Event-SAR

Event-SAR contains 416 reviewed, coordinate-aligned SAR patches around earthquake and volcanic events: 314 earthquake patches from 53 events and 102 volcano patches from 102 events. It complements NA-SAR using the same WebDataset tar/NumPy packaging approach.

All 416 accepted patches are retained, including patches with partial coverage. The latest final-pass GUI decisions are not used to remove samples. Original review scores (3–5), event information, source provenance, geospatial grids, and exact valid fractions are preserved. Additional matching acquisitions were processed with ASF HyP3 to recover available missing coverage. Missing observations remain missing.

train is a storage split containing the entire corpus, not a prescribed training/evaluation split. Multiple patches can belong to the same event and overlap; use event-aware, spatially and temporally appropriate splits for evaluation.

Browse samples online

Open the six-example gallery — no dataset download or Python required. Three earthquake and three volcano examples, with Fair, Good, and Excellent review labels. Each shows all five stages and available orbits.

Fair, Good, and Excellent describe the existing review ratings; coverage is a separate measure. These are downsampled display previews, not training arrays. Checkerboards indicate missing data. The gallery explains scales and sample selection.

Review rating distribution

Counts cover the full released dataset, not just the six examples. Review ratings and valid coverage are separate measures.

Event type	Fair	Good	Excellent	Total
Earthquake	97	82	135	314
Volcano	23	52	27	102

Download SVG · Counts CSV

Earthquake · Fair · us2000bqsn

Earthquake · Good · us7000irp8

Earthquake · Excellent · us20003k7w

Volcano · Fair · gvp_312030_22129

Volcano · Good · gvp_273010_22151

Volcano · Excellent · gvp_252120_22319

Contents

• One WebDataset sample per accepted spatial patch, with <sample_key>.npz.bin and <sample_key>.json members in a tar shard.
• Five ordered stages: pre0, pre1, event, post0, post1.
• 218 patches have one orbit; 198 patches have two orbits. Orbit IDs and their acquisition times/directions are in the sample metadata. Views are sorted by orbit ID; view_0 does not always mean ascending. No missing view is synthesized.
• Earthquake patches: 314 at 1024×1024 pixels.
• Volcano patches: 49 at 512×512, 47 at 1024×1024, and 6 at 2048×2048 pixels.
• All products within each patch share exactly the same CRS, affine transform, bounds, and dimensions. Alignment does not imply complete valid coverage.
• 23,570 source rasters, retained losslessly as arrays. Original wrapped phase is kept in addition to the derived cosine/sine representation.

For each available view v, with T=5 and the patch's original H,W:

Array key	Shape	Meaning
prime_rtc_view_v	(T,2,H,W)	Reference RTC channels [VV,VH], raw linear backscatter
secondary_rtc_view_v	(T,2,H,W)	Secondary RTC channels [VV,VH]
inc_angle_view_v	(T,1,H,W)	Incidence from local vertical, radians
ifg_view_v	(T,2,H,W)	Derived [cos(phase),sin(phase)]
wrapped_phase_view_v	(T,1,H,W)	Original wrapped phase, radians
coh_view_v	(T,1,H,W)	Coherence
dem	(1,H,W)	Original DEM elevation values, meters
slope_deg	(1,H,W)	Original terrain slope, degrees
common_valid	(1,H,W)	Joint reference/secondary VV + coherence + phase mask
common_valid_dem_vh	(1,H,W)	Joint mask also requiring DEM and reference/secondary VH
common_valid_all	(1,H,W)	Joint mask also requiring incidence and slope

All joint masks intersect every saved orbit and stage. They contain uint8 0/1; scientific arrays retain float32 values and NaNs. A completely unavailable VH acquisition remains NaN and can make the extended/all-product shared ratio zero. DEM source vertical-reference information is preserved; no datum conversion is applied. Incidence is not terrain-local incidence.

Relationship to NA-SAR

The view-key families, RTC channel order, cosine/sine phase channels, raw DEM, slope, tar shards, Parquet index, and shard manifest follow NA-SAR's conventions. Event-SAR adds a leading stage dimension, variable patch sizes and orbit counts, original wrapped phase, validity masks, and rich event metadata. Unlike NA-SAR's zero-padding policy, Event-SAR preserves nodata as NaN for exact coverage filtering. The NPZ payload uses the .npz.bin suffix so Hugging Face keeps it as binary bytes instead of automatically expanding large arrays into nested Python lists. It is an ordinary NumPy NPZ archive, readable with numpy.load. Rich metadata is held in a JSON string to avoid automatic schema inference altering variable source fields. Use the included Event-SAR loader rather than assuming NA-SAR's fixed-shape loader can consume these samples unchanged. No normalization, resizing, cropping, or coverage-based exclusion is applied during export.

Metadata and coverage filtering

metadata.parquet has one row per patch, including:

• sample_key, shard_path, shard_index, sample_index_in_shard;
• event type/ID, patch ID, original score, stages, orbit IDs, dimensions;
• CRS, affine transform, projected/geographic bounds, event metadata JSON;
• shared_valid_ratio, shared_valid_dem_vh_ratio, shared_valid_all_ratio;
• NPZ/JSON checksums and the original metadata checksum.

file_metadata.parquet has one row per source raster, with its exact valid_ratio, original GeoTIFF checksum, product/stage/orbit, and corresponding array key, stage index, and channel index inside the NPZ.

