Source code for torchgeo.datasets.sen12ms
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
"""SEN12MS dataset."""
import os
from typing import Any, Callable, Dict, List, Optional
import numpy as np
import pytorch_lightning as pl
import rasterio
import torch
from sklearn.model_selection import GroupShuffleSplit
from torch import Tensor
from torch.utils.data import DataLoader, Subset
from .geo import VisionDataset
from .utils import check_integrity
# https://github.com/pytorch/pytorch/issues/60979
# https://github.com/pytorch/pytorch/pull/61045
DataLoader.__module__ = "torch.utils.data"
class SEN12MS(VisionDataset):
"""SEN12MS dataset.
The `SEN12MS <https://doi.org/10.14459/2019mp1474000>`_ dataset contains
180,662 patch triplets of corresponding Sentinel-1 dual-pol SAR data,
Sentinel-2 multi-spectral images, and MODIS-derived land cover maps.
The patches are distributed across the land masses of the Earth and
spread over all four meteorological seasons. This is reflected by the
dataset structure. All patches are provided in the form of 16-bit GeoTiffs
containing the following specific information:
* Sentinel-1 SAR: 2 channels corresponding to sigma nought backscatter
values in dB scale for VV and VH polarization.
* Sentinel-2 Multi-Spectral: 13 channels corresponding to the 13 spectral bands
(B1, B2, B3, B4, B5, B6, B7, B8, B8a, B9, B10, B11, B12).
* MODIS Land Cover: 4 channels corresponding to IGBP, LCCS Land Cover,
LCCS Land Use, and LCCS Surface Hydrology layers.
If you use this dataset in your research, please cite the following paper:
* https://doi.org/10.5194/isprs-annals-IV-2-W7-153-2019
.. note::
This dataset can be automatically downloaded using the following bash script:
.. code-block:: bash
for season in 1158_spring 1868_summer 1970_fall 2017_winter
do
for source in lc s1 s2
do
wget "ftp://m1474000:m1474000@dataserv.ub.tum.de/ROIs${season}_${source}.tar.gz"
tar xvzf "ROIs${season}_${source}.tar.gz"
done
done
for split in train test
do
wget "https://raw.githubusercontent.com/schmitt-muc/SEN12MS/master/splits/${split}_list.txt"
done
or manually downloaded from https://dataserv.ub.tum.de/s/m1474000
and https://github.com/schmitt-muc/SEN12MS/tree/master/splits.
This download will likely take several hours.
""" # noqa: E501
BAND_SETS: Dict[str, List[int]] = {
"all": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14],
"s1": [0, 1],
"s2-all": [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14],
"s2-reduced": [3, 4, 5, 9, 12, 13],
}
filenames = [
"ROIs1158_spring_lc.tar.gz",
"ROIs1158_spring_s1.tar.gz",
"ROIs1158_spring_s2.tar.gz",
"ROIs1868_summer_lc.tar.gz",
"ROIs1868_summer_s1.tar.gz",
"ROIs1868_summer_s2.tar.gz",
"ROIs1970_fall_lc.tar.gz",
"ROIs1970_fall_s1.tar.gz",
"ROIs1970_fall_s2.tar.gz",
"ROIs2017_winter_lc.tar.gz",
"ROIs2017_winter_s1.tar.gz",
"ROIs2017_winter_s2.tar.gz",
"train_list.txt",
"test_list.txt",
]
light_filenames = [
"ROIs1158_spring",
"ROIs1868_summer",
"ROIs1970_fall",
"ROIs2017_winter",
"train_list.txt",
"test_list.txt",
]
md5s = [
"6e2e8fa8b8cba77ddab49fd20ff5c37b",
"fba019bb27a08c1db96b31f718c34d79",
"d58af2c15a16f376eb3308dc9b685af2",
"2c5bd80244440b6f9d54957c6b1f23d4",
"01044b7f58d33570c6b57fec28a3d449",
"4dbaf72ecb704a4794036fe691427ff3",
"9b126a68b0e3af260071b3139cb57cee",
"19132e0aab9d4d6862fd42e8e6760847",
"b8f117818878da86b5f5e06400eb1866",
"0fa0420ef7bcfe4387c7e6fe226dc728",
"bb8cbfc16b95a4f054a3d5380e0130ed",
"3807545661288dcca312c9c538537b63",
"0a68d4e1eb24f128fccdb930000b2546",
"c7faad064001e646445c4c634169484d",
]
[docs] def __init__(
self,
root: str = "data",
split: str = "train",
bands: List[int] = BAND_SETS["all"],
transforms: Optional[Callable[[Dict[str, Tensor]], Dict[str, Tensor]]] = None,
checksum: bool = False,
) -> None:
"""Initialize a new SEN12MS dataset instance.
