12 From counts to totals

Every score on this page is built from the same counts, in these steps:

One count per labelled cell, per day, per product: hit, miss, false alarm or correct dry. Every product, and fusion, is scored on the same labelled cells.
Per event. Counts are summed over each window’s cells and days, then over the event’s windows, and kept per event, land type and kind of day. Every number below is a sum of these.
Fitting, per fold. The counts of the four folds not being scored, per land type, give each source’s hit rate (map days, the only days with flood labels) and false-alarm rate (map days plus every day before and after), counted on floodable cells only, where fusion makes its calls.
Scoring. Each fold’s events are scored with the rates fitted on the other four folds. The map-day counts of all 170 events, each tested once, are summed, and recall, precision and F1 are computed from the sums: overall, per continent, per land type and per product.
Pooled, not averaged. A large flood has more cell-days, so it weighs more; five events hold 47% of all gold flooded cell-days, and one basin’s four Bangladesh floods 40% (named under Testing). That is why results are also shown per continent and land type, with intervals from resampling whole events (a bootstrap: the scores are recomputed on many random re-draws of the events, and their spread shows how much they depend on which events happened to be mapped).
Event level. One yes or no per event (caught or not) and one number (days to detection), reported as shares of events. There every event counts equally.