Every term the analysis uses, in plain words, grouped by topic.
A.1 Events and gold maps
term
meaning
gold map
an expert-drawn flood map from Copernicus EMS or UNOSAT, used as the reference answer
activation code
the ID a mapping service gives one flood response, e.g. EMSR838. One code is one event
examined area
the part of a gold map the analyst actually looked at. Outside it there is no label, flooded or dry
map day
a date on which an event, or one of its windows, has a gold map. A window is scored only on its own map days
event period
from an event’s first map day to its last
window
one area of an event, processed on its own. Maps within 0.05° of each other share a window. A window covers 30 days before the event period to 60 days after it
days before and after
the 30 days before the first map and the 60 after the last. Examined cells are labelled dry there, to measure false alarms on ordinary days, except cells flooded on a map day of their window, which get no label there
built
an event whose gold labels and source values have been prepared on the grid and saved (an event cube per window)
A.2 Grid and cells
term
meaning
cell
one 0.01° × 0.01° square of the grid, about 1.1 km across at the equator
cell-day
one cell on one day: the unit every count is made of. A window-day is one window on one day
sub-pixel
one of the 100 squares of 0.001° a gold map is drawn onto inside each cell
HAND
height above nearest drainage: how far a point sits above the nearest stream. From the GLO-30 elevation model
floodable
at least 30% of the cell lies less than 15 m above the nearest stream
permanent water
water at least 90% of the time over 1984–2021 (JRC Global Surface Water). Cells that are mostly permanent water are removed
domain
floodable cells that are not permanent water. Nothing outside it is predicted or scored
normally wet
a cell where at least 30% of the 30 m pixels were water in that calendar month in at least half of the years observed (JRC monthly recurrence, 1984–2021): a wet-season floodplain, wetland or rice paddy
reference water
a map of the water that is normally there, which a product or an analyst leaves out of “flood”. GFM uses its own; the gold analysts decide by eye, and do not always decide the same way
unusual water
water that is not normally there in that month: the flood every source agrees on
water-first design
a proposal: fuse the products’ water, then turn it into flood with one normal-water baseline, so the definition of flood is one visible setting
land type, stratum
open, vegetated or urban, from ESA WorldCover. A stratum is the group of cells with one land type; the model’s rates are fitted per stratum
snow season
a month in which a cell is usually snow-covered (ERA5-Land). Not used: almost all snow-season gold flood is in Europe, which is out of scope
A.3 Satellites and sources
term
meaning
radar (SAR)
the satellite sends a microwave pulse and reads the echo; calm water reflects it away and looks dark. Works through cloud and at night, but cannot see water under trees
Sentinel-1, GFM
the European radar satellites, and Copernicus’s Global Flood Monitoring product made from them. A given place gets an image only every 6 to 12 days
optical
sees reflected sunlight, like a camera. Blocked by cloud
VIIRS, MODIS
US optical sensors that image the whole Earth daily; NOAA and NASA make flood maps from them
passive microwave
measures the energy the ground itself gives off; wet ground gives off less. Sees through cloud every day, but only at about 9 km
FloodScan, SFED
AER’s product from passive microwave; SFED is the share of each 9 km cell that is flooded
GFDS
JRC/GDACS Global Flood Detection System, also passive microwave, about 9 km
couldn’t see
the product had no usable observation for that cell that day (also called NoData). Never counted as dry
microwave vote
FloodScan and GFDS combined: flood if either says flood, dry if at least one observed and neither says flood
optical vote
the same for VIIRS and MODIS (proposed)
CYGNSS
a NASA constellation measuring GPS signals reflected off the ground; the UC Berkeley daily watermask made from it is a rolling multi-week composite of surface water at 0.01°, within 37.4° of the equator, from August 2018
A.4 Fusion
term
meaning
hit rate
for one source and land type: the share of gold-flooded cells it observed that it called flood (docs/methods.md: sensitivity)
false-alarm rate
the share of gold-dry cells it observed that it called flood (1 − specificity)
fitting
counting hits, misses, false alarms and correct dries on the events used for fitting, to get those two rates
base flood rate (prior)
how often a cell of that land type is flooded on an ordinary day, from the global random sample’s 30 blocks in the four continents: about 0.23% in open land, 0.01% in vegetated and 0.005% in urban land
odds
P / (1 − P). P 0.002 is odds of about 1 to 500. Fusion multiplies the odds by each source’s hit rate over its false-alarm rate
log-odds
the logarithm of the odds; fusion adds evidence and fades yesterday’s value on this scale
P
fusion’s probability that a cell is flooded that day; P ≥ 0.5 counts as a flood call
memory
yesterday’s log-odds carry over at 0.9 of their distance from the base rate, so evidence builds up over days
cascade
a baseline that needs no fitting: for each cell-day, the first product that observed it, in the order VIIRS, GFM, microwave vote
A.5 Scoring and testing
term
meaning
hit, miss, false alarm, correct dry
product flood where gold is flooded; product dry (or couldn’t see) where gold is flooded; product flood where gold is dry; product dry where gold is dry
recall
the share of gold-flooded cells a product flags
precision
the share of a product’s flood calls that gold confirms
F1
the balance of recall and precision (their harmonic mean); high only when both are. The headline score
CSI
hits ÷ (hits + misses + false alarms); ranks products the same as F1. Used only for criterion (b)
all cells, observed cells
two scoring views: a product’s gaps count as misses / only the cells it saw are scored
pooled
counts summed over all cells, days and events before computing a score, so large floods weigh more (a micro-average)
basin-weighted
each river basin’s counts scaled so its flood counts once; stops one basin deciding the answer
reliability, calibration slope
whether probabilities mean what they say: of cells given P 0.6, about 60% should be flooded. Slope 1 is perfect; the target is 0.8–1.2
fold
one of five groups of events; each is scored with rates fitted on the other four
training, test (held-out) events
the first run’s one-off split: 149 events for fitting, 49 for testing
river basin
a river’s catchment, from HydroBASINS at level 4 (level 1 is a whole continent; each level splits basins into smaller sub-basins). Events in one basin stay in one fold
bootstrap interval
the range a score moves within when it is recomputed on many random re-draws of whole events; it shows how much the score depends on which events happened to be mapped
global random sample
53 blocks of 0.25° to 0.5° with no gold maps, with every day of 2023–2025 prepared: 30 random, 12 on cities and 11 reserves. The 30 blocks in Africa, Asia, South America and Oceania give the base flood rate; ordinary-day precision will be measured there
caught
a product flags at least 10% of an event’s gold-flooded floodable cells, and at least 5, on some day from 3 days before its first map to 3 days after its last. “Caught, no false event” also requires no false event
false event
the “caught” rule firing on ordinary days, when there was no mapped flood
A.6 Files and tools
term
meaning
STAC
a searchable catalogue of satellite files (GFM’s is at EODC)
COG
a cloud-optimised GeoTIFF: a raster file that can be read in parts over the internet
NetCDF
a file format for gridded data with several layers and a time axis; each event cube is one
blob
a file in the team’s Azure storage (dev or prod account)