Datasets

terratorch.datasets.biomassters

BioMasstersNonGeo

Bases: BioMassters

BioMassters Dataset for Aboveground Biomass prediction.

Dataset intended for Aboveground Biomass (AGB) prediction over Finnish forests based on Sentinel 1 and 2 data with corresponding target AGB mask values generated by Light Detection and Ranging (LiDAR).

Dataset Format:

• .tif files for Sentinel 1 and 2 data
• .tif file for pixel wise AGB target mask
• .csv files for metadata regarding features and targets

Dataset Features:

• 13,000 target AGB masks of size (256x256px)
• 12 months of data per target mask
• Sentinel 1 and Sentinel 2 data for each location
• Sentinel 1 available for every month
• Sentinel 2 available for almost every month (not available for every month due to ESA acquisition halt over the region during particular periods)

If you use this dataset in your research, please cite the following paper:

• https://nascetti-a.github.io/BioMasster/

.. versionadded:: 0.5

__init__(root='data', split='train', bands=BAND_SETS['all'], transform=None, mask_mean=63.4584, mask_std=72.21242, sensors=['S1', 'S2'], as_time_series=False, metadata_filename=default_metadata_filename, max_cloud_percentage=None, max_red_mean=None, include_corrupt=True, subset=1, seed=42, use_four_frames=False)

Initialize a new instance of BioMassters dataset.

If as_time_series=False (the default), each time step becomes its own sample with the target being shared across multiple samples.

Parameters:

Name	Type	Description	Default
root		root directory where dataset can be found	'data'
split	str	train or test split	'train'
sensors	Sequence[str]	which sensors to consider for the sample, Sentinel 1 and/or Sentinel 2 ('S1', 'S2')	['S1', 'S2']
as_time_series	bool	whether or not to return all available time-steps or just a single one for a given target location	False
metadata_filename	str	metadata file to be used	default_metadata_filename
max_cloud_percentage	float | None	maximum allowed cloud percentage for images	None
max_red_mean	float | None	maximum allowed red_mean value for images	None
include_corrupt	bool	whether to include images marked as corrupted	True

Raises:

Type	Description
AssertionError	if split or sensors is invalid
DatasetNotFoundError	If dataset is not found.

plot(sample, show_titles=True, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	a sample returned by :meth:__getitem__	required
show_titles	bool	flag indicating whether to show titles above each panel	True
suptitle	str | None	optional suptitle to use for figure	None

Returns:

Type	Description
Figure	a matplotlib Figure with the rendered sample

terratorch.datasets.burn_intensity

BurnIntensityNonGeo

Bases: NonGeoDataset

Dataset implementation for Burn Intensity classification.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, use_full_data=True, no_data_replace=0.0001, no_label_replace=-1, use_metadata=False)

Initialize the BurnIntensity dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train' or 'val'.	'train'
bands	Sequence[str]	Bands to output. Defaults to all bands.	BAND_SETS['all']
transform	Optional[Compose]	Albumentations transform to be applied.	None
use_metadata	bool	Whether to return metadata info (location).	False
use_full_data	bool	Wheter to use full data or data with less than 25 percent zeros.	True
no_data_replace	Optional[float]	Value to replace NaNs in images.	0.0001
no_label_replace	Optional[int]	Value to replace NaNs in labels.	-1

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by __getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Any	A matplotlib Figure with the rendered sample.

terratorch.datasets.carbonflux

CarbonFluxNonGeo

Bases: NonGeoDataset

Dataset for Carbon Flux regression from HLS images and MERRA data.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, gpp_mean=None, gpp_std=None, no_data_replace=0.0001, use_metadata=False, modalities=('image', 'merra_vars'))

Initialize the CarbonFluxNonGeo dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	'train' or 'test'.	'train'
bands	Sequence[str]	Bands to use. Defaults to all bands.	BAND_SETS['all']
transform	Optional[Compose]	Albumentations transform to be applied.	None
use_metadata	bool	Whether to return metadata (coordinates and date).	False
merra_means	Sequence[float]	Means for MERRA data normalization.	required
merra_stds	Sequence[float]	Standard deviations for MERRA data normalization.	required
gpp_mean	float	Mean for GPP normalization.	None
gpp_std	float	Standard deviation for GPP normalization.	None
no_data_replace	Optional[float]	Value to replace NO_DATA values in images.	0.0001

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Any]	A sample returned by __getitem__.	required
suptitle	str | None	Optional title for the figure.	None

Returns:

Type	Description
Any	A matplotlib figure with the rendered sample.

terratorch.datasets.forestnet

ForestNetNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for ForestNet.

