遇见数据集

LB-DLPU: An L-Band (NISAR/UAVSAR) Benchmark for InSAR Phase Unwrapping

收藏
Zenodo2026-08-10 更新2026-08-13 收录
官方服务:

资源简介:

LB-DLPU is a physically-simulated benchmark for L-band interferometric SAR (InSAR) phase unwrapping, calibrated to the NISAR (spaceborne, 20 m) and UAVSAR (airborne, 6 m) regimes. It contains 10,000 patches (256×256), split 8,000 / 1,000 / 1,000 (train / validation / test) by whole DEM tile so no terrain is shared across splits. Each patch is a gzip-compressed HDF5 file holding the wrapped phase (input), the ground-truth absolute phase (target), coherence, per-edge five-arc integer ambiguity labels (kx, ky ∈ {−2,…,+2}), the Goldstein loop-residue map, and a validity mask, plus per-patch attributes (sensor, difficulty, mean coherence, pixel spacing, looks, residue count, clean-oracle RMSE, and component provenance). A wrapped-phase PNG quicklook accompanies every patch. Two properties distinguish it from existing (C-band, RMSE-only) PU datasets. (1) Well-posedness certificate: every scene is provably recoverable — a noiseless-oracle minimum-cost-flow unwrapper reconstructs each patch to within 0.035 rad given the correct per-edge costs (100% of 10,000 scenes pass), so any error a method incurs is attributable to the method, not to an unsolvable target. (2) Sensor calibration: coherence (Beta fits), Cramér–Rao phase noise, and an ionospheric screen (NISAR, at the measured f^−2.4 spectral slope) are calibrated to real NISAR L2 GUNW and UAVSAR interferograms; the synthetic residue density brackets the real measurement in both regimes. Topography comes from 29 Copernicus GLO-30 DEM tiles spanning tectonic, volcanic, glacial, vegetated, and arid terrain worldwide, assigned whole to the splits (17 / 5 / 7). Intended use: training and benchmarking L-band phase-unwrapping methods, deep networks, and classical/minim-cost-flow solvers, with per-regime × difficulty evaluation. See README.md and MANIFEST.txt in this record for the full HDF5 schema, directory layout, loading examples, and citation. Mirrored on HuggingFace: https://huggingface.co/datasets/TheDagbanja/L-Band_DLPU

提供机构:
Zenodo
创建时间:
2026-08-03
二维码
社区交流群
二维码
科研交流群
商业服务