NA-SAR

Finalized compact release — October 24, 2026

The compact release is published on main. All 1,099,604 samples in 91 shards are retained; this is a storage reorganization, not a new sample selection or temporal filter.

Component	Size (decimal GB)
SAR WebDataset shards	982.97
Shared DEM and slope store (terrain/, 87,755 objects)	9.85
Optional environmental layers (auxiliary/)	50.80
Published release files, including metadata and loaders	1,043.73

Every sample passed reconstruction validation: retained SAR NPY payloads and restored terrain arrays are byte-identical to the previous release. Focused loader tests and real-data comparisons verified default RTC/InSAR outputs, distributed/worker partitioning, and opt-in environmental loading. Environmental layers remain disabled by default; they do not change sample selection.

Download data/, terrain/, the root metadata files, and the supplied Python modules together for the original InSAR loader behavior. Add auxiliary/ only when environmental layers are needed. Use the updated loader shipped with this release: a legacy loader that expects DEM inside each NPZ cannot reconstruct the compact terrain store. RTC-only loading does not require terrain.

Sizes describe the published release files, not Hugging Face's storage accounting for historical revisions. The CTI coverage and MODIS placeholder policy below apply to this finalized release.

Compact terrain and environmental auxiliary layers

This release retains all 1,099,604 samples, acquisition dates, quality selection, and SAR values. Repeated DEM/slope arrays now share storage under terrain/. The provided NASARWebInSARDataset restores them automatically before applying the original normalization. Default output keys, numerical values, temporal sampling, shuffling, and distributed partitioning are preserved. NASARWebRTCDataset retains its original behavior. Both loaders also support older shards with embedded terrain.

Storage-format change: the repacked NPZ files no longer directly contain dem or slope_deg. Generic WebDataset/NumPy readers see only the remaining SAR arrays. Use the provided loader and retain terrain/ for the original InSAR outputs, or explicitly join the terrain store by the exact sample key. Copying only the tar files is insufficient for InSAR terrain reconstruction.

The optional environmental layers occupy 50.8 GB under auxiliary/ and are losslessly deduplicated across exact grids and source years. They include:

Layer	Product and interpretation
Land cover and QC	MODIS MCD12Q1 v6.1 IGBP classes, annual maps matched to acquisition year except marked placeholders below
Aridity and 12 monthly PET maps	Corrected CGIAR Global-AI/PET v3.1; fixed 1970–2000 climatologies
Compound Topographic Index	USGS HDMA CTI with regional and Greenland gap repair; static terrain-derived wetness proxy

Final CTI coverage is 1,078,966 fully covered, 20,131 partially covered, and 507 without coverage. Repair improved 47,355 samples. The remaining completely missing samples are Canadian locations with source nodata. Missing pixels retain masks and NaN values; unsupported values were not fabricated. The compact store already incorporates the final repair and requires no separate CTI overlay.

2025 MODIS placeholder policy: observations acquired in 2025 use 2024 land-cover and QC maps. At the 2026-10-21 source-catalog check, all 23 required 2025 tiles were unavailable. This affects 186,736 samples with at least one 2025 observation. Other observations use their acquisition year's maps. Original acquisition dates remain unchanged.

Per-observation metadata retains acquisition_year, source_year, is_temporal_placeholder, temporal_mapping_reason, and array keys. A map named landcover_2025 is acquisition-year keyed; its source year is 2024 and its placeholder flag is true. These placeholders cannot represent land-cover changes in 2025. Analyses requiring exact-year labels should exclude flagged observations or replace them when suitable maps become available.

Static layers and each MODIS source-year bundle are shared only when exact grid, dtype, shape, and array bytes match. All samples were reconstructed and checked against the finalized originals. CTI is decoded with factor 0.01; aridity uses factor 0.0001 and the support mask from annual ET0. Categorical maps use nearest-neighbor resampling and continuous layers use bilinear resampling. Alignment does not increase the products' native resolution.

Opt-in environmental loading

Keep the supplied Python modules together beside the downloaded release. Install requirements.txt. Environmental auxiliary loading remains disabled unless auxiliary_root is provided:

from nasar_dataset import NASARWebRTCDataset, NASARWebInSARDataset

# Original model-facing behavior; terrain is found under the dataset root.
insar = NASARWebInSARDataset('/path/to/NA-SAR', require_dem=True)
rtc = NASARWebRTCDataset('/path/to/NA-SAR')

# Explicitly add separate environmental arrays and provenance fields.
insar_aux = NASARWebInSARDataset(
    '/path/to/NA-SAR', auxiliary_root='/path/to/NA-SAR/auxiliary'
)

Opt-in samples add auxiliary and auxiliary_metadata; no environmental channels are automatically concatenated or used to filter samples. Apply the same spatial transforms to SAR, auxiliary arrays, and masks. Use a suitable collator for variable land-cover year keys/metadata, or batch_size=None. CGIAR maps are climatologies, and CTI is static; neither measures weather or soil moisture at the SAR acquisition time.

