Source code for torchgeo.datasets.ssl4eo
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
"""Self-Supervised Learning for Earth Observation."""
import glob
import os
import random
from typing import Callable, Optional, TypedDict
import matplotlib.pyplot as plt
import numpy as np
import rasterio
import torch
from matplotlib.figure import Figure
from torch import Tensor
from .geo import NonGeoDataset
from .utils import check_integrity, download_url, extract_archive
[docs]class SSL4EO(NonGeoDataset):
"""Base class for all SSL4EO datasets.
Self-Supervised Learning for Earth Observation (SSL4EO) is a collection of
large-scale multimodal multitemporal datasets for unsupervised/self-supervised
pre-training in Earth observation.
.. versionadded:: 0.5
"""
[docs]class SSL4EOL(NonGeoDataset):
"""SSL4EO-L dataset.
Landsat version of SSL4EO.
The dataset consists of a parallel corpus (same locations and dates for SR/TOA)
for the following sensors:
.. list-table::
:widths: 10 10 10 10 10
:header-rows: 1
* - Satellites
- Sensors
- Level
- # Bands
- Link
* - Landsat 4--5
- TM
- TOA
- 7
- `GEE <https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LT05_C02_T1_TOA>`__
* - Landsat 7
- ETM+
- SR
- 6
- `GEE <https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LT05_C02_T1_L2>`__
* - Landsat 7
- ETM+
- TOA
- 9
- `GEE <https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LE07_C02_T1_TOA>`__
* - Landsat 8--9
- OLI+TIRS
- TOA
- 11
- `GEE <https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LC08_C02_T1_TOA>`__
* - Landsat 8--9
- OLI
- SR
- 7
- `GEE <https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LC08_C02_T1_L2>`__
Each patch has the following properties:
* 264 x 264 pixels
* Resampled to 30 m resolution (7920 x 7920 m)
* Single multispectral GeoTIFF file
.. note::
Each split is 300--400 GB and requires 3x that to concatenate and extract
tarballs. Tarballs can be safely deleted after extraction to save space.
The dataset takes about 1.5 hrs to download and checksum and another 3 hrs
to extract.
If you use this dataset in your research, please cite the following paper:
* https://arxiv.org/abs/2306.09424
.. versionadded:: 0.5
""" # noqa: E501
class _Metadata(TypedDict):
num_bands: int
rgb_bands: list[int]
metadata: dict[str, _Metadata] = {
"tm_toa": {"num_bands": 7, "rgb_bands": [2, 1, 0]},
"etm_toa": {"num_bands": 9, "rgb_bands": [2, 1, 0]},
"etm_sr": {"num_bands": 6, "rgb_bands": [2, 1, 0]},
"oli_tirs_toa": {"num_bands": 11, "rgb_bands": [3, 2, 1]},
"oli_sr": {"num_bands": 7, "rgb_bands": [3, 2, 1]},
}
url = "https://hf.co/datasets/torchgeo/ssl4eo_l/resolve/e2467887e6a6bcd7547d9d5999f8d9bc3323dc31/{0}/ssl4eo_l_{0}.tar.gz{1}" # noqa: E501
checksums = {
"tm_toa": {
"aa": "553795b8d73aa253445b1e67c5b81f11",
"ab": "e9e0739b5171b37d16086cb89ab370e8",
"ac": "6cb27189f6abe500c67343bfcab2432c",
"ad": "15a885d4f544d0c1849523f689e27402",
"ae": "35523336bf9f8132f38ff86413dcd6dc",
"af": "fa1108436034e6222d153586861f663b",
"ag": "d5c91301c115c00acaf01ceb3b78c0fe",
},
"etm_toa": {
"aa": "587c3efc7d0a0c493dfb36139d91ccdf",
"ab": "ec34f33face893d2d8fd152496e1df05",
"ac": "947acc2c6bc3c1d1415ac92bab695380",
"ad": "e31273dec921e187f5c0dc73af5b6102",
