Dataset Viewer
Auto-converted to Parquet Duplicate
row_index
int64
sample_id
string
grid_id
string
quarter
string
selection_role
string
grid_epsg
int32
grid_transform
list
valid_fraction
float64
rtc_score
float64
coh_block16_std
float64
modis_source_year
int16
prime_time
timestamp[us, tz=UTC]
secondary_time
timestamp[us, tz=UTC]
modis_class_fractions
list
sar_member
string
aux_shard
string
aux_member
string
sar_sha256
string
aux_sha256
string
sar_npz
unknown
aux_npz
unknown
0
32601:57:747|2024Q3
32601:57:747
2024Q3
original_primary
32,601
[ 30, 0, 437760, 0, -30, 5744640 ]
0.99881
0.843638
0.169337
2,021
2024-08-17T18:03:07
2024-08-29T18:03:07
[0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.01050912588438682,0.5256191389826519,0.1373584041259311,0.0,0.0,0(...TRUNCATED)
samples/32601_57_747/2024Q3.npz
aux-00000.tar
aux/32601_57_747.npz
6a0b4589fd119a47e5dc9fbdc600acb992b0ccb6d04be1e48699e83d1c939017
174fff15450054d9f2b8b0c9179d83a163657fd02bf303e3afb00ac0c86dc742
"UEsDBC0AAAAIAAAAIQDsuEBu//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAABvIQcAAAAAAJSaeTS(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQBh7NTr//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAALPTAgAAAAAA7L13UJTd2vVJzjknyUiOghg(...TRUNCATED)
1
32601:61:746|2022Q1
32601:61:746
2022Q1
original_primary
32,601
[ 30, 0, 468480, 0, -30, 5736960 ]
0.999954
0.801698
0.217796
2,021
2022-02-05T18:02:57
2022-02-17T18:02:57
[0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0051740873523811294,0.5230941138762104,0.05887452788353197,0.0,0.(...TRUNCATED)
samples/32601_61_746/2022Q1.npz
aux-00000.tar
aux/32601_61_746.npz
fca4e5e6c538635112d4a5501550bd58f4d331cfef72317b23ba29d6f98b099a
ccdc818552b788a657e2f0fe84864ef73233c5ff3a15e3564559d8acf880e28b
"UEsDBC0AAAAIAAAAIQAKWcHp//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAAARMwcAAAAAAJSY6Tu(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQB38EgI//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAAFDdAQAAAAAA7N0HUFZbui5qREVBMSEokgU(...TRUNCATED)
2
32601:62:746|2024Q3
32601:62:746
2024Q3
original_primary
32,601
[ 30, 0, 476160, 0, -30, 5736960 ]
0.999924
0.599655
0.126769
2,021
2024-08-17T18:03:07
2024-08-29T18:03:07
[0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.7219515475543089,0.04470242279698678,0.0,0.0,0.0,0.0,0.0,0.23(...TRUNCATED)
samples/32601_62_746/2024Q3.npz
aux-00000.tar
aux/32601_62_746.npz
ff4ff10c35fb61870c45e792fbf749c51780a4e53fcf1f4e99630d2c1955a53b
a142981a86825cf0820fe4f01acb9ed4cc2639911223112c261c1844838f4af3
"UEsDBC0AAAAIAAAAIQAvymLJ//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAABtFgcAAAAAAJSX919(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQCDZBc8//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAAFFeAwAAAAAA7LsHVJTbvuVrwLSNiBlzBDF(...TRUNCATED)
3
32601:63:746|2022Q1
32601:63:746
2022Q1
original_primary
32,601
[ 30, 0, 483840, 0, -30, 5736960 ]
1
0.690033
0.157277
2,021
2022-02-05T18:02:57
2022-02-17T18:02:57
[0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0025305091597443343,0.6026554464216998,0.06547886113865775,0.0,0.(...TRUNCATED)
samples/32601_63_746/2022Q1.npz
aux-00000.tar
aux/32601_63_746.npz
4c4c4a793cded57fcdc43fc10a5e079bcacfb75114b4ec97b10cf4a5d512cc88
71ce0995eb9c47707a9fd2c1612cbd7760139d9b7a82e1b1ddf7686ec66054fa
"UEsDBC0AAAAIAAAAIQCL+qXO//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAACWIgcAAAAAAJSY91v(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQDY810S//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAAAtKAwAAAAAAlLl1VFtR47WZmyB195a6uwt(...TRUNCATED)
4
32601:63:748|2024Q3
32601:63:748
2024Q3
original_primary
32,601
[ 30, 0, 483840, 0, -30, 5752320 ]
0.995773
0.99958
0.211939
2,021
2024-08-17T18:03:04
2024-08-29T18:03:04
[0.0025575385036713987,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.18167526074449833,0.2724097617855898,0.0,0.(...TRUNCATED)
samples/32601_63_748/2024Q3.npz
aux-00000.tar
aux/32601_63_748.npz
2271aeaf9b4f6b573b76b9e6e7f237d4292b05523ffd65d5bf6d1a3134d6f8e7
ba0f8048f448691a15f05487a5a1d597f152f7da8d708137c5e2228901310f18
