What does this map show?

6 August 2026 — we answered "how many fires do we miss?" with a number for the first time

• Detection completeness measured — against an independent source. Until now we said "we may miss small fires" without knowing how many. We compared against EFFIS burnt-area maps: we catch every fire above 1,000 hectares and 81% of those above 30 hectares. Because burnt area and active fire are two different measurements, this is not us grading ourselves with our own ruler. Added as the "How many fires do we actually see?" section above.
• We corrected the measurement twice, and both corrections changed the answer. The first figure was 73% — then we realised we had measured using only two of the four satellite sources the site actually uses. Adding the missing two gave 81%. Second: the date EFFIS records is systematically late, because a burn scar only appears in imagery after the fire has burned, which makes it easy to think we missed something.
• Three improvement ideas were tried; none held, and nothing on the map changed. We wanted to add Landsat as a second verification source — we measured it, and that product does not cover Türkiye at all. We tried tuning the cone radius by fuel type; forest and scrub already want the same setting. And we tried machine learning for direction once more, this time only on "clean" fires: it briefly looked like it worked, until a deliberately meaningless filter produced the same apparent gain, which showed the effect was not real. None of the three shipped.

5 August 2026 — we retested the "nine out of ten" claim against ten times the data; this time it held

• The coverage claim was independently confirmed. The site says "nine out of ten of the cells the fire advanced into stay inside the shape we draw". That claim turned out to be wrong once before (see the 2 August note), so we tested it again — this time not with 232 cases from Türkiye but with 2,383 cases and 41,103 cells from eight seasons across the Mediterranean basin. Result: 90.4% overall, 89.1% for Türkiye alone. When we picked the setting from data outside Türkiye — never seeing Türkiye at all — and then tested it there, we still got 89%, so the figure is not luck specific to one country. Holding out each season in turn gives 90% on average.
• We tried to shrink the cone, found no meaningful way, and chose to change nothing. Three separate routes to a narrower cone at the same reliability were tested: varying the safety margin with wind speed, letting the fire correct itself from its own previously observed advance, and predicting "this fire will not grow". Each gained only about 3% on its own, and combining them did not add up — they all trim the same excess. We did not break something that works for a gain smaller than our measurement uncertainty.
• Machine learning did not beat the physics model at predicting direction. We trained a model on eight seasons of data. In the geography it was trained on it was a few degrees better than the physics model, but in Türkiye the difference vanished into noise. We write this down because saying "we added AI" is easy and saying it did not work is hard.
• Nothing changed on the map in this round. The work was measuring and verifying.

