Source code for torchgeo.datasets.ssl4eo_benchmark
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
"""Self-Supervised Learning for Earth Observation Benchmark Datasets."""
import glob
import os
from typing import Callable, Optional
import matplotlib.pyplot as plt
import numpy as np
import rasterio
import torch
from matplotlib.figure import Figure
from torch import Tensor
from .cdl import CDL
from .geo import NonGeoDataset
from .nlcd import NLCD
from .utils import download_url, extract_archive
[docs]class SSL4EOLBenchmark(NonGeoDataset):
"""SSL4EO Landsat Benchmark Evaluation Dataset.
Dataset is intended to be used for evaluation of SSL techniques. Each
benchmark dataset consists of 25,000 images with corresponding land
cover classification masks.
Dataset format:
* Input landsat image and single channel mask
* 25,000 total samples split into train, val, test (70%, 15%, 15%)
* NLCD dataset version has 17 classes
* CDL dataset version has 134 classes
Each patch has the following properties:
* 264 x 264 pixels
* Resampled to 30 m resolution (7920 x 7920 m)
* Single multispectral GeoTIFF file
If you use this dataset in your research, please cite the following paper:
* https://arxiv.org/abs/2306.09424
.. versionadded:: 0.5
"""
url = "https://hf.co/datasets/torchgeo/ssl4eo-l-benchmark/resolve/da96ae2b04cb509710b72fce9131c2a3d5c211c2/{}.tar.gz" # noqa: E501
valid_sensors = ["tm_toa", "etm_toa", "etm_sr", "oli_tirs_toa", "oli_sr"]
valid_products = ["cdl", "nlcd"]
valid_splits = ["train", "val", "test"]
image_root = "ssl4eo_l_{}_benchmark"
img_md5s = {
"tm_toa": "8e3c5bcd56d3780a442f1332013b8d15",
"etm_toa": "1b051c7fe4d61c581b341370c9e76f1f",
"etm_sr": "34a24fa89a801654f8d01e054662c8cd",
"oli_tirs_toa": "6e9d7cf0392e1de2cbdb39962ba591aa",
"oli_sr": "0700cd15cc2366fe68c2f8c02fa09a15",
}
mask_dir_dict = {
"tm_toa": "ssl4eo_l_tm_{}",
"etm_toa": "ssl4eo_l_etm_{}",
"etm_sr": "ssl4eo_l_etm_{}",
"oli_tirs_toa": "ssl4eo_l_oli_{}",
"oli_sr": "ssl4eo_l_oli_{}",
}
mask_md5s = {
"tm": {
"cdl": "3d676770ffb56c7e222a7192a652a846",
"nlcd": "261149d7614fcfdcb3be368eefa825c7",
},
"etm": {
"cdl": "008098c968544049eaf7b307e14241de",
"nlcd": "9c031049d665202ba42ac1d89b687999",
},
"oli": {
"cdl": "1cb057de6eafeca975deb35cb9fb036f",
"nlcd": "9de0d6d4d0b94313b80450f650813922",
},
}
year_dict = {
"tm_toa": 2011,
"etm_toa": 2019,
"etm_sr": 2019,
"oli_tirs_toa": 2019,
"oli_sr": 2019,
}
rgb_indices = {
"tm_toa": [2, 1, 0],
"etm_toa": [2, 1, 0],
"etm_sr": [2, 1, 0],
"oli_tirs_toa": [3, 2, 1],
"oli_sr": [3, 2, 1],
}
split_percentages = [0.7, 0.15, 0.15]
cmaps = {"nlcd": NLCD.cmap, "cdl": CDL.cmap}
[docs] def __init__(
self,
root: str = "data",
sensor: str = "oli_sr",
product: str = "cdl",
split: str = "train",
classes: Optional[list[int]] = None,
transforms: Optional[Callable[[dict[str, Tensor]], dict[str, Tensor]]] = None,
download: bool = False,
checksum: bool = False,
) -> None:
"""Initialize a new SSL4EO Landsat Benchmark instance.
Args:
root: root directory where dataset can be found
sensor: one of ['etm_toa', 'etm_sr', 'oli_tirs_toa, 'oli_sr']
product: mask target, one of ['cdl', 'nlcd']
split: dataset split, one of ['train', 'val', 'test']
classes: list of classes to include, the rest will be mapped to 0
(defaults to all classes for the chosen product)
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 (
sensor in self.valid_sensors
), f"Only supports one of {self.valid_sensors}, but found {sensor}."
self.sensor = sensor
assert (
product in self.valid_products
), f"Only supports one of {self.valid_products}, but found {product}."
self.product = product
assert (
split in self.valid_splits
), f"Only supports one of {self.valid_splits}, but found {split}."
self.split = split
self.cmap = self.cmaps[product]
if classes is None:
classes = list(self.cmap.keys())
assert (
set(classes) <= self.cmap.keys()
), f"Only the following classes are valid: {list(self.cmap.keys())}."