__init__(data_root, split='train', label_map=default_label_map, transform=None, fraction=1.0, bands=BAND_SETS['all'], use_metadata=False)

Initialize the ForestNetNonGeo dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
label_map	Dict[str, int]	Mapping from label names to integer labels.	default_label_map
transform	Compose | None	Transformations to be applied to the images.	None
fraction	float	Fraction of the dataset to use. Defaults to 1.0 (use all data).	1.0

map_label(index)

Map the label name to an integer label.

terratorch.datasets.fire_scars

FireScarsHLS

Bases: RasterDataset

RasterDataset implementation for fire scars input images.

FireScarsNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for fire scars.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, no_data_replace=0, no_label_replace=-1, use_metadata=False)

Constructor

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
bands	list[str]	Bands that should be output by the dataset. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Should end with ToTensorV2(). If used through the corresponding data module, should not include normalization. Defaults to None, which applies ToTensorV2().	None
no_data_replace	float | None	Replace nan values in input images with this value. If None, does no replacement. Defaults to 0.	0
no_label_replace	int | None	Replace nan values in label with this value. If none, does no replacement. Defaults to -1.	-1
use_metadata	bool	whether to return metadata info (time and location).	False

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	a sample returned by :meth:__getitem__	required
suptitle	str | None	optional string to use as a suptitle	None

Returns:

Type	Description
Figure	a matplotlib Figure with the rendered sample

FireScarsSegmentationMask

Bases: RasterDataset

RasterDataset implementation for fire scars segmentation mask. Can be easily merged with input images using the & operator.

terratorch.datasets.landslide4sense

Landslide4SenseNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for Landslide4Sense.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None)

Initialize the Landslide4Sense dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'validation', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None

terratorch.datasets.m_eurosat

MEuroSATNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-EuroSAT.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default')

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_bigearthnet

MBigEarthNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-BigEarthNet.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default')

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_brick_kiln

MBrickKilnNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-BrickKiln.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default')

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_forestnet

MForestNetNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-ForestNet.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default', use_metadata=False)

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'
use_metadata	bool	Whether to return metadata info (time and location).	False

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_so2sat

MSo2SatNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-So2Sat.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default')

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_pv4ger

MPv4gerNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-PV4GER.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default', use_metadata=False)

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'
use_metadata	bool	Whether to return metadata info (location coordinates).	False

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_cashew_plantation

MBeninSmallHolderCashewsNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-BeninSmallHolderCashews.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default', use_metadata=False)

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'
use_metadata	bool	Whether to return metadata info (time).	False

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_nz_cattle

MNzCattleNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-NZ-Cattle.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default', use_metadata=False)

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'
use_metadata	bool	Whether to return metadata info (time and location).	False

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_chesapeake_landcover

MChesapeakeLandcoverNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-ChesapeakeLandcover.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default')

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_pv4ger_seg

MPv4gerSegNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-PV4GER-SEG.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default', use_metadata=False)

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'
use_metadata	bool	Whether to return metadata info (location coordinates).	False

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_SA_crop_type

MSACropTypeNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-SA-Crop-Type.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, partition='default')

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.m_neontree

MNeonTreeNonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for M-NeonTree.

__init__(data_root, split='train', bands=rgb_bands, transform=None, partition='default')

Initialize the dataset.