4 August 2026 — a new satellite, pixel honesty, and a failure we had been silent about

• The reach shape now changes with the wind — and our earlier measurement was broken. We drew the shape as "2.4 times longer ahead than behind", the same at every wind speed. When we measured that number we took the fire's direction from its centre of mass; in a 30-kilometre fire, the direction that looks "backwards" from the centre can be the flank of the front. We redid it per 375-metre cell, each cell measured from its own nearest burnt neighbour (5,719 cells): in light wind a fire really does advance almost equally in every direction (1.2×, nearly a circle), while above 15 km/h it barely moves backwards at all (5.5×). One fixed shape got both wrong. The new shape both covers more (88% → 91%) and is smaller — in strong wind the area we draw shrank by a third, because we no longer spend it on the side the fire does not go.
• A "Clean" button, and information bands you can dismiss. Clean view hides the top strip, the list on the left, the legend and the data-source notes, leaving just the map — for screenshots, or simply for looking at the map. The small handle at the top right, or ESC, brings it all back. Each information band at the top now also has an × of its own, so you can dismiss the ones you have read; on a phone that gives the map noticeably more room. One exception is the unverified heat source warning: dismissing it applies only to that alert, and it returns when the satellite sees a new source. We did not want the site's earliest warning to be silenced for good. In clean view the Algow signature sits at the bottom left and a very faint yangin.algow.net watermark covers the map: the platform is free and open source, but its imagery should not be sold on as someone else's work.
• Heat sources in Syria and Iraq moved to the bottom of the list. Our satellite window also covers neighbouring countries, and most of the large power readings there are not fires: it is gas flared at oil wells, industrial heat that never goes out. A 1,171-megawatt flare near Mosul was pushing a real 512-megawatt forest fire in Çankırı down the list. We did not remove them from the map — we show what the satellite sees — but they now sort last. Greece and the Balkans are excluded from this: a large fire there is a real fire.
• Fire weather had been broken for months in Greece and the Balkans — now fixed. When we widened the map westwards we had written the boundary by hand in six separate places and forgot to update five of them. For every fire west of 24.9° east, wind, humidity, fire weather index, smoke forecast, slope and fuel type came back silently empty. Nothing errored, so it simply looked like "no data". Some of the largest fires on the map were in exactly that area. The boundary now lives in one place and a test stops it being copied again.
• We now draw the satellite pixel at its real size. Zoom in and each detection gets a dashed ellipse around it: that is the cell the heat sits inside. The dot does not mean "the fire is exactly here". On VIIRS the cell is 375 m at best; on MODIS it grows to 4 km at the edge of the swath — so you can now see why a position sometimes looks off.
• The active fire count dropped a little, because it used to be slightly inflated. Every detection carries a confidence flag and we were ignoring it. Low-confidence daytime detections are now drawn faint and left out of the counter. Daytime false alarms are mostly sun glint: greenhouse sheeting, metal roofs, water. At night that mechanism does not exist, so night detections are not downgraded — fires grow at night too.
• A new satellite: Sentinel-3. Europe's SLSTR adds about four passes a day at 1 km. Sharper than Meteosat but slower: we measured roughly two hours from sensing to publication. So it does not close the blind gap, it adds another pass to the sharp layer. It also reports its own error margin, which is why we can say "30 ± 4 MW".
• We say so when the sensor saturates. In a very intense fire the satellite's heat channel saturates and the power it reports becomes a lower bound. The card then reads "VERY INTENSE": the figure shown is the smallest the fire could be.
• Where a fire is heading is now written out, not just drawn. The projected reach was already on the map but stayed abstract for anyone not looking at it. It now reads "in this direction: X ~7 km". This is not an evacuation warning and does not mean the fire will get there — wind turns and crews intervene.
• Smoke: a measurement when there is one, the model otherwise. Until now the smoke line was entirely model output (ECMWF/CAMS) and we never said so. If a ground station within 25 km of the fire is reporting, its real value now appears on its own line. To be honest about it: in Türkiye that line will usually be missing, because the national air quality network's open feed stopped in May 2023 — of 406 registered stations only 8 still report and all of them are in Istanbul. It works in Greece and the rest of Europe. When there is no station, the line below says the figure is model output only.
• Sharing a link now shows our opening screen. The preview card in chat apps and social media is the same screen you see when the site opens. And for search engines and browsers without JavaScript, the page summary now comes from the server: the age of the newest detection, the number of active events and the emergency numbers stay readable even if the map never loads.

3 August 2026 — Greece now covered, firefighting aircraft on the map

• The area the map covers now extends further west. Previously we only saw Greece's Aegean coast and its eastern islands; the western mainland, the Peloponnese, the Ionian islands and the south of Crete were outside the box entirely. On the first day of the wider coverage the largest fire there measured 2,645 MW — twice the size of the largest fire we had ever recorded in Türkiye. Fires in neighbouring countries carry an "ABROAD" badge and are excluded from the Türkiye counter.
• The event list is now ordered by fire size. Cross-border fires used to be pushed to the bottom unconditionally. Once coverage widened, that rule started hiding information: 1,230 MW was burning at Corinth while a 182 MW fire sat at the top of the list. Which country a fire is in is now told by the badge on the card, not by its position.
• A new layer shows the aircraft and helicopters working a fire. The source is an open ADS-B network run by volunteers. We genuinely know the aircraft type and registration; we do not know its mission — so read this layer as "there is a response here", not as an official tasking record. An absent aircraft does not mean there is no response either: not every aircraft broadcasts, and coverage is weak over mountains. It is not counted as an active fire.
• Our own mistake: part of the wind map was silently coming up empty. Widening the coverage pushed us past the per-minute limit of the service we get wind data from, and the northern part of the grid could not be filled. That matters, because a fire with no wind data gets no spread forecast either — people were seeing an incomplete map without knowing why. We made the grid slightly coarser; the whole coverage now fills. The cost is about 11 degrees of extra coarseness in the direction calculation, which is well inside the forecast's own margin of error, so it is the right trade.

