Is the signal the storm?
A radar change map can light up for reasons that have nothing to do with wind: rain, a wet season, a field harvested. So NISAR's change across the storm was set against NISAR's change over two 12-day windows with no storm in them.
AUC measures how well a layer separates damaged from intact forest; 0.5 is no better than chance and 1.0 is perfect.
How well NISAR HV change on one ascending track separates damaged from intact forest (0.5 is chance), for the 12 days across the storm and for two 12-day windows with no storm, with the held-out areas left out. Upland means blocks that are at least 90 percent pine or hardwood on ground more than 5 m above the nearest stream.
Across the storm, NISAR HV scores 0.82 at 100 m resolution. The two no-storm windows score 0.57 and 0.58, so roughly three quarters of the separation above chance comes from the storm. On upland forest, where floodwater cannot pool, it still scores 0.72 against 0.52 and 0.58. NISAR's HH channel carries no detectable storm signal on upland forest. These numbers leave out the held-out stretch and the aerial strip; with them included, the earlier analysis gave 0.83 against 0.54 and 0.57.
Against Sentinel-1 with images matched in time, NISAR's ascending track does better at 100 m (0.84 against 0.80, held-out areas excluded) but the two NISAR tracks averaged together only tie, because the descending track is weak over pine.
Downed trees or floodwater?
Isaias dropped up to 11 inches of rain. In bottomland forest, standing water under the trees also brightens radar, and it would look a lot like damage. Water drains; fallen trees stay down. So the test is whether the signal holds over time.
Mean radar change in damaged forest minus intact forest at 30 m, by date after the storm, with the held-out areas left out. The gray band is NISAR HV over the two no-storm windows.
On upland forest, NISAR HV fades only slightly from 17 to 29 October (+0.51 to +0.44 dB), far above the no-storm windows, which looks more like structure than water. In woody wetland, both NISAR channels grow instead of fading, led by HH (+0.94 to +1.07 dB), which is also what lasting floodwater under the canopy would do. There, downed trees and water cannot be told apart from these data. Sentinel-1 fades more on upland (+0.36 to +0.18 dB) than in wetland (+0.97 to +0.86 dB), the opposite of a simple drying pattern. Held-out areas are excluded throughout, and NISAR's 29 October scene comes from a newer processor (see Methods).
Wind and damage
A northward-moving hurricane usually does its worst damage right of the track, near landfall. Isaias did the opposite: the damage sits 45 to 75 km west of the track, on its left. East of the track the satellite map and both radars show very little, although NWS Mobile reported widespread tree damage there in Covington County.
Share of forest in the Sentinel-2 damage mask by distance from the NHC best track, in 5 km bins, on both sides of the track, with the held-out areas left out. Triangles are station peak gusts; hollow ones reported fewer than ten times during the storm and may have failed.
West of the track, 15 percent of forest 40 to 80 km out is in the damage mask, up to 45 percent of woody wetland. East of the track, mapped with the same method out to 75 km, it is 0.4 percent within 40 km, and under 1 percent of upland forest at every distance; the only rise is in what looks like a river floodplain at the far eastern edge, where flooding could also be the cause. The band is not an artifact of how the images were taken. Part of its eastern edge runs close to a satellite tile boundary, but the change does not jump across that boundary where the band is absent, the band survives a rebuild from a single clear image before and after, and both radars put it in the same place. The radars, which do not depend on the optical mask, also show the change west of the track and not east of it.
Coastal gusts near Pensacola
Some of the strongest gusts were measured near the coast around Pensacola: 101 kt at Oriole Beach on the Gulf Breeze peninsula and 80 kt at Pensacola airport, 48 to 58 km west of the track. Media coverage of Isaias discussed a sting jet as a possible cause of its unusual winds: the WALA FOX10 summary of the NWS report attributed the unusual inland winds to a sting jet that set up as the storm transitioned. Asked by WEAR-TV whether there was a sting jet, NHC Deputy Director Jamie Rhome said: "It remains to be seen." He added that strong winds on the west side are not by themselves necessarily a sting jet. NHC has not confirmed one.