"ae": "43390a47d138593095e9a6775ae7dc75",
"af": "082881464ca6dcbaa585f72de1ac14fd",
"ag": "de2511aaebd640bd5e5404c40d7494cb",
"ah": "124c5fbcda6871f27524ae59480dabc5",
"ai": "12b5f94824b7f102df30a63b1139fc57",
},
"etm_sr": {
"aa": "baa36a9b8e42e234bb44ab4046f8f2ac",
"ab": "9fb0f948c76154caabe086d2d0008fdf",
"ac": "99a55367178373805d357a096d68e418",
"ad": "59d53a643b9e28911246d4609744ef25",
"ae": "7abfcfc57528cb9c619c66ee307a2cc9",
"af": "bb23cf26cc9fe156e7a68589ec69f43e",
"ag": "97347e5a81d24c93cf33d99bb46a5b91",
},
"oli_tirs_toa": {
"aa": "4711369b861c856ebfadbc861e928d3a",
"ab": "660a96cda1caf54df837c4b3c6c703f6",
"ac": "c9b6a1117916ba318ac3e310447c60dc",
"ad": "b8502e9e92d4a7765a287d21d7c9146c",
"ae": "5c11c14cfe45f78de4f6d6faf03f3146",
"af": "5b0ed3901be1000137ddd3a6d58d5109",
"ag": "a3b6734f8fe6763dcf311c9464a05d5b",
"ah": "5e55f92e3238a8ab3e471be041f8111b",
"ai": "e20617f73d0232a0c0472ce336d4c92f",
},
"oli_sr": {
"aa": "ca338511c9da4dcbfddda28b38ca9e0a",
"ab": "7f4100aa9791156958dccf1bb2a88ae0",
"ac": "6b0f18be2b63ba9da194cc7886dbbc01",
"ad": "57efbcc894d8da8c4975c29437d8b775",
"ae": "2594a0a856897f3f5a902c830186872d",
"af": "a03839311a2b3dc17dfb9fb9bc4f9751",
"ag": "6a329d8fd9fdd591e400ab20f9d11dea",
},
}
[docs] def __init__(
self,
root: str = "data",
split: str = "oli_sr",
seasons: int = 1,
transforms: Optional[Callable[[dict[str, Tensor]], dict[str, Tensor]]] = None,
download: bool = False,
checksum: bool = False,
) -> None:
"""Initialize a new SSL4EOL instance.
Args:
root: root directory where dataset can be found
split: one of ['tm_toa', 'etm_toa', 'etm_sr', 'oli_tirs_toa', 'oli_sr']
seasons: number of seasonal patches to sample per location, 1--4
transforms: a function/transform that takes input sample and its target as
entry and returns a transformed version
download: if True, download dataset and store it in the root directory
checksum: if True, check the MD5 after downloading files (may be slow)
Raises:
AssertionError: if any arguments are invalid
RuntimeError: if ``download=False`` but dataset is missing or checksum fails
"""
assert split in self.metadata
assert seasons in range(1, 5)
self.root = root
self.subdir = os.path.join(root, f"ssl4eo_l_{split}")
self.split = split
self.seasons = seasons
self.transforms = transforms
self.download = download
self.checksum = checksum
self._verify()
self.scenes = sorted(os.listdir(self.subdir))
[docs] def __getitem__(self, index: int) -> dict[str, Tensor]:
"""Return an index within the dataset.
Args:
index: index to return
Returns:
image sample
"""
root = os.path.join(self.subdir, self.scenes[index])
subdirs = os.listdir(root)
subdirs = random.sample(subdirs, self.seasons)
images = []
for subdir in subdirs:
directory = os.path.join(root, subdir)
filename = os.path.join(directory, "all_bands.tif")
with rasterio.open(filename) as f:
image = f.read()
images.append(torch.from_numpy(image.astype(np.float32)))
sample = {"image": torch.cat(images)}
if self.transforms is not None:
sample = self.transforms(sample)
return sample
[docs] def __len__(self) -> int:
"""Return the number of data points in the dataset.
Returns:
length of the dataset
"""
return len(self.scenes)
def _verify(self) -> None:
"""Verify the integrity of the dataset.