3 August 2026 — the fire we missed, news reports, first alarm

• On 2 August we failed to show you the fire at Bayramiç — even though we had seen it. The high-resolution satellites (VIIRS/MODIS) never saw it: the fire started and was contained between their passes. The geostationary Meteosat saw it at 16:08 local time, 91 minutes before the first news report, peaking at 303 MW. But that detection was only a temporary orange ring on the map: it never entered the fire list, was not counted, raised no alert, and vanished once the fire died down. We had the information and did not tell you.
• First alarm added. Heat sources that Meteosat can see but a high-resolution satellite has not yet confirmed are now called out above the map. They are unconfirmed and are not counted as active fires; known industrial sources are excluded. Clicking a Meteosat ring now says "possible fire in this area" and explains what remains uncertain.
• News report layer added. For the small, short-lived fires satellites cannot see, news is now watched as a second channel. These are unverified reports: they are drawn as an approximate area rather than a point, and are not counted as active fires. Clicking a circle shows which report it rests on, how many outlets carried it, and whether the fire has been contained.
• The Bayramiç fire has been added to the archive — the only record there built entirely from Meteosat detections, because no other satellite left a trace of it.

2 August 2026 — English language support

• The whole platform is now available in English. The map, the fire panel, the province pages, the season statistics and the archive all have an English version under /en. Turkish addresses did not change: every link that was shared before still works. Place names stay Turkish in both languages — a name is an identity, and translating it would break the link between the map and what people say on the ground.

2 August 2026 — permanent heat sources, province pages, sharing

• We stopped counting industrial facilities as fires — our counter was inflated. A satellite sees heat, not flames; refineries, steel plants and power stations are hot every day. We were counting those as "active fires". We measured it: there are 50 places seen hot on more than 40 separate daysat the same point. For comparison, the longest measured forest fire in Türkiye lasted 16.5 days. Those points stay on the map but are now marked as a "fixed source" and left out of the fire count. That is why the figure in the header dropped — the old one was wrong, the new one is right.
• A separate page was opened for each of the 81 provinces. Every province page carries that province's season data: how many detections, how it compares with past seasons, when and where the peak heat was measured. Permanent heat sources are subtracted from those numbers — otherwise the figure in industrial regions comes out many times higher than the truth.
• Fires can now be shared. When you select a fire and send its link, the other side sees that fire directly; in chat apps the place name, radiative power and duration appear too. The numbers on the card are not written into the link — they are produced from real data each time, so that a fabricated card cannot be made.
• A smoke forecast was added. We show when the fine particulate matter (PM2.5) in the air at the fire's location will peak over the next 48 hours. We do not compute the dispersion ourselves; we relay the output of CAMS, Europe's atmosphere monitoring service.
• An archive of past fires was opened. The 2021 Manavgat, Marmaris and Milas fires can be replayed from beginning to end. The forecast shape is deliberately not drawn in the archive: we do not present a forecast that was never made that day as if it had been.
• Fires on agricultural land can be hidden, and a province / district search arrived. Detections whose land cover is agricultural drop out of the list and the map with one button. Events whose cover could not be queried are not hidden, and how many of them there are is written out — a silently cleaned map would be misleading.
• The reach shape is drawn at close zoom. At the country view the shape was only a few pixels; even when drawn it could not be read and just left a smudge. It now appears as you approach a fire, and the reason it is not shown is written out. The outermost 6-hour boundary was also made distinct — that was the line that most needed reading, yet it was the faintest.
• The area the satellite saw as hot is shown. In hectares. This is not an official burnt area: smouldering sections lose their heat signature, and a detected pixel may not have burned in full. We named it accordingly.

