How it was done, step by step

1. Sentinel-2 picturesSentinel-1 radarMERIT Hydro terrain
   1Put everything on one 20 m grid

   Before and after pictures from the same season, radar and terrain, square by square.
2. 2Measure five things in every square

   Lost green, colour change, bareness, radar change, height above the river.
3. 3Make a first guess at the rule

   Settings picked by hand, without the colour-change test.

   17.7 km², F1 0.61–0.79
4. Copernicus damage outlineLangtang valley (untouched)
   4Try all 1,008 combinations of settings

   Score each one against the experts' map, and count its false alarms in an untouched valley.
5. 5Choose the settings fairly

   Drop any that flag more than 2 % of the untouched valley, test on areas held out, keep the best.

   The chosen settings
6. 6Make the final map

   Run the chosen rule along the whole corridor, tile by tile.

   Footprint 22.0 km²
7. Copernicus damage outline
   7Check it against Copernicus

   Lay the final map over the experts' map, square by square, and score the match.

   F1 0.83–0.89right 93 %–98 % of the time
8. Copernicus graded assetsOpen BuildingsOpenStreetMap
   8Count what lay inside

   Lay buildings, roads, bridges and power plants over the footprint.

   92 % of destroyed buildings30 of 30 bridges
9. Video, gauges, SMS timeMERIT river profile
   9Time the flood front

   Fit a simple flood-travel rule through the few real times, 10,000 times over.

   Syabrubesi 08:4934–41 min with an automatic alert

•   chosen in step 5
•   the first guess of step 3

How each version was scored

1. aNo crying wolf. Any version that marked more than 2 % of the untouched Langtang valley floor was thrown out.The best version without this rule (ΔNDVI below −0.05, d above 0.03) scored well against Copernicus but marked 5.3 % of the untouched valley, so it was rejected.
2. bTest on an area it never saw. Settings were chosen on two areas and scored on the third, three times over, so each area was once the unseen test.Leave-one-area-out cross-validation. If the method only memorised the answers, the score on the unseen area would drop. It did not.
3. cKeep the best of what is left. Among the survivors, the version with the highest average F1 over all three areas became the final rule.Final: ΔNDVI below −0.10, d above 0.08, NDVI after below 0.35, |ΔVH| above 4 dB, HAND below 150 m; it marks 1.5 % of the control valley.
4. dCheck that it holds. The three tests chose nearly the same settings, and the unseen-area scores (0.90, 0.84, 0.84) match the normal ones. The rule works on ground it was not tuned on.Two of the three folds chose exactly the final settings; the Bidur fold differed only in ΔNDVI and NDVI after.

Three pieces

Four scores, one picture

1. PPrecision: right when we say damaged. Both ÷ ours. At Syabrubesi: 0.90 ÷ 0.92 = 0.98.Of the ground we marked, the share the experts marked too. Low precision would mean false alarms.
2. RRecall: how much of their damage we found. Both ÷ Copernicus. At Syabrubesi: 0.90 ÷ 1.11 = 0.81.Of the ground the experts marked, the share we found. Low recall means misses.
3. F1F1: the overall match. Both ÷ the average size of the two maps. At Syabrubesi: 0.90 ÷ 1.02 = 0.89.2 × P × R ÷ (P + R). High only when precision and recall are both high, so marking the whole valley, which finds everything, would not fool it.
4. IoUIoU: the strict overlap. Both ÷ everything either map marked. At Syabrubesi: 0.90 ÷ 1.13 = 0.80.Intersection over union. Every false alarm and every miss counts against it in full, so it is always lower than F1 (IoU = F1 ÷ (2 − F1)).

Did it hold up on ground it had never seen?

What the check says

1. 1Careful. When our map says damaged, the experts agree 93 % to 98 % of the time. False alarms are rare, and the untouched control valley stays 1.5 % marked.
2. 2A little small. It finds 73 % to 81 % of what the experts drew. The misses lie mostly at the edges, most likely strips thinner than our 20 m squares, ground that changed only faintly, and gorges radar cannot see into.
3. 3Not tuned to the answers. Areas hidden while choosing the settings score as well as the rest, so the result is not an artefact of fitting.
4. 4Checked only where Copernicus looked. Three areas and one control valley. Copernicus is a human drawing too, so a few “misses” or “extras” may be theirs rather than ours.

1. BBuildings count as inside when the 20 m square under them is marked.Copernicus graded building points (3,922) and Google Open Buildings centroids with confidence 0.75 or more.
2. RRoads are measured by the length that runs through marked squares.Each line is sampled every 5 m; Copernicus roads graded destroyed and OpenStreetMap roads by type.
3. BrBridges count when they are within 40 m of the footprint, since they cross the river at its edge.Copernicus bridges graded destroyed; OpenStreetMap road and foot bridges.
4. HHydropower sites count when they are within 100 m of the footprint.OpenStreetMap power plants, generators, dams and weirs.

What we do know

Filling in the places in between

What “10,000 times” means

How k is found

Why it matters

python -m rasuwa export      # 1, 6  the footprint, tile by tile
python -m rasuwa tune        # 4, 5  the 1,008-version sweep and the fair choice
python -m rasuwa validate    # 7     the check against Copernicus
python -m rasuwa buildings   # 8
python -m rasuwa infrastructure
python -m rasuwa leadtime    # 9