Hermanni/sen12mscr

SEN12MS-CR

Reorganized mirror of the SEN12MS-CR dataset in Parquet format.

Quick Start

from datasets import load_dataset
import numpy as np

ds = load_dataset("Hermanni/sen12mscr", streaming=True)

for sample in ds["train"]:
    sar = np.frombuffer(sample["sar"], dtype=np.float32).reshape(sample["sar_shape"])
    cloudy = np.frombuffer(sample["cloudy"], dtype=np.int16).reshape(sample["opt_shape"])
    target = np.frombuffer(sample["target"], dtype=np.int16).reshape(sample["opt_shape"])

    # Optical tensors are stored as HWC: (256, 256, 13)
    # Convert to CHW if needed:
    # cloudy = np.transpose(cloudy, (2, 0, 1))
    # target = np.transpose(target, (2, 0, 1))
    break

Notes

• sar is stored as float32
• cloudy and target are stored as int16
• opt_shape is stored in HWC order, typically (256, 256, 13)
• The dtype column is a legacy field and should not be used for decoding cloudy or target

Full Download

ds = load_dataset("Hermanni/sen12mscr", split="train")

PyTorch Example

from torch.utils.data import Dataset, DataLoader
from datasets import load_dataset
import numpy as np
import torch

class SEN12MSCR(Dataset):
    def __init__(self, hf_dataset, normalize=True, chw_optical=True):
        self.ds = hf_dataset
        self.normalize = normalize
        self.chw_optical = chw_optical

    def __len__(self):
        return len(self.ds)

    def __getitem__(self, idx):
        s = self.ds[idx]

        sar = np.frombuffer(s["sar"], dtype=np.float32).reshape(s["sar_shape"]).astype(np.float32)
        cloudy = np.frombuffer(s["cloudy"], dtype=np.int16).reshape(s["opt_shape"]).astype(np.float32)
        target = np.frombuffer(s["target"], dtype=np.int16).reshape(s["opt_shape"]).astype(np.float32)

        if self.chw_optical:
            cloudy = np.transpose(cloudy, (2, 0, 1))
            target = np.transpose(target, (2, 0, 1))

        sar = torch.from_numpy(sar.copy())
        cloudy = torch.from_numpy(cloudy.copy())
        target = torch.from_numpy(target.copy())

        if self.normalize:
            cloudy /= 10000.0
            target /= 10000.0

        return {"sar": sar, "cloudy": cloudy, "target": target}

ds = load_dataset("Hermanni/sen12mscr", split="train")
loader = DataLoader(SEN12MSCR(ds), batch_size=8, shuffle=True, num_workers=4)

Contents

• ~122,218 triplets
• SAR: Sentinel-1, 2 channels, float32
• Cloudy: Sentinel-2, 13 channels, int16
• Target: Sentinel-2, 13 channels, int16
• 4 seasons, 175 global ROIs (2018)

Columns

Column	Type	Description
sar	binary	SAR bytes, decode as float32, reshape with sar_shape
cloudy	binary	Cloudy S2 bytes, decode as int16, reshape with opt_shape
target	binary	Cloud-free S2 bytes, decode as int16, reshape with opt_shape
sar_shape	list[int]	SAR shape, typically [2, 256, 256]
opt_shape	list[int]	Optical shape, typically [256, 256, 13]
dtype	string	Legacy field from SAR export; do not use for optical decoding
season	string	spring / summer / fall / winter
scene	string	Scene number
patch	string	Patch ID

License

CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)

Source

• mediaTUM (ID: 1554803) (https://mediatum.ub.tum.de/1554803)