Results
Floods under forest, seen by NISAR: every test, including the failures
Full abstract
Figures and tables
Hover or tap any chart for its values, and open a figure's details for what each panel shows. Every figure also downloads as SVG, PNG or the data it plots.
Table 1 Radar against a lidar map
How GLOWfm compares on the latest test's passes with a lidar-only map, which simply calls land under 2 m above the river flooded.
Details
The lidar map uses no radar or gauges, and since the reference it's scored against is built from that same lidar, the comparison favors it.
Table 2 Without lidar
Most of the world has no lidar, so two more frozen tests swapped in global elevation data, first treetop heights (Copernicus) and then bare earth (FABDEM).
Details
Both ran on the seven judged passes and were scored against the same gauge-and-lidar reference, with each column giving NISAR's score for one source of ground height.
More figures: study design, how soon NISAR saw floods, the sealed test, water level under the canopy
Table 3 A test sealed before the passes were acquired
Details
Methods, data and code
Methods
Data
We used NISAR's Level-2 GCOV backscatter (HH and HV, provisional release P05023, 40 MHz scenes) on a 50 m grid, comparing each flood pass with earlier passes on the same track. Land cover comes from ESA WorldCover 2021, ground height from USGS 3DEP lidar (or Copernicus GLO-30 and FABDEM in the tests without it), river levels from USGS gauges, rainfall from IMERG, and the official flood map from Copernicus GFM.
GLOWfm's method (version 4, frozen)
For every pixel, GLOWfm takes the median of each same-track scene within 60 days as its dry baseline and measures how much the HH backscatter changed, smoothed over 3×3 pixels. It switches to HH minus HV when higher valley forest brightened by more than 0.5 dB, as rain on the canopy does. It only looks at forest, wetland and other land within 6 m above the river and 4 km of the valley, where it splits the signal with Otsu's method at no less than 0.5 dB, so each scene sets its own cut-off without any reference data. It holds back wherever the nearest gauge is more than six days past its crest.
Trust light
The light turns green when low valley forest (under 2 m above the river) brightened more than higher forest by over 0.5 dB and the pass wasn't rainy, and red otherwise. It was fixed before its first fair test, where it was right on 3 of 4 judgeable passes (), which is too few to prove it.
The reference ("the river's answer")
Gauge levels at the time of the pass are interpolated along the valley and compared with lidar ground height, so a pixel counts as flooded where the water stands more than 0.5 m above the ground and dry where the ground sits 0.5–5 m above the water. It's kept to a 4 km corridor within 10 km of a gauge and outside permanent water, which makes it an estimate rather than an observation, and tests it. Two checks against photos, fixed in advance, supported it where the ground was visible on the Satilla (water at 9 of 14 flooded points and 0 of 11 dry), but under closed canopy photos can't tell water from shadow, so mapping under trees is checked against this reference alone.
Scoring
CSI is hits divided by hits, misses and false alarms, counted on mapped, gauge-supported pixels outside permanent water, with 95% intervals from a 2 km block bootstrap (2,000 resamples) and paired resamples for differences between maps. A pass is judged when the reference holds at least 500 ha of flooding, and recomputing the intervals with 10 km blocks widens most of them ().
Frozen one-time tests
Before each test, its protocol, scene list, site files, run code and scientific modules were hashed into a receipt, and the run refuses to start if any of them changed or if it has already run once. Each protocol set out in advance what would count as success or failure, and any fix made after scoring is labeled post hoc without replacing the first number.
Without lidar
Two frozen tests swapped lidar for global elevation data, first Copernicus GLO-30, which includes the treetops, and then FABDEM, which removes them, keeping everything else in version 4 and scoring the same pixels against the same reference. The bare-earth test checked in advance that any gain didn't just come from mapping more ground, and as it predicted, flooded forest on the Satilla sat a median 0.39 m above the river by FABDEM, against 4.92 m by Copernicus and −0.39 m by lidar.
Coverage and delay ()
For every flood GDACS reported over 60 days, we searched NASA's catalog for NISAR scenes covering the reported point while the flood was being reported, keeping each scene's recording time and the date it reached the catalog (scripts/build_coverage.py).
Water level under the canopy ()
This is an exploratory check, made after the frozen tests and not scored. NISAR's GUNW product compares the radar phase of two passes 12 days apart, and where flooded swamp stays coherent, a phase change is a change in water level of about 16 cm per cycle. We read a dry-season control and two pairs spanning Hurricane Isaias over the Choctawhatchee delta, adding up the phase steps from west to east and setting them to zero at the Bunker gauge (scripts/posthoc_swamp_fringes.py).
Descriptive figures
and were computed after the tests from the saved pixel layers of the twelve frozen version-4 runs, and the build checks them against the frozen records so nothing is re-scored or re-tuned. ROC curves (area under the curve 0.54–0.91) and agreement by depth, height and land cover come from the same layers, and their data is in data/figures/separability.json and strata.json.
Data and code availability
All code is open source (MIT), and the rules, results and hash receipts are in the repository's research/ folder, starting with OVERBANK_RESULTS.md, BARE_EARTH_RESULTS.md, BLIND_SPOT_AUDIT.md and BLIND_SPOT_ATLAS.md. FABDEM is CC BY-NC-SA 4.0 (Hawker et al. 2022, Environ. Res. Lett. 17, 024016), so no FABDEM layer is included and only its scores are shown. The figure data is built by scripts/build_figures.py, which checks it against the frozen records, and every figure downloads with the data it plots.