Raises:
RuntimeError: if ``download=False`` but dataset is missing or checksum fails
"""
# Check if the extracted files already exist
path = os.path.join(self.subdir, "00000*", "*", "all_bands.tif")
if glob.glob(path):
return
# Check if the tar.gz files have already been downloaded
exists = []
for suffix in self.checksums[self.split]:
path = self.subdir + f".tar.gz{suffix}"
exists.append(os.path.exists(path))
if all(exists):
self._extract()
return
# Check if the user requested to download the dataset
if not self.download:
raise RuntimeError(
f"Dataset not found in `root={self.root}` and `download=False`, "
"either specify a different `root` directory or use `download=True` "
"to automatically download the dataset."
)
# Download the dataset
self._download()
self._extract()
def _download(self) -> None:
"""Download the dataset."""
for suffix, md5 in self.checksums[self.split].items():
download_url(
self.url.format(self.split, suffix),
self.root,
md5=md5 if self.checksum else None,
)
def _extract(self) -> None:
"""Extract the dataset."""
# Concatenate all tarballs together
chunk_size = 2**15 # same as torchvision
path = self.subdir + ".tar.gz"
with open(path, "wb") as f:
for suffix in self.checksums[self.split]:
with open(path + suffix, "rb") as g:
while chunk := g.read(chunk_size):
f.write(chunk)
# Extract the concatenated tarball
extract_archive(path)
[docs] def plot(
self,
sample: dict[str, Tensor],
show_titles: bool = True,
suptitle: Optional[str] = None,
) -> Figure:
"""Plot a sample from the dataset.
Args:
sample: a sample returned by :meth:`__getitem__`
show_titles: flag indicating whether to show titles above each panel
suptitle: optional string to use as a suptitle
Returns:
a matplotlib Figure with the rendered sample
"""
fig, axes = plt.subplots(
ncols=self.seasons, squeeze=False, figsize=(4 * self.seasons, 4)
)
num_bands = self.metadata[self.split]["num_bands"]
rgb_bands = self.metadata[self.split]["rgb_bands"]
for i in range(self.seasons):
image = sample["image"][i * num_bands : (i + 1) * num_bands].byte()
image = image[rgb_bands].permute(1, 2, 0)
axes[0, i].imshow(image)
axes[0, i].axis("off")
if show_titles:
axes[0, i].set_title(f"Split {self.split}, Season {i + 1}")
if suptitle is not None:
plt.suptitle(suptitle)
return fig
[docs]class SSL4EOS12(NonGeoDataset):
"""SSL4EO-S12 dataset.
`Sentinel-1/2 <https://github.com/zhu-xlab/SSL4EO-S12>`_ version of SSL4EO.
The dataset consists of unlabeled patch triplets (Sentinel-1 dual-pol SAR,
Sentinel-2 top-of-atmosphere multispectral, Sentinel-2 surface reflectance
multispectral) from 251079 locations across the globe, each patch covering
2640mx2640m and including four seasonal time stamps.
If you use this dataset in your research, please cite the following paper:
* https://arxiv.org/abs/2211.07044
.. note::
This dataset can be downloaded using:
.. code-block:: console
$ export RSYNC_PASSWORD=m1660427.001
$ rsync -av rsync://m1660427.001@dataserv.ub.tum.de/m1660427.001/ .
The dataset is about 1.5 TB when compressed and 3.7 TB when uncompressed, and
takes roughly 36 hrs to download, 1 hr to checksum, and 12 hrs to extract.
.. versionadded:: 0.5
"""
size = 264
class _Metadata(TypedDict):
filename: str
md5: str
bands: list[str]
metadata: dict[str, _Metadata] = {
"s1": {
"filename": "s1.tar.gz",
"md5": "51ee23b33eb0a2f920bda25225072f3a",
"bands": ["VV", "VH"],
},
"s2c": {
"filename": "s2_l1c.tar.gz",
"md5": "b4f8b03c365e4a85780ded600b7497ab",
"bands": [
"B1",
"B2",
"B3",
"B4",
"B5",
"B6",
"B7",
"B8",
"B8A",
"B9",
"B10",
"B11",
"B12",
],
},
"s2a": {
"filename": "s2_l2a.tar.gz",
"md5": "85496cd9d6742aee03b6a1c99cee0ac1",
"bands": [
"B1",
"B2",
"B3",
"B4",
"B5",
"B6",
"B7",
"B8",
"B8A",
"B9",
"B11",
"B12",
],
},
}
[docs] def __init__(
self,
root: str = "data",
split: str = "s2c",
seasons: int = 1,
transforms: Optional[Callable[[dict[str, Tensor]], dict[str, Tensor]]] = None,
checksum: bool = False,
) -> None:
"""Initialize a new SSL4EOS12 instance.