2 August 2026 — Cyprus and visuals

• Place names were set to Turkish and the TRNC boundary was added. Although this was a Turkish product, the map showed names in the local language — Greek in Cyprus, English for the country name. It now reads Lefkoşa, Girne, Gazimağusa, Larnaka, Baf and Türkiye. The TRNC land boundary is drawn as well; the island had looked like a single piece.
• The "growing" label was tested, failed, and its wording was corrected. The trend label on the fire card read like a forecast. We measured it over 820 advances: its discriminating power is a coin flip (AUC 0.502). A fire labelled "growing" does not advance more than one labelled "receding"; what is more, risen heat usually falls back. The label now says what it is: heat rose / heat fell — a description of the past, not a forecast of the future.
• The promotional screenshots were refreshed. The screenshots on the project page at algow.net were updated to this version; the new reach shape, the terrain layer and the honesty notices are visible. The previous version's images were not deleted and remain accessible in the archive.

2 August 2026 — audit

• Province boundaries are now complete. The basemap's province boundaries arrived broken and in some regions were not drawn at all. We now draw the boundaries from our own data (Natural Earth, public domain) — 81 provinces, complete at every zoom. To keep the first load quick, they are fetched after the map settles.
• Regression tests were added to the direction calculation. The direction of the shape is now a multi-layered computation; a single sign error in it could have reversed the shape without visibly breaking anything. 25 tests were added, and we verified that they really catch it by injecting a deliberate error.
• We tested our own claim, it came out wrong, and we fixed it. We had tuned the "90%" threshold of the rings on the same data we measured it with — that is, the claim was confirming itself. Holding each season out in turn and testing on it showed that the ring covered a fire's furthest point only 81% of the time. The rings were enlarged; with the new setting, coverage on unseen seasons is 90%.

2 August 2026 — map pass

• Topography was added. The "Terrain" button turns on hillshading: valleys, ridges and slope aspects become visible. Since half of fire behaviour is terrain, this also makes it readable why the shape leans the way it does.
• The satellite basemap was renewed, and a daily image arrived. The previous basemap was a mosaic that could be years old. In its place came Sentinel-2 cloudless (10 m, cloud-free). In addition, the "Today" button opens NASA GIBS daily true colour imagery — the resolution is coarse (250 m) but the date is today, so smoke from large fires can be seen.
• Doubled city names were fixed. The setting we used to show labels at an early zoom was opening city layers that should exclude one another at the same time; the same city was written twice.

2 August 2026 — second pass

• The reach shape was measured: it is a teardrop now, not a circle. We measured the question "how far did fires advance in directions deviating by so much from the forecast direction?". Fires travel 2.4 times further ahead than behind. A symmetric circle threw that information away; the shape now draws the measured envelope, and the arrow at its tip makes the lean readable.
• The rings are drawn with hourly forecast wind. Previously it was assumed that "the current wind holds for six hours". Each ring now uses the wind of its own hour; if the wind turns, the outer rings bend accordingly and the panel says by how many degrees.
• Land cover is shown. The CORINE land class of the selected fire is written out. Stubble burning on agricultural land is now marked as "not a forest fire".
• Satellite observation gap warning. Pass windows are measured from the data itself. If we are in a gap it says so in the top bar and in the panel: no detection does not mean the fire is out.

2 August 2026

• The size of the shape was measured and reduced. Until now the radius of the rings rested on no measurement and was 8–12 times the truth. We measured it with six seasons of data: we were drawing the 3-hour ring at 7.6 km, while 90% of fires had advanced less than 2.5 km from their own edge in that time. The rings were pulled back to that measurement.
• The angle of the shape reflects the real uncertainty. The old narrow 15–30° wedge covered only 30% of the observed deviations. The angle now comes from the measured distribution; when the wind is weak the direction is nearly undetermined, so a disc is drawn instead of a wedge.
• Terrain slope was added to the direction forecast. The direction is now the resultant of wind and slope. On steep ground the rate of completely reversed forecasts fell from 22% to 11%.
• The validation was redone and our old verdict corrected. This page used to say, on the basis of twenty-three samples, that "the forecast is no better than random". Repeating it with six seasons and 468 thousand detections showed that to be a small-sample error.

1 August 2026

• The Meteosat 15-minute layer was added; the observation gap fell from 5 hours to about 35 minutes.
• The fire weather index (FWI), smoke (PM2.5), terrain slope, burnt area and danger layers were added.
• Fire alerts near me were added — saved places never leave the device.
• The platform went live; the source code was opened.