The ``bands`` argument allows for the subsetting of bands returned by the
dataset. Integers in ``bands`` index into a stack of Sentinel 1 and Sentinel 2
imagery. Indices 0 and 1 correspond to the Sentinel 1 imagery where indices 2
through 14 correspond to the Sentinel 2 imagery.
Args:
root: root directory where dataset can be found
split: one of "train" or "test"
bands: a list of band indices to use where the indices correspond to the
array index of combined Sentinel 1 and Sentinel 2
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 data is not found in ``root``, or checksums don't match
"""
assert split in ["train", "test"]
self.root = root
self.split = split
self.bands = torch.tensor(bands).long() # type: ignore[attr-defined]
self.transforms = transforms
self.checksum = checksum
if checksum:
if not self._check_integrity():
raise RuntimeError("Dataset not found or corrupted.")
else:
if not self._check_integrity_light():
raise RuntimeError("Dataset not found or corrupted.")
with open(os.path.join(self.root, split + "_list.txt")) as f:
self.ids = [line.rstrip() for line in f.readlines()]
[docs] def __getitem__(self, index: int) -> Dict[str, Tensor]:
"""Return an index within the dataset.
Args:
index: index to return
Returns:
data and label at that index
"""
filename = self.ids[index]
lc = self._load_raster(filename, "lc")
s1 = self._load_raster(filename, "s1")
s2 = self._load_raster(filename, "s2")
image = torch.cat(tensors=[s1, s2], dim=0) # type: ignore[attr-defined]
image = torch.index_select( # type: ignore[attr-defined]
image, dim=0, index=self.bands
)
sample: Dict[str, Tensor] = {"image": image, "mask": lc}
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.ids)
def _load_raster(self, filename: str, source: str) -> Tensor:
"""Load a single raster image or target.
Args:
filename: name of the file to load
source: one of "lc", "s1", or "s2"
Returns:
the raster image or target
"""
parts = filename.split("_")
parts[2] = source
with rasterio.open(
os.path.join(
self.root,
"{0}_{1}".format(*parts),
"{2}_{3}".format(*parts),
"{0}_{1}_{2}_{3}_{4}".format(*parts),
)
) as f:
array = f.read().astype(np.int32)
tensor: Tensor = torch.from_numpy(array) # type: ignore[attr-defined]
return tensor
def _check_integrity_light(self) -> bool:
"""Checks the integrity of the dataset structure.
Returns:
True if the dataset directories and split files are found, else False
"""
for filename in self.light_filenames:
filepath = os.path.join(self.root, filename)
if not os.path.exists(filepath):
return False
return True
def _check_integrity(self) -> bool:
"""Check integrity of dataset.
Returns:
True if dataset files are found and/or MD5s match, else False
"""
for filename, md5 in zip(self.filenames, self.md5s):
filepath = os.path.join(self.root, filename)
if not check_integrity(filepath, md5 if self.checksum else None):
return False
return True
class SEN12MSDataModule(pl.LightningDataModule):
"""LightningDataModule implementation for the SEN12MS dataset.
Implements 80/20 geographic train/val splits and uses the test split from the
classification dataset definitions. See :func:`setup` for more details.
Uses the Simplified IGBP scheme defined in the 2020 Data Fusion Competition. See
https://arxiv.org/abs/2002.08254.
"""
#: Mapping from the IGBP class definitions to the DFC2020, taken from the dataloader
#: here https://github.com/lukasliebel/dfc2020_baseline.
DFC2020_CLASS_MAPPING = torch.tensor( # type: ignore[attr-defined]
[
0, # maps 0s to 0
1, # maps 1s to 1
1, # maps 2s to 1
1, # ...
1,
1,
2,
2,
3,
3,
4,
5,
6,
7,
6,
8,
9,
10,
]
)
[docs] def __init__(
self,
root_dir: str,
seed: int,
band_set: str = "all",
batch_size: int = 64,
num_workers: int = 0,
**kwargs: Any,
) -> None:
"""Initialize a LightningDataModule for SEN12MS based DataLoaders.
