AugmentedUrbanSARFloods

UrbanSARFloods imagery and labels augmented with pre-event and co-event VV wrapped-phase products. This is a reduced subset, not the complete original benchmark.

Contents

5,505 entries: 3,831 training chips, 1,671 validation chips (512 × 512 pixels), and 3 large test mosaics. Approximately 96.4 GB of imagery, labels, phase and provenance. Original split assignments and raster values are preserved. 970 original entries lacking one or both phase pairs are excluded. Sixteen failed processing segments were excluded; details are in release_selection.json.

samples.jsonl is the entry point. sar_path, gt_path, each phase directory, and output_sha256 keys are relative to this repository root.

samples/<split>/<sample_id>/
  sar.tif
  label.tif
  pre/
    phase_cos_sin.tif
    validity.tif
    source_segment.tif
    complete.json
  co/
    phase_cos_sin.tif
    validity.tif
    source_segment.tif
    complete.json

The released SAR raster has eight bands, in order: pre-event coherence VH, VV; co-event coherence VH, VV; pre-event intensity VH, VV; post-event intensity VH, VV. Labels retain their original class IDs. The same event label applies to both temporal views.

The reconstruction uses three acquisitions: t0 and t1 before the flood, and t2 at the flood observation. Pre-event phase uses t0–t1; co-event phase uses t1–t2. Each phase raster stores cosine then sine of VV wrapped phase on the original sample grid. Recover the angle with atan2(sine, cosine). Validity rasters identify available phase pixels; source-segment rasters record one-based contributing-segment provenance.

Load one co-event sample

Install the loader dependencies:

pip install numpy rasterio torch pyarrow huggingface_hub

The example downloads the loader code and the training Parquet index, then fetches the raster files needed for one sample. It returns t1/t2 VV–VH intensity, co-event phase/coherence, the original label, and validity masks. No t0 intensity or pre-event phase files are needed.

import sys
from huggingface_hub import snapshot_download, hf_hub_download

repo = "Ehzoahis/AugmentedUrbanSARFloods"

# Download only the loader code.
code_dir = snapshot_download(
    repo_id=repo,
    repo_type="dataset",
    allow_patterns=["dataloader/**"],
)
sys.path.insert(0, f"{code_dir}/dataloader")

from augmented_urban_sar_floods import UrbanSARFloodParquetDataset

# Download the training index; imagery is fetched when accessed.
index = hf_hub_download(
    repo_id=repo,
    repo_type="dataset",
    filename="default/train/0000.parquet",
    revision="refs/convert/parquet",
)

dataset = UrbanSARFloodParquetDataset(
    [index],
    split="train",
    mode="co_event",
    crop_size=128,  # Shared center crop; omit for the full 512×512 chip.
)

sample = dataset[0]

print("ID:", sample["patch_id"])
print("t1 VV/VH:", sample["prime"].shape)       # (2, 128, 128)
print("t2 VV/VH:", sample["secondary"].shape)   # (2, 128, 128)
print("cos/sin/coh:", sample["insar"].shape)    # (3, 128, 128)
print("Label:", sample["label"].shape)         # (128, 128)

valid = sample["valid"][0] & sample["label_valid"]
print("Usable labeled pixels:", valid.mean())

Outputs are NumPy arrays; a PyTorch DataLoader converts them into batched tensors automatically. Intensities are converted from power dB to linear and normalized using the pretrained model's NA-SAR convention; this does not establish radiometric equivalence. InSAR channel order is cosine, sine, VV coherence. Use valid for loss and evaluation.

HF's converted Parquet is an index of individual TIFF files, not one row per sample. The loader groups them using the sample manifest and pins raster downloads to the source commit recorded in the index. If HF adds shards, supply all Parquet files for the chosen split. Access downloads complete raster files before applying the crop. See the dataloader documentation for more details.

Processing and limitations

SNAP-based processing uses 4 × 1 multilooking, Goldstein filtering, modeled flat-Earth/topographic phase removal, complex geocoding, and measured registration corrections with independent alignment checks. Failed segments were not accepted by relaxing validation thresholds.

3,849 retained entries have full phase coverage; 1,656 have partial phase coverage. Always use the validity masks. Numeric validity does not guarantee high coherence or identify layover/shadow. Exact identity between reconstructed acquisition pairs and the pairs used for released coherence remains unverified. Original coherence is retained rather than recomputed. Final independent scientific QA remains pending; the gallery and successful processing do not establish downstream utility.

Nested segment provenance and completion receipts retain historical producer paths and hashes. Those paths are audit metadata, not downloadable dependencies. Use the top-level portable manifest paths to load the data. Raw SLCs and processing intermediates are not included.

Source attribution

Original imagery, labels, coherence and splits: UrbanSARFloods, https://arxiv.org/abs/2406.04111. This repository adds derived phase and its provenance; it does not claim authorship of the original dataset. No new license is asserted over the original materials; consult the original authors' distribution terms before reuse.