Each geo-location in the original dataset has 4 seasonal acquisitions. This version pre-selects one season per sample so that training notebooks no longer need to decode 4 seasons at runtime.

Season assignment is deterministic and reproducible:

import hashlib
season = int(hashlib.md5(sample_key.encode()).hexdigest(), 16) % 4

Format

WebDataset .tar shards under train/.
Each sample contains two files:

File	Description
`{key}.bands.b2`	blosc2/zstd-compressed `[14, 264, 264]` float16 array. Channels 0-11 = S2L2A, 12-13 = S1GRD.
`{key}.season.txt`	Single ASCII digit recording which season (0–3) was selected.

Loading a sample

import blosc2, numpy as np, webdataset as wds

N_CHANNELS, H, W = 14, 264, 264

def decode_sample(sample):
    raw = blosc2.decompress(sample['bands.b2'])
    arr = np.frombuffer(raw, dtype=np.float16).reshape(N_CHANNELS, H, W)
    return arr   # [14, 264, 264] float16

dataset = wds.WebDataset('path/to/shard.tar').map(decode_sample)

Source statistics

~250,000 geo-locations across 477 paired shards
14 channels: 12 × Sentinel-2 L2A + 2 × Sentinel-1 GRD
Spatial resolution: 264 × 264 pixels (10 m/px for S2, 10 m/px for S1)

Downloads last month: 79

Total file size:

272 GB