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	One of 'train', 'val', or 'test'.	'train'
bands	Sequence[str]	Bands to be used. Defaults to RGB bands.	rgb_bands
transform	Compose | None	Albumentations transform to be applied. Defaults to None, which applies default_transform().	None
partition	str	Partition name for the dataset splits. Defaults to 'default'.	'default'

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	A sample returned by :meth:__getitem__.	required
suptitle	str | None	Optional string to use as a suptitle.	None

Returns:

Type	Description
Figure	matplotlib.figure.Figure: A matplotlib Figure with the rendered sample.

terratorch.datasets.multi_temporal_crop_classification

MultiTemporalCropClassification

Bases: NonGeoDataset

NonGeo dataset implementation for multi-temporal crop classification.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, no_data_replace=None, no_label_replace=None, expand_temporal_dimension=True, reduce_zero_label=True, use_metadata=False, metadata_file_name='chips_df.csv')

Constructor

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	one of 'train' or 'val'.	'train'
bands	list[str]	Bands that should be output by the dataset. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Should end with ToTensorV2(). If used through the corresponding data module, should not include normalization. Defaults to None, which applies ToTensorV2().	None
no_data_replace	float | None	Replace nan values in input images with this value. If None, does no replacement. Defaults to None.	None
no_label_replace	int | None	Replace nan values in label with this value. If none, does no replacement. Defaults to None.	None
expand_temporal_dimension	bool	Go from shape (time*channels, h, w) to (channels, time, h, w). Defaults to True.	True
reduce_zero_label	bool	Subtract 1 from all labels. Useful when labels start from 1 instead of the expected 0. Defaults to True.	True
use_metadata	bool	whether to return metadata info (time and location).	False

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	a sample returned by :meth:__getitem__	required
suptitle	str | None	optional string to use as a suptitle	None

Returns:

Type	Description
Figure	a matplotlib Figure with the rendered sample

terratorch.datasets.open_sentinel_map

OpenSentinelMap

Bases: NonGeoDataset

Pytorch Dataset class to load samples from the OpenSentinelMap dataset, supporting multiple bands and temporal sampling strategies.

__init__(data_root, split='train', bands=None, transform=None, spatial_interpolate_and_stack_temporally=True, pad_image=None, truncate_image=None, target=0, pick_random_pair=True)

Parameters:

Name	Type	Description	Default
data_root	str	Path to the root directory of the dataset.	required
split	str	Dataset split to load. Options are 'train', 'val', or 'test'. Defaults to 'train'.	'train'
bands	list of str	List of band names to load. Defaults to ['gsd_10', 'gsd_20', 'gsd_60'].	None
transform	Compose	Albumentations transformations to apply to the data.	None
spatial_interpolate_and_stack_temporally	bool	If True, the bands are interpolated and concatenated over time. Default is True.	True
pad_image	int	Number of timesteps to pad the time dimension of the image. If None, no padding is applied.	None
truncate_image	int	Number of timesteps to truncate the time dimension of the image. If None, no truncation is performed.	None
target	int	Specifies which target class to use from the mask. Default is 0.	0
pick_random_pair	bool	If True, selects two random images from the temporal sequence. Default is True.	True

terratorch.datasets.openearthmap

OpenEarthMapNonGeo

Bases: NonGeoDataset

OpenEarthMapNonGeo Dataset for non-georeferenced imagery.

This dataset class handles non-georeferenced image data from the OpenEarthMap dataset. It supports configurable band sets and transformations, and performs cropping operations to ensure that the images conform to the required input dimensions. The dataset is split into "train", "test", and "val" subsets based on the provided split parameter.

__init__(data_root, bands=BAND_SETS['all'], transform=None, split='train', crop_size=256, random_crop=True)

Initialize a new instance of the OpenEarthMapNonGeo dataset.

Parameters:

Name	Type	Description	Default
data_root	str	The root directory containing the dataset files.	required
bands	Sequence[str]	A list of band names to be used. Default is BAND_SETS["all"].	BAND_SETS['all']
transform	Compose or None	A transformation pipeline to be applied to the data. If None, a default transform converting the data to a tensor is applied.	None
split	str	The dataset split to use ("train", "test", or "val"). Default is "train".	'train'
crop_size	int	The size (in pixels) of the crop to apply to images. Must be greater than 0. Default is 256.	256
random_crop	bool	If True, performs a random crop; otherwise, performs a center crop. Default is True.	True

Raises:

Type	Description
Exception	If the provided split is not one of "train", "test", or "val".
AssertionError	If crop_size is not greater than 0.

terratorch.datasets.pastis

PASTIS

Bases: NonGeoDataset

" Pytorch Dataset class to load samples from the PASTIS dataset, for semantic and panoptic segmentation.