"UEsDBC0AAAAIAAAAIQCQwpdo//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAAAOOwcAAAAAAJSZeVR(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQDu8rkd//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAAF1mAgAAAAAA7Lr5U02PA/9fSiVpT6SdNoo(...TRUNCATED)
5
32601:64:748|2022Q1
32601:64:748
2022Q1
original_primary
32,601
[ 30, 0, 491520, 0, -30, 5752320 ]
0.999954
0.999995
0.094174
2,021
2022-02-05T18:02:54
2022-02-17T18:02:54
[0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.12334109546225028,0.30051293602994145,0.0,0.0,0.0,0.006407747(...TRUNCATED)
samples/32601_64_748/2022Q1.npz
aux-00000.tar
aux/32601_64_748.npz
4c62a4982b4f9a662eaaf1bd5f30e1a513a58d8c157c3e3617938386e4776e62
4c7c295da720620f9c41ea05df4b1f3cdcf6148f372625a8449e1b19abac4999
"UEsDBC0AAAAIAAAAIQDwE0QK//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAAA2UQcAAAAAAJSX91+(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQBrY3vV//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAAN/bAQAAAAAA7LxnUBVv36ULKgbMAXNWzDl(...TRUNCATED)
6
32601:66:745|2024Q3
32601:66:745
2024Q3
original_primary
32,601
[ 30, 0, 506880, 0, -30, 5729280 ]
0.999664
0.759106
0.259234
2,021
2024-08-17T18:03:07
2024-08-29T18:03:07
[0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.005076290881858229,0.41495399269140415,0.13190225882436615,0.0,0.(...TRUNCATED)
samples/32601_66_745/2024Q3.npz
aux-00000.tar
aux/32601_66_745.npz
ec2973103a765e27b2f6f656890ffd6a251db932a9431a62d0bfd16d7e83d513
c4a2afb499add48e6b9dc025de482276db834512e7fa201fda6817dc78436f2c
"UEsDBC0AAAAIAAAAIQDoPXqj//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAADUMQcAAAAAAJSX9ze(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQB0fbXw//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAAH41AgAAAAAAzJt3VJT7kq6bzoGmaTIIihF(...TRUNCATED)
7
32601:66:746|2022Q1
32601:66:746
2022Q1
original_primary
32,601
[ 30, 0, 506880, 0, -30, 5736960 ]
0.999756
0.735042
0.181804
2,021
2022-02-05T18:02:57
2022-02-17T18:02:57
[0.0,0.0,0.0,0.0,0.0,0.0,0.005068181968098397,0.0,0.007602272952147591,0.38406747382554995,0.2028106(...TRUNCATED)
samples/32601_66_746/2022Q1.npz
aux-00000.tar
aux/32601_66_746.npz
6cabf7535ac4a617b531ff69982194959828d254b1fa2b6ca8402bf496862701
2060cb8ffe2ca1f1937b184d117ed97bb78f43db7a24feaa2870e0509a454451
"UEsDBC0AAAAIAAAAIQDdcDSx//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAAArNwcAAAAAAJSXh1f(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQDE3gq8//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAANaNAgAAAAAAzLsFVFZvvOcrAordXdgBimI(...TRUNCATED)
8
32601:67:745|2024Q3
32601:67:745
2024Q3
original_primary
32,601
[ 30, 0, 514560, 0, -30, 5729280 ]
1
0.676756
0.224079
2,021
2024-08-17T18:03:07
2024-08-29T18:03:07
[0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.007583575841752913,0.23734153071519443,0.2841974362933635,0.0,0.0(...TRUNCATED)
samples/32601_67_745/2024Q3.npz
aux-00000.tar
aux/32601_67_745.npz
cca0e4be1a02c72aac765486df30d9be132f521701e2b452577c63591f1bc91f
93c413e5a8d3019441955055db683ff5565c841908c82ea0727846e8b79c4455
"UEsDBC0AAAAIAAAAIQAdasCq//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAADfIgcAAAAAAJSX9z/(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQCb20g7//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAAJ+iAgAAAAAAlLh1VFRv/LZLCyKh2F2oX7G(...TRUNCATED)
9
32601:67:747|2022Q1
32601:67:747
2022Q1
original_primary
32,601
[ 30, 0, 514560, 0, -30, 5744640 ]
0.999847
0.762714
0.180905
2,021
2022-01-07T17:54:47
2022-01-19T17:54:46
[0.0025197595301958953,0.0,0.0,0.0,0.010039969799517667,0.0,0.0003011729206799374,0.0,0.0,0.60796281(...TRUNCATED)
samples/32601_67_747/2022Q1.npz
aux-00000.tar
aux/32601_67_747.npz
fcd32fc23d3532e203c1c8d08dd4d64aafce0ea34e14eb0172f4293ab2abd9f9
d22cb784ec1a9b4adb34b16e113686f914c21eb8f6c065cdce009747360b9c8a
"UEsDBC0AAAAIAAAAIQARQodS//////////8UABQAcHJpbWVfcnRjX3ZpZXdfMC5ucHkBABAAgAAIAAAAAAC+LgcAAAAAAJSXdzi(...TRUNCATED)
"UEsDBC0AAAAIAAAAIQD5A7/K//////////8HABQAZGVtLm5weQEAEACAAAQAAAAAAM/XAgAAAAAAvLoHUFXfv+VJFEkSBVEUMSA(...TRUNCATED)
End of preview. Expand in Data Studio