Args:
root: root directory where dataset can be found
split: one of "s1" (Sentinel-1 dual-pol SAR), "s2c" (Sentinel-2 Level-1C
top-of-atmosphere reflectance), and "s2a" (Sentinel-2 Level-2a surface
reflectance)
seasons: number of seasonal patches to sample per location, 1--4
transforms: a function/transform that takes input sample and its target as
entry and returns a transformed version
checksum: if True, check the MD5 of the downloaded files (may be slow)
Raises:
AssertionError: if ``split`` argument is invalid
RuntimeError: if dataset is missing or checksum fails
"""
assert split in self.metadata
assert seasons in range(1, 5)
self.root = root
self.split = split
self.seasons = seasons
self.transforms = transforms
self.checksum = checksum
self.bands = self.metadata[self.split]["bands"]
self._verify()
[docs] def __getitem__(self, index: int) -> dict[str, Tensor]:
"""Return an index within the dataset.
Args:
index: index to return
Returns:
image sample
"""
root = os.path.join(self.root, self.split, f"{index:07}")
subdirs = os.listdir(root)
subdirs = random.sample(subdirs, self.seasons)
images = []
for subdir in subdirs:
directory = os.path.join(root, subdir)
for band in self.bands:
filename = os.path.join(directory, f"{band}.tif")
with rasterio.open(filename) as f:
image = f.read(out_shape=(1, self.size, self.size))
images.append(torch.from_numpy(image.astype(np.float32)))
sample = {"image": torch.cat(images)}
if self.transforms is not None:
sample = self.transforms(sample)
return sample
[docs] def __len__(self) -> int:
"""Return the number of data points in the dataset.
Returns:
length of the dataset
"""
return 251079
def _verify(self) -> None:
"""Verify the integrity of the dataset.
Raises:
RuntimeError: if dataset is missing or checksum fails
"""
# Check if the extracted files already exist
directory_path = os.path.join(self.root, self.split)
if os.path.exists(directory_path):
return
# Check if the zip files have already been downloaded
filename = self.metadata[self.split]["filename"]
zip_path = os.path.join(self.root, filename)
md5 = self.metadata[self.split]["md5"] if self.checksum else None
integrity = check_integrity(zip_path, md5)
if integrity:
self._extract()
else:
raise RuntimeError(f"Dataset not found in `root={self.root}`")
def _extract(self) -> None:
"""Extract the dataset."""
filename = self.metadata[self.split]["filename"]
extract_archive(os.path.join(self.root, filename))
[docs] def plot(
self,
sample: dict[str, Tensor],
show_titles: bool = True,
suptitle: Optional[str] = None,
) -> Figure:
"""Plot a sample from the dataset.
Args:
sample: a sample returned by :meth:`__getitem__`
show_titles: flag indicating whether to show titles above each panel
suptitle: optional string to use as a suptitle
Returns:
a matplotlib Figure with the rendered sample
"""
nrows = 2 if self.split == "s1" else 1
fig, axes = plt.subplots(
nrows=nrows,
ncols=self.seasons,
squeeze=False,
figsize=(4 * self.seasons, 4 * nrows),
)
for i in range(self.seasons):
image = sample["image"][i * len(self.bands) : (i + 1) * len(self.bands)]
if self.split == "s1":
axes[0, i].imshow(image[0])
axes[1, i].imshow(image[1])
else:
image = image[[3, 2, 1]].permute(1, 2, 0)
image = torch.clamp(image / 3000, min=0, max=1)
axes[0, i].imshow(image)
axes[0, i].axis("off")
if show_titles:
axes[0, i].set_title(f"Split {self.split}, Season {i + 1}")
if suptitle is not None:
plt.suptitle(suptitle)
return fig
