Interactive radar view of the 29 October 2024 DANA flood in Valencia, Spain, from Sentinel-1 C-band RTC, plus Sen1Floods11 chips, NISAR L-band browse imagery, and an honest water-IoU table.
↔
29 Oct 2024 DANA
VV dB difference (post minus pre). Cyan is a decrease in backscatter, often standing water or a smoother surface. Amber is an increase. Same orbit, same grid, real pixels — not a model output.
lower VV (water-like) higher VV
Official Sen1Floods11 hand-labelled test chips. Images are real VV/VH (already in dB). The cyan overlay is the hand label; amber is the dataset's per-chip Otsu baseline; blue is JRC permanent water.
Hand-labelled water Otsu water JRC permanent water
These scenes are C-band Sentinel-1 (5.4 GHz, ~5.6 cm wavelength) from 2024. NISAR L-band (~1.25 GHz, ~24 cm) was not yet imaging Valencia during the 29 Oct 2024 flood. The NISAR browses below are later (2025–2026) public previews of the same region, used here to explain the physics we would apply to a Saturday .h5 granule — not as images of the DANA event itself.
VV, VH, and the VV/VH ratio
A SAR instrument sends a microwave pulse and listens for the echo (backscatter). Polarization names the electric-field direction: VV is vertical transmit and vertical receive; VH is vertical transmit and horizontal receive (cross-pol).
Smooth open water acts like a mirror. The pulse reflects away from the satellite (specular reflection), so VV backscatter drops and water looks dark. Cross-pol (VH) is already weak over smooth water, so the VV/VH ratio (VV minus VH in dB) helps separate wind-roughened water, wet soil, and vegetation.
The false-color product in the viewer is exactly that: red = VV, green = VH, blue = VV/VH, stretched with a shared 2–98 percentile per orbit so the time-lapse does not flicker.
Dielectric constant and roughness
Radar is most sensitive to two things: surface roughness relative to the wavelength, and the dielectric constant. Liquid water has a very high dielectric constant, so wet soil and inundation change the return even when the geometry stays the same. That is why a dark VV patch after 29 October is physically meaningful, not just a contrast stretch.
Why NISAR L-band can see flood under a canopy
C-band (Sentinel-1) interacts strongly with leaves and small branches. In orchards, marshes, and the Albufera wetland, C-band often never reaches the water under the vegetation, so a flood can be missed. L-band (NISAR) is about four times longer. It penetrates deeper into the canopy and can return a double-bounce signature from trunk and water — a brighter, not darker, response in inundated forest.
That contrast — C-band dark open water versus L-band canopy-penetrating inundation — is the story this challenge is built for. Saturday's Earthdata .h5 hook is how we put real NISAR polarizations on the same map.
What this viewer is (and is not)
Every Valencia frame is a real Sentinel-1 RTC GeoTIFF, downsampled for the web, with VV, VH, and VV/VH rendered from the same pixels. Before/after pairs are always the same orbit and the same grid. The change map is a VV dB difference, not a neural-network mask.
The U-Net numbers in the Metrics tab were measured on Sen1Floods11, not on these Valencia tiles. Valencia is the storytelling AOI; Sen1Floods11 is the labelled test set. We do not pretend otherwise.
Test water IoU on Sen1Floods11 (official hand-labelled test split, 90 chips)
Channel ablation (same U-Net recipe, test water IoU)
Adding VH, then the VV/VH ratio, is the only change between these three arms. The ratio arm is the strongest of the three on this seed.
CPU inference on a 512-pixel tile (Colab CPU, not an edge device)
Saturday hooks
The trained ONNX weights and any NISAR .h5 granules are not in this tree (the free T4 runtime ended before the export was saved; .h5 files need an Earthdata login). The CLI hooks are ready:
Tiled 512 px inference with 64 px cosine blending, CPU only, no credentials in the repo.
Public NISAR GCOV browse PNGs and CMR metadata, staged locally. These are L-band previews over Valencia, Toronto, and Jasper from 2025–2026 — later than the DANA flood. Full geocoded polarimetry lives in the companion .h5 granules (typically 1.5–7 GB) at the ASF DAAC and is not shipped here.
Data and software
Original software in this folder is licensed under Apache License 2.0. NASA content remains in the public domain. Sen1Floods11 is CC-BY-4.0. Builds on our earlier code and open-source libraries.