Args:
root_dir: The ``root`` arugment to pass to the SEN12MS Dataset classes
seed: The seed value to use when doing the sklearn based ShuffleSplit
band_set: The subset of S1/S2 bands to use. Options are: "all",
"s1", "s2-all", and "s2-reduced" where the "s2-reduced" set includes:
B2, B3, B4, B8, B11, and B12.
batch_size: The batch size to use in all created DataLoaders
num_workers: The number of workers to use in all created DataLoaders
"""
super().__init__() # type: ignore[no-untyped-call]
assert band_set in SEN12MS.BAND_SETS.keys()
self.root_dir = root_dir
self.seed = seed
self.band_set = band_set
self.band_indices = SEN12MS.BAND_SETS[band_set]
self.batch_size = batch_size
self.num_workers = num_workers
[docs] def custom_transform(self, sample: Dict[str, Any]) -> Dict[str, Any]:
"""Transform a single sample from the Dataset.
Args:
sample: dictionary containing image and mask
Returns:
preprocessed sample
"""
sample["image"] = sample["image"].float()
if self.band_set == "all":
sample["image"][:2] = sample["image"][:2].clamp(-25, 0) / -25
sample["image"][2:] = sample["image"][2:].clamp(0, 10000) / 10000
elif self.band_set == "s1":
sample["image"][:2] = sample["image"][:2].clamp(-25, 0) / -25
else:
sample["image"][:] = sample["image"][:].clamp(0, 10000) / 10000
sample["mask"] = sample["mask"][0, :, :].long()
sample["mask"] = torch.take( # type: ignore[attr-defined]
self.DFC2020_CLASS_MAPPING, sample["mask"]
)
return sample
[docs] def setup(self, stage: Optional[str] = None) -> None:
"""Create the train/val/test splits based on the original Dataset objects.
The splits should be done here vs. in :func:`__init__` per the docs:
https://pytorch-lightning.readthedocs.io/en/latest/extensions/datamodules.html#setup.
We split samples between train and val geographically with proportions of 80/20.
This mimics the geographic test set split.
Args:
stage: stage to set up
"""
season_to_int = {"winter": 0, "spring": 1000, "summer": 2000, "fall": 3000}
self.all_train_dataset = SEN12MS(
self.root_dir,
split="train",
bands=self.band_indices,
transforms=self.custom_transform,
checksum=False,
)
self.all_test_dataset = SEN12MS(
self.root_dir,
split="test",
bands=self.band_indices,
transforms=self.custom_transform,
checksum=False,
)
# A patch is a filename like: "ROIs{num}_{season}_s2_{scene_id}_p{patch_id}.tif"
# This patch will belong to the scene that is uniquelly identified by its
# (season, scene_id) tuple. Because the largest scene_id is 149, we can simply
# give each season a large number and representing a `unique_scene_id` as
# `season_id + scene_id`.
scenes = []
for scene_fn in self.all_train_dataset.ids:
parts = scene_fn.split("_")
season_id = season_to_int[parts[1]]
scene_id = int(parts[3])
scenes.append(season_id + scene_id)
train_indices, val_indices = next(
GroupShuffleSplit(test_size=0.2, n_splits=2, random_state=self.seed).split(
scenes, groups=scenes
)
)
self.train_dataset = Subset(self.all_train_dataset, train_indices)
self.val_dataset = Subset(self.all_train_dataset, val_indices)
self.test_dataset = Subset(
self.all_test_dataset, range(len(self.all_test_dataset))
)
[docs] def train_dataloader(self) -> DataLoader[Any]:
"""Return a DataLoader for training.
Returns:
training data loader
"""
return DataLoader(
self.train_dataset,
batch_size=self.batch_size,
num_workers=self.num_workers,
shuffle=True,
)
[docs] def val_dataloader(self) -> DataLoader[Any]:
"""Return a DataLoader for validation.
Returns:
validation data loader
"""
return DataLoader(
self.val_dataset,
batch_size=self.batch_size,
num_workers=self.num_workers,
shuffle=False,
)
[docs] def test_dataloader(self) -> DataLoader[Any]:
"""Return a DataLoader for testing.
Returns:
testing data loader
"""
return DataLoader(
self.test_dataset,
batch_size=self.batch_size,
num_workers=self.num_workers,
shuffle=False,
)