Sources: MODIS, CGIAR v3.1, USGS HDMA.

NA-SAR is a North America synthetic aperture radar pretraining dataset built from NASA OPERA products. It is intended for self-supervised and masked-image pretraining of SAR foundation models.

This release is an unsplit pretraining corpus. There are no train/validation/test partitions because the dataset is not intended to define an evaluation protocol. Downstream evaluations should define their own geographically and temporally appropriate splits.

Browse the current NA-SAR release

Open the six-sample gallery to inspect corrected SAR, shared terrain, MODIS land cover, CGIAR aridity/PET, and finalized CTI together.

These previews were regenerated directly from the compact release's NPZ shards and exact-key terrain/auxiliary stores. They show the corrected phase/coherence. The same six examples span automated quality-score percentiles 10, 25, 40, 60, 75, and 90, with distinct anchor bursts; they are illustrative, not a random sample.

The quality distribution and scores are unchanged by the repack and phase/coherence correction. They describe the original automated selection, not manual review.

Each panel shows both SAR views, shared DEM/slope, prime-view-0 annual land cover, static CTI, aridity, and January/July PET climatologies. All 12 PET months are available in the data. Land-cover acquisition/source years are recorded in sample provenance; 2025 observations use marked 2024 placeholders. Gray indicates masked/non-finite data or nonpositive RTC, including zero-padded missing polarization.

RTC uses log1p(20x) and fixed display limits across samples, dates, and views: co-pol 0.15–2.39 and cross-pol 0.01–1.09. DEM, aridity, CTI, and PET use fixed shared p2–p98 ranges estimated from 256 uniformly sampled released observations (seed 20261024), with at most 512 valid values per product per observation. All 12 PET months were pooled to establish one monthly PET range. This is a sample-weighted estimate, not a full-corpus or location-balanced statistic. Exact ranges and sampled keys are provided for reproducibility.

Display layer	Shared range
DEM	1.52–2177.08 m
Slope	0–45 degrees
Aridity P/ET0	0.03494–2.13745
CTI	5.75896–13.38149
Monthly PET	7.00–301.56 mm/month
Coherence	0–1
Wrapped phase	−π–π radians
Incidence	0–90 degrees

Colors saturate outside these display limits; stored arrays are unchanged. Nearly uniform AI/PET colors indicate small local variation relative to the shared scale, not missing data. The MODIS panel overlays discrete classes at 40% opacity on prime-view-0 co-pol RTC; white lines show native MODIS cell edges transformed into the SAR grid. Overlay colors blend with SAR brightness. Class legends are included on each image and below. MODIS's approximately 463 m source cells do not become 30 m labels when resampled.

Sample 000000370819 · automated quality 0.581

Sample 000000189514 · automated quality 0.724

Phase and coherence correction — October 2026

All 91 shards (1,099,604 samples, both views) now contain recomputed phase and conventional coherence from the source OPERA CSLC acquisitions. Coherence uses 5 × 5 local complex cross moments and powers on the RTC grid. Phase retains 3 × 5 circular smoothing with corrected invalid-input handling; no Goldstein filter is applied. Stored coherence remains in [0,1], and phase remains cos/sin. Other sample arrays and sample membership are unchanged.

Full coherence statistics: finite-pixel mean 0.4734929473796252, population standard deviation 0.275366784057543. These describe both views of the released subset, including overlapping pixels. There are 27 NaN coherence pixels; no infinities or finite out-of-range values. Statistics do not imply that the stored arrays have been standardized.

The gallery previews now show the corrected phase and coherence from the current release. Metadata quality scores and sample selection remain those of the original release; they were not recomputed.

Dataset Contents

• Samples: 1,099,604
• Shards: 91 WebDataset tar shards
• Patch size: 128 x 128 pixels
• RTC views: two spatial views per sample
• Temporal RTC acquisitions: prime and secondary
• RTC channels: co-pol and cross-pol, with missing polarization zero padded
• Incidence angle: one channel per spatial view
• InSAR phase: stored as cos(phi), sin(phi)
• Coherence: one channel per spatial view
• DEM elevation: raw elevation in meters
• DEM slope: raw terrain slope in degrees

The logical sample contains the arrays below. dem and slope_deg are reconstructed from the shared terrain store; the other arrays remain in the NPZ shards:

Key	Shape	Description
prime_rtc_view_0	(2, 128, 128)	Prime RTC view 0, co-pol/cross-pol
secondary_rtc_view_0	(2, 128, 128)	Secondary RTC view 0, co-pol/cross-pol
prime_rtc_view_1	(2, 128, 128)	Prime RTC view 1, co-pol/cross-pol
secondary_rtc_view_1	(2, 128, 128)	Secondary RTC view 1, co-pol/cross-pol
inc_angle_view_0	(1, 128, 128)	Incidence angle for view 0
inc_angle_view_1	(1, 128, 128)	Incidence angle for view 1
ifg_view_0	(2, 128, 128)	InSAR phase for view 0 as cos/sin
coh_view_0	(1, 128, 128)	Coherence for view 0
ifg_view_1	(2, 128, 128)	InSAR phase for view 1 as cos/sin
coh_view_1	(1, 128, 128)	Coherence for view 1
dem	(1, 128, 128)	Raw DEM elevation in meters
slope_deg	(1, 128, 128)	Raw terrain slope angle in degrees

dem_relief and normalized slope are not stored in the WebDataset. They are derived at loading time from the raw dem and slope_deg arrays.

Metadata

metadata.parquet is the publishable metadata table for the sharded release. It includes OPERA provenance, patch geometry, quality score, polarization availability, and the following sharding columns:

• sample_key: sample key inside the WebDataset shard
• shard_path: relative path to the tar shard
• shard_index: integer shard id
• sample_index_in_shard: row order within that shard

The metadata is filtered to samples with quality_score > 0.53.

webdataset_summary.json records the shard count and DEM storage policy. For this release, raw terrain arrays remain dem and slope_deg, with terrain_storage identifying their shared store.

Loading With Hugging Face Datasets

The SAR-only NPZ payload can be streamed with a generic WebDataset loader. This example does not reconstruct shared DEM/slope; use the provided PyTorch loader below for the original InSAR terrain outputs:

from io import BytesIO

import numpy as np
from datasets import load_dataset

ds = load_dataset(
    "webdataset",
    data_files={"train": "data/nasar-train-*.tar"},
    split="train",
    streaming=True,
)

sample = next(iter(ds))
arrays = np.load(BytesIO(sample["npz"]))
print(arrays.files)

Loading With PyTorch

The release includes nasar_dataset.py, a lightweight PyTorch loader for the local shard layout:

from nasar_dataset import NASARWebRTCDataset, NASARWebInSARDataset

rtc_ds = NASARWebRTCDataset("/path/to/NA-SAR-HF")
insar_ds = NASARWebInSARDataset("/path/to/NA-SAR-HF", require_dem=True)

For InSAR, NASARWebInSARDataset returns SAR arrays plus DEM-derived channels:

Loader output key	Source array	Default normalization
dem_view_0/1	dem	global elevation, (-100 m, 4000 m) -> [0, 1]
dem_relief_view_0/1	dem	patch-local p5/p95 relief -> [0, 1]
slope_view_0/1	slope_deg	slope degrees, (0 deg, 45 deg) -> [0, 1]

These normalization ranges can be changed through the loader constructor.

Preprocessing Notes

RTC arrays are stored as raw linear backscatter with invalid, non-finite, and negative values set to zero. The helper loader applies the pretraining-time RTC transform by default:

x = log1p(20.0 * x)
co_pol = clip_and_rescale(co_pol, p1=0.15, p99=2.39)
cross_pol = clip_and_rescale(cross_pol, p1=0.01, p99=1.09)

The co-pol and cross-pol RTC channels use separate normalization ranges because cross-pol backscatter is substantially darker. Fully missing polarizations remain zero-padded after preprocessing.

InSAR phase is stored as cosine and sine channels instead of wrapped phase radians to avoid phase discontinuities at the wrap boundary.

DEM arrays are static for the spatial patch and shared across both orbit views. dem is raw elevation in meters. slope_deg is raw terrain slope in degrees. Derived terrain features, including global-normalized DEM, patch-local relief, and normalized slope, are intentionally computed during loading rather than stored as processed arrays.

Source and Scope

The source SAR data comes from NASA OPERA products over North America. Terrain comes from aligned DEM patches. Users should follow the applicable terms and citation guidance for OPERA and DEM source products.

Intended Use

This dataset is intended for SAR and InSAR representation learning, especially self-supervised pretraining. It is not an evaluation benchmark by itself.

Limitations

• Coverage is geographically focused on North America.
• Quality filtering removes lower-quality samples and can bias the corpus toward easier or cleaner acquisitions.
• Missing RTC polarizations are represented by zero-padded channels. The metadata includes polarization availability flags so users can distinguish true zeros from missing channels.
• DEM fields are aligned to the 128 x 128 patch grid and should not be treated as a standalone high-resolution terrain product.