fidopt Interactive view
Optical to SAR (2D), same place

Optical to SAR (SpaceNet 9, BRIGHT)

Optical (fixed) and radar (SAR, moving) satellite images of one town after an earthquake, explosion or eruption, at 0.3 to 0.45 m a pixel, with point pairs placed by people on the ground.

6 test pairsearthsame patientfixed OPTICAL, moving SAR2D, distances in mdelivered alignedresearch only
Look at the images
Read before comparing (12)
  • Modalities. A colour photograph of the ground against the strength of radar echoes, with speckle: the same ground with unrelated intensities, as a CT and an MR. A case's 'same patient' is the same place.
  • Layover. Raised objects shift towards the radar and away from the optical camera, so roofs never meet under a smooth transform; the ground does. The points lie on ground features seen in both, so buildings, slopes in layover and ground the disaster changed are not measured.
  • SpaceNet 9. WorldView-2 colour and Umbra SAR images of Türkiye after the 2023 earthquakes: gaziantep01 (151 points), gaziantep02 (104) and adiyaman01 (161); its two public test pairs come without points. The two Gaziantep sections share no ground, 806 m apart, but come from the same two acquisitions.
  • BRIGHT. Pre-event optical and post-event SAR images, before BRIGHT's own registration: Noto 2024 (Umbra SAR, 64 points), Bata 2021 (Capella SAR, 255 points) and La Palma 2021 (Capella SAR, 308 points in two groups with none across the middle).
  • Truth. SpaceNet 9's team placed its points in QGIS, labelling then review. BRIGHT's are the control points its Earth-observation experts placed and cross-checked to register each pair, which its authors call a proxy rather than absolute truth. Neither says how far apart two people would place one point.
  • Consistency. A thin-plate spline through a case's other points predicts each point's displacement to within a median of 2.5 to 4.2 m a case (Bata 2.5, the two Gaziantep pairs 3.5, La Palma 4.0, Noto 4.1, Adıyaman 4.2): an upper bound on the points' own error.
  • Alignment. Each image comes placed on the map by its provider, kept as delivered: no transform is computed. As delivered the points are 32.2, 26.1 and 31.9 m apart on gaziantep01, gaziantep02 and adiyaman01, 17.0, 8.1 and 9.5 m on noto, bata and lapalma. An elastix translation from the intensities left bata at 583 m and was not used.
  • Baseline. With no registration the six test pairs score a TRE of 20.8 m (8.1 to 32.2), rTRE 0.0049 (0.0009 to 0.0099), well above each pair's consistency bound.
  • Borders. Outside the acquisition the images hold 0, as does some dark SAR ground: 31 to 49% of the optical images other than Gaziantep's, 22 to 67% of the SAR images. No mask is given, so a method sees these borders as image.
  • Size. Up to 29971 pixels a side. A method that reads no OME-Zarr gets each image at its finest level within 2^24 pixels; the landmarks are scored at full resolution.
  • Ranked apart. It misses the truth bar with six test pairs, where a main board needs ten, and passes the two other checks.
  • Outside numbers. None compare: SpaceNet 9 ranks its own test pairs, and BRIGHT starts from the pair as delivered with methods at half resolution.

No method has a result on every pair of this dataset yet.

AccuracySmoothnessCost
#MethodTRE mChangerTRESDlogJFolds %Time sRAM GBGPU GBImages
No registration20.78 ±11.330.0049 ±0.00420.0000.00

TRE: distance between matching landmarks. rTRE: the TRE over the length of the image's diagonal, to compare datasets of other sizes and units. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.

Pair by pair

TRE m123456
No registration31.878.1432.1526.089.4616.97

Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.

About the data

Test pairs
All six pairs are test pairs: the three training pairs of the SpaceNet 9 challenge and the three of BRIGHT's image matching benchmark, the only pairs of the two whose points are public, one a place. None is left for development, since it would share a place with a test case.
Truth
landmarks by experts the providers' point pairs, placed by people on ground features seen in both images: SpaceNet 9's tie points (placed in QGIS by the SpaceNet team in two rounds, labelling then review, buildings avoided) and BRIGHT's control points (selected in QGIS by Earth-observation experts and cross-checked), 64 to 308 a pair
Licence
CC BY-NC 4.0