__init__(data_root, norm=True, target='semantic', folds=None, reference_date='2018-09-01', date_interval=(-200, 600), class_mapping=None, transform=None, truncate_image=None, pad_image=None, satellites=['S2'])

Parameters:

Name	Type	Description	Default
data_root	str	Path to the dataset.	required
norm	bool	If true, images are standardised using pre-computed channel-wise means and standard deviations.	True
reference_date	(str, Format)	'YYYY-MM-DD'): Defines the reference date based on which all observation dates are expressed. Along with the image time series and the target tensor, this dataloader yields the sequence of observation dates (in terms of number of days since the reference date). This sequence of dates is used for instance for the positional encoding in attention based approaches.	'2018-09-01'
target	str	'semantic' or 'instance'. Defines which type of target is returned by the dataloader. * If 'semantic' the target tensor is a tensor containing the class of each pixel. * If 'instance' the target tensor is the concatenation of several signals, necessary to train the Parcel-as-Points module: - the centerness heatmap, - the instance ids, - the voronoi partitioning of the patch with regards to the parcels' centers, - the (height, width) size of each parcel, - the semantic label of each parcel, - the semantic label of each pixel.	'semantic'
folds	list	List of ints specifying which of the 5 official folds to load. By default (when None is specified), all folds are loaded.	None
class_mapping	dict	A dictionary to define a mapping between the default 18 class nomenclature and another class grouping. If not provided, the default class mapping is used.	None
transform	callable	A transform to apply to the loaded data (images, dates, and masks). By default, no transformation is applied.	None
truncate_image	int	Truncate the time dimension of the image to a specified number of timesteps. If None, no truncation is performed.	None
pad_image	int	Pad the time dimension of the image to a specified number of timesteps. If None, no padding is applied.	None
satellites	list	Defines the satellites to use. If you are using PASTIS-R, you have access to Sentinel-2 imagery and Sentinel-1 observations in Ascending and Descending orbits, respectively S2, S1A, and S1D. For example, use satellites=['S2', 'S1A'] for Sentinel-2 + Sentinel-1 ascending time series, or satellites=['S2', 'S1A', 'S1D'] to retrieve all time series. If you are using PASTIS, only S2 observations are available.	['S2']

terratorch.datasets.sen1floods11

Sen1Floods11NonGeo

Bases: NonGeoDataset

NonGeo dataset implementation for sen1floods11.

__init__(data_root, split='train', bands=BAND_SETS['all'], transform=None, constant_scale=0.0001, no_data_replace=0, no_label_replace=-1, use_metadata=False)

Constructor

Parameters:

Name	Type	Description	Default
data_root	str	Path to the data root directory.	required
split	str	one of 'train', 'val' or 'test'.	'train'
bands	list[str]	Bands that should be output by the dataset. Defaults to all bands.	BAND_SETS['all']
transform	Compose | None	Albumentations transform to be applied. Should end with ToTensorV2(). Defaults to None, which applies ToTensorV2().	None
constant_scale	float	Factor to multiply image values by. Defaults to 0.0001.	0.0001
no_data_replace	float | None	Replace nan values in input images with this value. If None, does no replacement. Defaults to 0.	0
no_label_replace	int | None	Replace nan values in label with this value. If none, does no replacement. Defaults to -1.	-1
use_metadata	bool	whether to return metadata info (time and location).	False

plot(sample, suptitle=None)

Plot a sample from the dataset.

Parameters:

Name	Type	Description	Default
sample	dict[str, Tensor]	a sample returned by :meth:__getitem__	required
suptitle	str | None	optional string to use as a suptitle	None

Returns:

Type	Description
Figure	a matplotlib Figure with the rendered sample

terratorch.datasets.sen4agrinet

Sen4AgriNet

Bases: NonGeoDataset

__init__(data_root, bands=None, scenario='random', split='train', transform=None, truncate_image=4, pad_image=4, spatial_interpolate_and_stack_temporally=True, seed=42)

Pytorch Dataset class to load samples from the Sen4AgriNet dataset, supporting multiple scenarios for splitting the data.

