TheDagbanja/L-Band_DLPU

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

[image]

LB-DLPU is a physically-simulated benchmark for interferometric SAR (InSAR) phase unwrapping at L-band, calibrated to the NISAR (spaceborne, 20 m) and UAVSAR (airborne, 6 m) regimes. Each of the 10,000 patches ships with the wrapped phase, the ground-truth absolute phase, coherence, per-edge integer ambiguity labels, residues, and a validity mask — everything needed to train, validate, and test learning-based and classical unwrappers.

Two properties set it apart from existing (C-band, RMSE-only) PU datasets:

• —Well-posedness certificate. Every scene is provably recoverable: a noiseless-oracle minimum-cost-flow (MCF) unwrapper reconstructs each patch to within 0.035 rad given the correct per-edge costs (100% of 10,000 scenes pass; max clean-oracle RMSE = 0.0347 rad). Any error a method incurs is attributable to the method, not to an unsolvable target.
• —L-band-specific difficulty. Calibrated coherence (Beta fits to real granules), Cramér–Rao phase noise, an ionospheric screen (NISAR), and steep near-fault gradients that push the true per-edge ambiguity into the five-arc range {−2,…,+2}.

Dataset at a glance

Per-regime statistics (from datasheet.md):

Directory layout

LB_DLPU/
├── README.md              # this card
├── datasheet.md           # auto-generated statistics + well-posedness certificate
├── dem_manifest.csv       # DEM tile → split assignment (tile, split, region, regime)
├── index.jsonl            # one JSON record per patch (metadata, no arrays)
├── assets/                # figures used in this card
│   ├── preview.png
│   └── dem_tiles_map.png
├── sim/
│   ├── train/  000000.h5 … 007999.h5      (8,000)
│   ├── val/    008000.h5 … 008999.h5      (1,000)
│   └── test/   009000.h5 … 009999.h5      (1,000)
└── sim_wrapped_png/       # wrapped-phase quicklooks, mirroring sim/
    ├── train/  000000.png … 007999.png
    ├── val/    008000.png …
    └── test/   009000.png …

Patch ids are shared across sim/<split>/<id>.h5 and sim_wrapped_png/<split>/<id>.png.

Per-patch HDF5 schema

Each .h5 file (≈0.7 MB) contains:

psi is the noisy wrapped observation and phi the clean target; they are not exactly congruent (that is the noise the unwrapper must survive). The per-edge labels satisfy Δφ_e = W(Δψ)_e + 2π·k_e, where W wraps to (−π, π].

Per-patch attributes (HDF5 .attrs): sensor (nisar|uavsar), difficulty (smooth|mixed|dense), mean_coherence, px_m (pixel spacing), NL (looks), residue_count, residues_per_mp, max_grad_rad_per_px, frac_edges_k1, frac_edges_k2, label_clip_frac, water_frac, clean_oracle_rmse, seed, and components (JSON: topography / deformation / atmosphere / ionosphere provenance).

index.jsonl

One record per patch with the same metadata as the HDF5 attributes plus id and split, for fast filtering without opening every file:

{"id": "000000", "split": "train", "sensor": "nisar", "difficulty": "smooth",
 "mean_coherence": 0.56, "NL": 8, "px_m": 20.0, "residue_count": 2182,
 "residues_per_mp": 33294.7, "clean_oracle_rmse": 0.0, "seed": 939529293,
 "components": {"topo_src": "dem", "topo": {"B_perp_m": 32.97}, ...}}

Splits

Train / val / test are disjoint by whole DEM tile: the 29 Copernicus GLO-30 tiles (worldwide tectonic, volcanic, and glacial terrain) are partitioned 17 / 5 / 7, so no terrain is shared across splits and the test set measures generalization to unseen geography. The exact tile → split assignment is in dem_manifest.csv.

