Non-Geospatial Datasets

Non-geospatial datasets are pre-defined benchmark datasets without coordinate information. They are designed for standard computer vision tasks using PyTorch’s DataLoader.

Base Classes

NonGeoDataset

Abstract base class for datasets lacking geospatial information.

from torchgeo.datasets import NonGeoDataset

Key Features:

Integer indexing (not spatial queries)
Pre-defined image chips
Compatible with PyTorch DataLoader
Suitable for benchmarking and competitions

Methods:

__getitem__

(index: int) -> Sample

Retrieve sample by integer index: dataset[0]

__len__

() -> int

Return number of samples in dataset

Example:

from torch.utils.data import DataLoader

dataset = MyNonGeoDataset(root='data', split='train')
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

for sample in dataloader:
    image = sample['image']  # [B, C, H, W]
    label = sample['label']  # [B]

NonGeoClassificationDataset

Base class for classification datasets organized in folder structure (one folder per class).

from torchgeo.datasets import NonGeoClassificationDataset

Constructor:

root

str | PathLike

default:"'data'"

Root directory containing class subdirectories

transforms

Callable[[Sample], Sample] | None

default:"None"

Transform function applied to each sample

loader

Callable[[str], Any]

default:"pil_loader"

Function to load images from file paths

is_valid_file

Callable[[str], bool] | None

default:"None"

Function to filter valid image files

Expected Directory Structure:

root/
├── class_1/
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ...
├── class_2/
│   ├── image1.jpg
│   └── ...
└── ...

Output Format:

{
    'image': Tensor,  # [C, H, W]
    'label': Tensor   # Scalar class index
}

Common Parameters

Most non-geospatial datasets support these parameters:

root

str | PathLike

default:"'data'"

Root directory where dataset is stored or will be downloaded

split

str

default:"'train'"

Dataset split: typically ‘train’, ‘val’, or ‘test’

transforms

Callable[[Sample], Sample] | None

default:"None"

Transform function for data augmentation

download

bool

default:"False"

If True, download the dataset if not found locally

checksum

bool

default:"False"

If True, verify file integrity using MD5 checksums

Classification Datasets

Benchmark datasets for image classification tasks.

EuroSAT

Sentinel-2 satellite imagery for land use classification.

from torchgeo.datasets import EuroSAT

dataset = EuroSAT(
    root='data/eurosat',
    split='train',
    download=True
)

Properties:

27,000 labeled images
13 spectral bands (Sentinel-2)
64x64 pixel images
10 classes: Annual Crop, Forest, Herbaceous Vegetation, Highway, Industrial, Pasture, Permanent Crop, Residential, River, Sea/Lake

Variants:

EuroSAT: Full 13-band version
EuroSAT100: 100-sample subset for testing
EuroSATSpatial: Spatial split variant

RESISC45

High-resolution remote sensing image scene classification.

from torchgeo.datasets import RESISC45

dataset = RESISC45(
    root='data/resisc45',
    split='train',
    download=True
)

Properties:

31,500 images
RGB aerial imagery
256x256 pixels
45 scene classes (airplane, beach, bridge, forest, etc.)

UC Merced

Land use classification from high-resolution imagery.

from torchgeo.datasets import UCMerced

dataset = UCMerced(
    root='data/ucmerced',
    download=True
)

Properties:

2,100 images
RGB aerial imagery
256x256 pixels
21 classes (agricultural, airplane, baseball diamond, beach, etc.)

Additional Classification Datasets

Satellite Imagery:

BigEarthNet: Large-scale Sentinel-2 benchmark (590,326 patches)
BigEarthNetV2: Updated version with corrected labels
So2Sat: Sentinel-1 & Sentinel-2 for local climate zones
PatternNet: 38-class remote sensing scene classification
MillionAID: Large-scale scene classification (1M images)

Aerial Imagery:

ADVANCE: Aerial scene understanding
AID: Aerial Image Dataset (10,000 images, 30 classes)

Specialized:

CV4AKenyaCropType: Crop type classification (Kenya)
SouthAfricaCropType: Crop type classification (South Africa)
SSL4EOL: Self-supervised learning benchmark
TreeSatAI: Tree species classification

Segmentation Datasets

Benchmark datasets for semantic segmentation tasks.