Parameters:

Name	Type	Description	Default
data_root	str	Root directory of the dataset.	required
bands	list of str	List of band names to load. Defaults to all available bands.	None
scenario	str	Defines the splitting scenario to use. Options are: - 'random': Random split of the data. - 'spatial': Split by geographical regions (Catalonia and France). - 'spatio-temporal': Split by region and year (France 2019 and Catalonia 2020).	'random'
split	str	Specifies the dataset split. Options are 'train', 'val', or 'test'.	'train'
transform	Compose	Albumentations transformations to apply to the data.	None
truncate_image	int	Number of timesteps to truncate the time dimension of the image. If None, no truncation is applied. Default is 4.	4
pad_image	int	Number of timesteps to pad the time dimension of the image. If None, no padding is applied. Default is 4.	4
spatial_interpolate_and_stack_temporally	bool	Whether to interpolate bands and concatenate them over time	True
seed	int	Random seed used for data splitting.	42

terratorch.datasets.sen4map

Sen4MapDatasetMonthlyComposites

Bases: Dataset

Sen4Map Dataset for Monthly Composites.

Dataset intended for land-cover and crop classification tasks based on monthly composites derived from multi-temporal satellite data stored in HDF5 files.

Dataset Format:

• HDF5 files containing multi-temporal acquisitions with spectral bands (e.g., B2, B3, …, B12)
• Composite images computed as the median across available acquisitions for each month.
• Classification labels provided via HDF5 attributes (e.g., 'lc1') with mappings defined for:
  • Land-cover: using land_cover_classification_map
  • Crops: using crop_classification_map

Dataset Features:

• Supports two classification tasks: "land-cover" (default) and "crops".
• Pre-processing options include center cropping, reverse tiling, and resizing.
• Option to save the keys HDF5 for later filtering.
• Input channel selection via a mapping between available bands and input bands.

__init__(h5py_file_object, h5data_keys=None, crop_size=None, dataset_bands=None, input_bands=None, resize=False, resize_to=[224, 224], resize_interpolation=InterpolationMode.BILINEAR, resize_antialiasing=True, reverse_tile=False, reverse_tile_size=3, save_keys_path=None, classification_map='land-cover')

Initialize a new instance of Sen4MapDatasetMonthlyComposites.

This dataset loads data from an HDF5 file object containing multi-temporal satellite data and computes monthly composite images by aggregating acquisitions (via median).

Parameters:

Name	Type	Description	Default
h5py_file_object	File	An open h5py.File object containing the dataset.	required
h5data_keys		Optional list of keys to select a subset of data samples from the HDF5 file. If None, all keys are used.	None
crop_size	None | int	Optional integer specifying the square crop size for the output image.	None
dataset_bands	list[HLSBands | int] | None	Optional list of bands available in the dataset.	None
input_bands	list[HLSBands | int] | None	Optional list of bands to be used as input channels. Must be provided along with dataset_bands.	None
resize		Boolean flag indicating whether the image should be resized. Default is False.	False
resize_to		Target dimensions [height, width] for resizing. Default is [224, 224].	[224, 224]
resize_interpolation		Interpolation mode used for resizing. Default is InterpolationMode.BILINEAR.	BILINEAR
resize_antialiasing		Boolean flag to apply antialiasing during resizing. Default is True.	True
reverse_tile		Boolean flag indicating whether to apply reverse tiling to the image. Default is False.	False
reverse_tile_size		Kernel size for the reverse tiling operation. Must be an odd number >= 3. Default is 3.	3
save_keys_path		Optional file path to save the list of dataset keys.	None
classification_map		String specifying the classification mapping to use ("land-cover" or "crops"). Default is "land-cover".	'land-cover'

Raises:

Type	Description
ValueError	If input_bands is provided without specifying dataset_bands.
ValueError	If an invalid classification_map is provided.

reverse_tiling_pytorch(img_tensor, kernel_size=3)

Upscales an image where every pixel is expanded into kernel_size*kernel_size pixels. Used to test whether the benefit of resizing images to the pre-trained size comes from the bilnearly interpolated pixels, or if the same would be realized with no interpolated pixels.

---

Last update: March 23, 2025