NA-SAR-X

NA-SAR-X contains 100,000 paired SAR samples from 83,134 distinct 256 × 256 local UTM grids at 30 m spacing. There are 66,268 locations with one quarter and 16,866 with two different quarters. Each sample contains primary and secondary RTC, interferometric phase cosine and sine, coherence, incidence angle, validity masks, georeferencing, and 2021 MODIS land-cover fractions.

The default Parquet split contains 500 shards of 200 samples each. Each row includes both the SAR and auxiliary rasters with their masks. All 100,000 rows were read back and both payload hashes were checked against the release manifest. See verification results and the Parquet dataloader guide.

Selection required a valid-pixel fraction of at least 0.995 and either an RTC quality score of at least 0.6 or a standard deviation of 16 × 16 coherence block means of at least 0.1. The additional 16,866 second-quarter samples prioritize urban locations and reduce nearby repeat sampling of forest, cropland, and grassland.

Pixel-aligned auxiliary layers

Each Parquet row includes its matching auxiliary patch in aux_npz, with the same grid_id as the SAR payload. The 100,000 samples share 83,134 auxiliary patches. The two quarters at the same location use the same terrain, climatology and 2021 land-cover rasters. Each auxiliary patch has the same local UTM EPSG, 30 m affine transform and 256 × 256 shape as its SAR sample.

Field under sample["auxiliary"] Bands Source and units
dem 1 Copernicus GLO-30 elevation, metres
slope_deg 1 DEM-derived slope, degrees
aridity_index 1 CGIAR Global-AI_PET v3.1 annual aridity, P/ET₀
cti 1 USGS HDMA compound topographic index, dimensionless
pet_monthly 12 CGIAR v3.1 monthly reference ET₀ climatology, January–December, mm/month
landcover 1 2021 MCD12Q1 LC_Type1 (IGBP), categorical
landcover_qc 1 2021 MCD12Q1 QC, categorical

Each field has a boolean mask named <field>_valid with the same shape. Continuous invalid pixels are NaN; categorical invalid pixels have value 255. Use the masks for training. The 30 m output grid preserves the registration of coarser products; their native spatial resolution remains unchanged. The MODIS class raster complements the original 18 full-patch area fractions.

Read the Parquet dataset

The default train split contains the 100,000 final samples in manifest order. Each row contains the original compressed SAR and auxiliary NPZ bytes, their SHA-256 digests, and searchable metadata. Samples from two quarters at the same location each include that location's auxiliary payload, so either row can be read independently. The root metadata.parquet contains the row metadata and shard addresses for filtering without scanning the raster payloads.

Install datasets, numpy, and huggingface_hub and run:

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

stream = load_dataset("GFM-Bench/NA-SAR-X", split="train", streaming=True)
row = next(iter(stream))
print(row["sample_id"], row["grid_epsg"], row["valid_fraction"])

for group in ("sar", "aux"):
    assert hashlib.sha256(row[group + "_npz"]).hexdigest() == row[group + "_sha256"]

with np.load(io.BytesIO(row["sar_npz"]), allow_pickle=False) as sar:
    coherence = sar["coh_view_0"]            # [1, 256, 256]
    valid = sar["valid"]                     # [256, 256]

with np.load(io.BytesIO(row["aux_npz"]), allow_pickle=False) as aux:
    dem = aux["dem"]                         # [1, 256, 256], metres
    monthly_pet = aux["pet_monthly"]         # [12, 256, 256], Jan–Dec
    dem_valid = aux["dem_valid"]

For all fields together, download nasar_parquet.py and call decode_sample(row, verify_sha=True). For PyTorch batching and lookup by sample ID, see DATALOADER.md.

