best_d is not None and best_d <= COLOR_MATCH_MAX else None

def rain_mmh(dbz):                      # Marshall-Palmer  Z = 200 R^1.6
    z = 10 ** (dbz / 10.0)
    return round((z / 200.0) ** (1 / 1.6), 2)

Where does the data come from, and where is it stored when finished?

Flow, the same for every source:  Source website → fetch (collector container, on a timer) → read (the rules above, per pixel / per place) → store (PostgreSQL database, plus the picture itself) → serve (/api/…, cached 1–5 min) → the public map and this admin page.

Everything lives in the PostgreSQL database weather_data on the weather server. A live check of every source (is it answering, how old is the newest picture, how old is what we hold) is in Custom Prediction → Satellite & radar data sources.

Growth

Row counts are PostgreSQL's own estimates (instant, usually within a few %). Press Count on a table for the exact figure. Monthly partitions of the readings table are added into one line.

Website feedback

Messages sent with the feedback form on the public About page (/about#feedback). Each one passed the "I am not a robot" check; the sender is kept only as a scrambled code (no address, no position). Newest first.

Testing addresses (not counted)

Find an address by its start, its provider or its city, then exclude it as testing: its network (the /24 of an IPv4 address, the /64 of an IPv6 one) is never counted again and its past visits are deleted, the daily totals rebuilt. Count again undoes the exclusion from now on; deleted visits do not come back.

Visitors in Cambodia by province and district

By GPS: only visitors who let the front page or Travel use their position, counted once a page; the server keeps the district it falls in, never the coordinates (green circles on the map). By internet address: the city the provider is in, approximate (blue circles). The period is the map’s.