[image]

Loading

import h5py, glob

def load_patch(path):
    with h5py.File(path, "r") as f:
        return {k: f[k][:] for k in f}, dict(f.attrs)

for p in sorted(glob.glob("sim/test/*.h5"))[:1]:
    arrays, attrs = load_patch(p)
    psi, phi = arrays["psi"], arrays["phi"]          # input, target
    print(attrs["sensor"], attrs["difficulty"], psi.shape)

Quicklooks are plain PNGs:

from PIL import Image
Image.open("sim_wrapped_png/test/009000.png")        # wrapped-phase preview

Intended use

Training and benchmarking L-band phase-unwrapping methods — deep networks (wrap-count regression/classification, gradient estimation) and classical / minimum-cost-flow solvers — with per-regime × difficulty evaluation. The well-posedness certificate makes the test split a fair ceiling reference; the per-edge labels support both pixel-wise and edge-wise supervision.

Reference baselines

A method-blind evaluation harness scores every unwrapper on the identical test split, per regime × difficulty, on RMSE, MAE, PSNR, SSIM, a cycle-slip (jump) rate, residue count, and five-arc |k|≥2 edge accuracy. Reference findings across classical, minimum-cost-flow, and in-domain-trained deep baselines:

• —Deep networks train cleanly on this data and outperform classical and statistical solvers (e.g. SNAPHU) — especially on the dense/high-gradient stratum, which the benchmark is designed to stress.
• —The minimum-cost-flow family is residue-free by construction; with the correct per-edge costs the certified oracle ceiling reaches near-zero, residue-free error, while gradient-domain methods leave residues.
• —The benchmark is not saturated: a gap to the oracle ceiling remains for every deployable method, and the strata form a genuine difficulty gradient (error widens smooth → mixed → dense).

The full leaderboard, metric definitions, and significance tests are in the accompanying code/paper.

Limitations and considerations

• —Synthetic ground truth. phi is physically modelled (topography from real Copernicus DEMs; coherence, noise, and ionosphere calibrated to real NISAR / UAVSAR granules) but is not field-validated absolute truth. It is intended for supervised training and controlled benchmarking; real-scene generalization should be assessed separately.
• —Two regimes. Sensor characteristics are approximated for NISAR-like and UAVSAR-like acquisitions; other L-band sensors may differ.
• —No decorrelation-only patches. Every scene is certified recoverable given correct costs; the benchmark isolates cost/prior estimation, not irrecoverable-noise regimes.

Provenance and licensing

• —Topography: Copernicus GLO-30 DEM (© ESA / Copernicus; free and open, attribution required).
• —Noise / coherence / ionosphere models: calibrated to real NISAR L2 GUNW and UAVSAR granules.
• —Simulated phase, labels, and quicklooks: this release.

License: CC-BY-4.0. Free to use, share, and adapt with attribution. The Copernicus GLO-30 DEM attribution above must be retained. If you intend a different license, update both this line and the license: field in the card metadata before publishing.

Citation

If you use LB-DLPU, please cite both the dataset and the accompanying paper.

Dataset (Zenodo):

@dataset{dagbanja_lbdlpu_data_2026,
  title     = {LB-DLPU: An L-Band (NISAR/UAVSAR) Benchmark for InSAR Phase Unwrapping},
  author    = {Dagbanja, S. and Qian, J.},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21768604},
  url       = {https://huggingface.co/datasets/TheDagbanja/L-Band_DLPU}
}

Paper:

@article{dagbanja_lbdlpu_paper_2026,
  title   = {LB-DLPU: A Recoverability-Certified L-Band InSAR Phase-Unwrapping Benchmark for Reliable Deformation Retrieval},
  author  = {Dagbanja, S., Qian, J. and Haitao, L.},
  journal = {#Will be updated upon publication},
  year    = {2026},
  note    = {under review}
}

Preprint

Dagbanja, Simon and Qian, Jiang and Lv, Haitao, LB-DLPU: A Recoverability-Certified L-Band InSAR Phase-Unwrapping Benchmark for Reliable Deformation Retrieval (August 08, 2026). Available at SSRN: https://ssrn.com/abstract=7358019 or http://dx.doi.org/10.2139/ssrn.7358019