For catalog-only analysis, use load_dataset("GFM-Bench/NA-SAR-X", "metadata", split="train", streaming=True). The catalog includes both acquisition timestamps in UTC, local UTM grid and affine transform, the 18 MODIS fractions, quality metrics, SHA-256 digests, and the Parquet shard and row for each sample.

Spatial distribution

The map counts samples by the location of their patch center in 1° longitude × 1° latitude cells. The final list occupies 2,104 cells. Each retained quarter contributes one count, so a location with two quarters contributes twice. The color scale is logarithmic. The geographic bins are for this figure; the patches themselves use their local UTM grids. Land outlines are from Natural Earth.

Sample counts by 1-degree cell across North America

Land-cover class weights

The figure shows sample-weighted votes from the 2021 MODIS land-cover fractions. We merge IGBP forest classes 1–5 first. A sample divides one vote equally among the classes covering strictly more than 20% of its patch. NoData receives no vote. Of the 100,000 samples, 99,986 contributed votes and 14 had no qualifying class. Grassland accounts for 21.17%, forest 11.63%, cropland 9.55%, and urban land 3.04% of the votes. Cropland and crop mosaic are reported separately.

Share of votes for all 13 merged land-cover classes

Four random samples

Each row is one 256 × 256 patch, labeled by grid ID and quarter. Four rows were drawn uniformly without replacement from the final manifest in its published order, using Python's random.Random(20261114).sample(range(100000), 4). Each NPZ was checked against its manifest SHA-256 when read.

The columns show primary and secondary co-polarized RTC on the same −25 to 0 dB display scale, coherence from 0 to 1, and wrapped phase from −π to π radians. RTC uses 10 * log10(linear RTC) for display. Phase uses atan2(sin_phase, cos_phase). Gray marks masked pixels; valid RTC values outside the display range are clipped in the figure.

Four random paired RTC, coherence and phase examples

SAR and auxiliary grid alignment

The panels below use the same four manifest rows as the gallery above. Both the SAR and auxiliary NPZ payloads were SHA-256 checked. For each pair the loader verified the grid ID, UTM EPSG and all six affine coefficients. Every panel shows the full 256 × 256 pixel array with the same row and column orientation, covering 7.68 km on each side.

The first row shows primary and secondary RTC, coherence and wrapped phase. The second row shows DEM elevation, slope, aridity and CTI. The third row shows 2021 land cover, raw QC codes, July reference ET₀ and primary RTC with cyan DEM elevation contours. The links under each panel show all twelve ET₀ climatology months. Each monthly sheet uses a common scale across months. Continuous auxiliary color ranges use the patch's 2nd and 98th percentiles for display. Land-cover and QC colors indicate the codes on the legend. Gray pixels denote masked values. Coarser land-cover and climate products retain their native detail on the 30 m grid.

32606:82:923 | 2021Q1

Matched SAR and auxiliary rasters for 32606:82:923 | 2021Q1

Full-size panel · January–December reference ET₀

32613:39:387 | 2025Q3

Matched SAR and auxiliary rasters for 32613:39:387 | 2025Q3

Full-size panel · January–December reference ET₀

32620:44:738 | 2025Q3

Matched SAR and auxiliary rasters for 32620:44:738 | 2025Q3

Full-size panel · January–December reference ET₀

32611:72:734 | 2025Q4

Matched SAR and auxiliary rasters for 32611:72:734 | 2025Q4

Full-size panel · January–December reference ET₀

Read a batch with PyTorch

Install the dependencies and download the reader into the directory where you run Python.

pip install datasets numpy torch huggingface_hub
hf download GFM-Bench/NA-SAR-X nasar_parquet.py --local-dir .
from functools import partial
from datasets import load_dataset
from torch.utils.data import DataLoader
from nasar_parquet import collate_samples

rows = load_dataset("GFM-Bench/NA-SAR-X", split="train", streaming=True)
loader = DataLoader(rows, batch_size=2, num_workers=0,
                    collate_fn=partial(collate_samples, verify_sha=True))
batch = next(iter(loader))
print(batch["sample_id"], batch["sar"].shape)       # [2, 8, 256, 256]
print(batch["auxiliary"]["pet_monthly"].shape)     # [2, 12, 256, 256]
print(batch["metadata"][0]["prime_time"])
Downloads last month
105