The inland swath
The damage band itself lies inland, running some 120 km north from about 30.5°N in Santa Rosa County into south Alabama. It may be a separate feature from the coastal gusts. At the stations nearest it, Whiting Field South’s 99 kt gust and Jay’s highest 15-minute wind (the time of Jay’s 102 kt gust is not public) came from the north-northeast to northeast about one to one and a half hours before the center passed, while Isaias was still a hurricane, and before dry air on satellite water vapor reached the band. At the time, NHC noted damaging gusts "in a band wrapping around the western side of the circulation." That timing fits the outer western wind field or a mesovortex better than a sting jet, which usually arrives with the dry air as a storm loses its tropical character. None of these is established here, and NHC's Tropical Cyclone Report for Isaias is still pending.
Sources: NHC advisory archive (Discussion 14); NWS Mobile report and county impact narratives; WALA FOX10 summary (14 October); WEAR-TV (20 October); KMOB and KEVX radar from the NEXRAD Level II archive; GOES-19 water vapor from NOAA's open data.
Results
Five ways of mapping the damage are compared on a sealed 20 km stretch of the band: simple thresholds on NISAR (A) and Sentinel-1 (B), and models trained on NISAR alone (C), Sentinel-1 alone (D), and both (E). The models come in two versions: v1 as first designed, and v2 after a documented change to the training labels (see Methods). All were frozen and their files hashed before scoring.
| Arm | Upland AUC | Wetland AUC | F1, damaged |
|---|---|---|---|
| A NISAR HV threshold | Scored after held-out labeling | ||
| B Sentinel-1 VH threshold | Scored after held-out labeling | ||
| C NISAR model, v1 | Scored after held-out labeling | ||
| C NISAR model, v2 | Scored after held-out labeling | ||
| D Sentinel-1 model, v1 | Scored after held-out labeling | ||
| D Sentinel-1 model, v2 | Scored after held-out labeling | ||
| E both sensors, v1 | Scored after held-out labeling | ||
| E both sensors, v2 | Scored after held-out labeling | ||
Until then, the only scores are on validation tiles inside the training region, and they measure agreement with the Sentinel-2 mask, not with hand labels. On the validation cut that mirrors the held-out sample (inside the band, damaged forest against all other forest), the v2 U-Net models scored 0.82 on upland and 0.82 on wetland with both radars, 0.76 and 0.80 with NISAR alone, and 0.78 and 0.76 with Sentinel-1 alone. The simple thresholds scored 0.54 to 0.60. These v2 models were tuned on that same cut, so the numbers are somewhat optimistic.
550 cells, 400 in the held-out stretch and 150 under the NOAA aerial survey, are waiting to be labeled by hand from before-and-after imagery. The scores above will be filled in once, all of them, including any that go against NISAR.
Methods and honesty notes
The reference is optical and provisional
The damage mask comes from Sentinel-2 greenness and burn-ratio change against the same weeks of 2025, with clearcuts and salvage flagged out. There is no field or aerial truth inside the damage band yet, and the hand labels are drawn from the same Sentinel-2 imagery, so agreement with them is partly agreement with Sentinel-2.
The mask sees stands, not single trees
The damage mask detects stand-level damage: a 10 m pixel has to lose greenness and burn ratio well beyond the same weeks of 2025. Scattered single trees down along roads and power lines, such as those NWS Mobile reported in Covington County, Alabama, may not show up in it.
Validation scores were inflated
The first models scored 0.97 to 0.9996 AUC on validation tiles, but the training labels pitted clear damage against clean intact forest and left out the cells in between. On the cut that mirrors the held-out sample, inside the band, damage against all other forest, they scored 0.63 to 0.78, and the simple thresholds 0.55 to 0.61.
Amendment 1 changed only the labels
About 93 to 95 percent of the out-of-mask cells in the hand-label sample were cell types the first models never saw. Before any model was scored on the held-out area and before labeling began, a written amendment added v2 models trained on every forest cell. v1 stays frozen; both will be scored side by side.
Processor versions are flagged
NISAR's newest scene (29 October) comes from a newer processor and sits about 0.6 dB lower in HV and 1 dB in HH. The 2025 no-storm scenes come from an older one and sit about 2 dB lower in HV. Each layer's forest-wide median is removed, which takes out uniform offsets but not uneven ones.
Not verified
The Jay gust against raw FAWN data (the public file has no gusts); the location and data of the Oriole Beach StickNet; why the worst damage fell west of the track; floodwater against structure in woody wetland; and the size of the selection advantage v2 gained by being tuned on the in-band cut.