assert 0 in classes, "Classes must include the background class: 0"
self.root = root
self.classes = classes
self.transforms = transforms
self.download = download
self.checksum = checksum
self.ordinal_map = torch.zeros(max(self.cmap.keys()) + 1, dtype=torch.long)
self.ordinal_cmap = torch.zeros((len(self.classes), 4), dtype=torch.uint8)
self.img_dir_name = self.image_root.format(self.sensor)
self.mask_dir_name = self.mask_dir_dict[self.sensor].format(self.product)
self._verify()
self.sample_collection = self.retrieve_sample_collection()
# train, val, test split
np.random.seed(0)
sizes = (np.array(self.split_percentages) * len(self.sample_collection)).astype(
int
)
cutoffs = np.cumsum(sizes)[:-1]
sample_indices = np.arange(len(self.sample_collection))
np.random.shuffle(sample_indices)
groups = np.split(sample_indices, cutoffs)
split_indices = {"train": groups[0], "val": groups[1], "test": groups[2]}[
self.split
]
self.sample_collection = [self.sample_collection[idx] for idx in split_indices]
# Map chosen classes to ordinal numbers, all others mapped to background class
for v, k in enumerate(self.classes):
self.ordinal_map[k] = v
self.ordinal_cmap[v] = torch.tensor(self.cmap[k])
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
img_pathname = os.path.join(self.root, self.img_dir_name, "**", "all_bands.tif")
exists = []
exists.append(bool(glob.glob(img_pathname, recursive=True)))
mask_pathname = os.path.join(
self.root,
self.mask_dir_name,
"**",
f"{self.product}_{self.year_dict[self.sensor]}.tif",
)
exists.append(bool(glob.glob(mask_pathname, recursive=True)))
if all(exists):
return
# Check if the tar.gz files have already been downloaded
exists = []
img_pathname = os.path.join(self.root, f"{self.img_dir_name}.tar.gz")
exists.append(os.path.exists(img_pathname))
mask_pathname = os.path.join(self.root, f"{self.mask_dir_name}.tar.gz")
exists.append(os.path.exists(mask_pathname))
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."""
# download imagery
download_url(
self.url.format(self.img_dir_name),
self.root,
md5=self.img_md5s[self.sensor] if self.checksum else None,
)
# download mask
download_url(
self.url.format(self.mask_dir_name),
self.root,
md5=(
self.mask_md5s[self.sensor.split("_")[0]][self.product]
if self.checksum
else None
),
)
def _extract(self) -> None:
"""Extract the dataset."""
img_pathname = os.path.join(self.root, f"{self.img_dir_name}.tar.gz")
extract_archive(img_pathname)
mask_pathname = os.path.join(self.root, f"{self.mask_dir_name}.tar.gz")
extract_archive(mask_pathname)
[docs] def __getitem__(self, index: int) -> dict[str, Tensor]:
"""Return an index within the dataset.
Args:
index: index to return
Returns:
image and sample
"""
img_path, mask_path = self.sample_collection[index]
sample = {
"image": self._load_image(img_path),
"mask": self._load_mask(mask_path),
}
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.sample_collection)
[docs] def retrieve_sample_collection(self) -> list[tuple[str, str]]:
"""Retrieve paths to samples in data directory."""
img_paths = glob.glob(
os.path.join(self.root, self.img_dir_name, "**", "all_bands.tif"),
recursive=True,
)
img_paths = sorted(img_paths)
sample_collection: list[tuple[str, str]] = []
for img_path in img_paths:
mask_path = img_path.replace(self.img_dir_name, self.mask_dir_name).replace(
"all_bands.tif", f"{self.product}_{self.year_dict[self.sensor]}.tif"
)
sample_collection.append((img_path, mask_path))
return sample_collection
def _load_image(self, path: str) -> Tensor:
"""Load the input image.
Args:
path: path to input image
Returns:
image
"""
with rasterio.open(path) as src:
image = torch.from_numpy(src.read()).float()
return image
def _load_mask(self, path: str) -> Tensor:
"""Load the mask.
Args:
path: path to mask
Retuns:
mask
"""
with rasterio.open(path) as src:
mask = torch.from_numpy(src.read()).long()
mask = self.ordinal_map[mask]
return mask
[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
"""
ncols = 2
image = sample["image"][self.rgb_indices[self.sensor]].permute(1, 2, 0)
image = image / 255
mask = sample["mask"].squeeze(0)
showing_predictions = "prediction" in sample
if showing_predictions:
pred = sample["prediction"].squeeze(0)
ncols = 3
fig, ax = plt.subplots(ncols=ncols, figsize=(4 * ncols, 4))
ax[0].imshow(image)
ax[0].axis("off")
ax[1].imshow(self.ordinal_cmap[mask], interpolation="none")
ax[1].axis("off")
if show_titles:
ax[0].set_title("Image")
ax[1].set_title("Mask")
if showing_predictions:
ax[2].imshow(self.ordinal_cmap[pred], interpolation="none")
if show_titles:
ax[2].set_title("Prediction")
if suptitle is not None:
plt.suptitle(suptitle)
return fig