Each sample's JSON has a fixed outer schema with identity and coverage fields. Its metadata_json string preserves the complete patch/event/acquisition metadata and adds hf_export: the view-to-orbit mapping, array shapes, source-to-array mapping, and GeoTIFF units, scales, offsets, descriptions, and tags. Machine-specific path prefixes are replaced by source/diskN/ and repository/; these are provenance identifiers, not URLs. The local original GeoTIFF release remains separate.

All ratios are exact fractions in [0,1], not rounded percentages:

Shared VV + InSAR coverage	Patches
Exactly 100%	49
At least 99%	347
Above 95%	385
At or below 95%	31

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download('GFM-Bench/Event-SAR', 'metadata.parquet', repo_type='dataset')
metadata = pd.read_parquet(path)
selected = metadata[metadata.shared_valid_ratio > 0.95]  # 385 patches

Streaming with Hugging Face Datasets

from io import BytesIO
import numpy as np
from datasets import load_dataset

stream = load_dataset('GFM-Bench/Event-SAR', split='train', streaming=True)
sample = next(iter(stream))
with np.load(BytesIO(sample['npz.bin']), allow_pickle=False) as archive:
    rtc = archive['prime_rtc_view_0']   # (5,2,H,W)
    mask = archive['common_valid']     # (1,H,W)
import json
metadata = json.loads(sample['json']['metadata_json'])
print(metadata['event_id'], metadata['common_valid_fraction'])

The explicit equivalent is load_dataset('webdataset', data_files={'train':'hf://datasets/GFM-Bench/Event-SAR/data/event-sar-train-*.tar'}, split='train', streaming=True). Each large sample contains all of its stages; streaming does not decompress the entire dataset in advance.

Local PyTorch loading

Download the published folder with huggingface_hub.snapshot_download, or use an existing local export. Import the included event_sar_dataset.py:

from torch.utils.data import DataLoader
from event_sar_dataset import EventSARWebDataset

ds = EventSARWebDataset('/path/to/Event-SAR-HF', min_shared_valid_ratio=0.95)
# Threshold applies inclusively to the cropped common_valid mask.
loader = DataLoader(ds, batch_size=None, num_workers=2)
sample = next(iter(loader))
rtc = sample['arrays']['prime_rtc_view_0']

# To select one stage and remove its leading dimension:
event_only = EventSARWebDataset('/path/to/Event-SAR-HF', stage='event')
# event_only yields prime_rtc_view_0 with shape (2,H,W).

Variable sizes and orbit counts require an appropriate batching/collation policy. The loader preserves raw values; models should handle NaNs using the supplied masks and apply their own normalization.

The loader supports aligned crops and product selection:

from event_sar_dataset import EventSARDataset, event_sar_collate

ds = EventSARDataset(
    '/path/to/Event-SAR-HF',
    event_type=['earthquake', 'volcano'],
    mode='all',                         # rtc, insar, or all (default)
    stages=['event'],
    crop_size=512,
    crop_mode='all',                    # random, central, or all
    min_shared_valid_ratio=0.95,
)
loader = DataLoader(ds, batch_size=4, collate_fn=event_sar_collate)

rtc returns reference/secondary VV/VH; insar returns cos/sin phase, wrapped phase and coherence. all includes both plus DEM, slope and incidence. Explicit include_dem, include_slope, include_inc_angle flags override auxiliary defaults (off for rtc/insar, on for all). Unselected arrays are not decompressed.

Filtering uses only the cropped common_valid mask, intersecting VV/InSAR across all original stages/views regardless of mode or selected stages. This mask does not require valid VH or auxiliary products. No channel-specific thresholds are used. A failing crop is skipped, with no random retries. The output shared_valid_ratio reports actual crop coverage. Scores and score filters are not exposed: the historical 3–5 review ratings are not final decisions. They remain in raw source metadata solely as provenance.

all crop mode produces separate nonoverlapping tiles (1024 -> four 512 tiles); dimensions must divide exactly. random produces one repeatable crop per sample and epoch. Call set_epoch(epoch) before iteration; recreate persistent workers after changing epochs. central selects the image center. event_ids and sample_keys support custom selections. Raw values and NaNs remain unchanged.

Outputs include event date, center coordinates, magnitude/type (Mw or VEI), stages, crop window and adjusted grid. Original patch coordinates are saved in Parquet (transform, CRS, bounds, dimensions) and JSON (grid). Earthquake JSON also preserves reviewed_box row/column offsets; volcano grids are event-centered. Loader-created crop_window coordinates are relative to the native exported patch and are generated at runtime, not stored on HF.

The collator returns a list to accommodate variable orbit counts. When filtering, length is unknown until iteration; num_candidate_crops is an upper bound. Shards are partitioned across workers/ranks; training must handle uneven counts.

Provenance and integrity

SAR observations are Sentinel-1 products processed with ASF HyP3. Exact source acquisitions, orbit groups, stage assignments, event catalog attributes, processing parameters, and terrain provenance are recorded per sample. Original provider attribution and source-product terms remain applicable.

webdataset_summary.json summarizes the export and verification. shard_manifest.jsonl records each tar shard's sample count, size, and SHA256. SHA256SUMS covers the published release files. Every source raster checksum/grid was checked, every raw exported array was round-trip verified, and the joint masks were reconstructed and compared before packaging. Cosine/sine phase is derived; the original phase array is retained for exact reconstruction.

Remaining nodata, partial footprints, variable acquisition intervals, and absent VH are part of this release. A high valid ratio measures data availability, not event visibility or coherence quality. Event labels do not constitute pixel-level damage or deformation annotations.