Inria Aerial Image Labeling

Building segmentation from aerial imagery.

from torchgeo.datasets import InriaAerialImageLabeling

dataset = InriaAerialImageLabeling(
    root='data/inria',
    split='train',
    download=True
)

Properties:

RGB aerial imagery + binary building masks
5000x5000 pixel tiles
180 cities across the world
Binary segmentation (building/background)

Output Format:

{
    'image': Tensor,  # [3, H, W]
    'mask': Tensor    # [H, W] - binary mask
}

DeepGlobe Land Cover

Multi-class land cover segmentation.

from torchgeo.datasets import DeepGlobeLandCover

dataset = DeepGlobeLandCover(
    root='data/deepglobe',
    split='train',
    download=True
)

Properties:

RGB satellite imagery
2448x2448 pixels
7 classes: Urban, Agriculture, Rangeland, Forest, Water, Barren, Unknown

LoveDA

Urban/rural scene segmentation.

from torchgeo.datasets import LoveDA

dataset = LoveDA(
    root='data/loveda',
    split='train',
    scene=['urban', 'rural'],
    download=True
)

Properties:

RGB imagery
1024x1024 pixels
7 classes: Background, Building, Road, Water, Barren, Forest, Agriculture
Urban and rural scenes

Potsdam2D

High-resolution urban segmentation.

from torchgeo.datasets import Potsdam2D

dataset = Potsdam2D(
    root='data/potsdam',
    split='train'
)

Properties:

RGB + NIR imagery
6000x6000 pixel tiles
6 classes: Impervious surfaces, Building, Low vegetation, Tree, Car, Background

Vaihingen2D

High-resolution urban segmentation.

from torchgeo.datasets import Vaihingen2D

dataset = Vaihingen2D(
    root='data/vaihingen',
    split='train'
)

Properties:

RGB + NIR + DSM
Variable image sizes
6 classes: Same as Potsdam2D

Additional Segmentation Datasets

Urban/Building:

Chesapeake: Land cover for Chesapeake Bay watershed
ChesapeakeCVPR: Competition variant
LandCoverAI: Building/woodland/water/road segmentation
SpaceNet: Building footprint extraction (multiple challenges)

Agricultural:

PASTIS: Panoptic segmentation of satellite image time series
AgriFieldNet: Agricultural field boundary delineation
FieldsOfTheWorld: Global field boundary dataset

Change Detection:

OSCD: Onera Satellite Change Detection
LEVIRCD: Building change detection
xBD / XView2: Building damage assessment

Specialized:

SEN12MS: Multi-modal (Sentinel-1/2, Landsat 8) segmentation
DFC2022: Data Fusion Contest 2022
EnviroAtlas: Multi-label land cover
GID15: Large-scale land cover

Object Detection Datasets

Benchmark datasets for object detection tasks.

VHR-10

Very high-resolution object detection.

from torchgeo.datasets import VHR10

dataset = VHR10(
    root='data/vhr10',
    download=True
)

Properties:

800 RGB images
10 classes: Airplane, Ship, Storage tank, Baseball diamond, Tennis court, Basketball court, Ground track field, Harbor, Bridge, Vehicle
Bounding box annotations

Output Format:

{
    'image': Tensor,      # [3, H, W]
    'bbox_xyxy': Tensor,  # [N, 4] - boxes in (xmin, ymin, xmax, ymax)
    'label': Tensor       # [N] - class indices
}

DIOR

Object detection in optical remote sensing images.

from torchgeo.datasets import DIOR

dataset = DIOR(
    root='data/dior',
    split='train'
)

Properties:

23,463 images
RGB imagery
800x800 pixels
20 object categories

DOTA

Oriented object detection in aerial images.

from torchgeo.datasets import DOTA

dataset = DOTA(
    root='data/dota',
    split='train',
    version='v1.0'
)

Properties:

Large-scale dataset
Oriented bounding boxes
15-18 categories (depending on version)
Very high resolution

Additional Detection Datasets

Object Detection:

FAIR1M: Fine-grained object recognition (1M instances, 37 categories)
COWC: Cars Overhead With Context
NASAMarineDebris: Marine debris detection
xView: 1M objects, 60 classes

Specialized:

RwandaFieldBoundary: Instance segmentation of crop fields
IDTReeS: Individual tree crown detection

Multi-Task Datasets

Datasets supporting multiple task types.

BigEarthNet

Large-scale multi-label classification.

from torchgeo.datasets import BigEarthNet, BigEarthNetV2

dataset = BigEarthNetV2(
    root='data/bigearthnet',
    split='train',
    bands='s2',  # or 's1' or 'all'
    download=True
)

Properties:

590,326 Sentinel-1 and Sentinel-2 patches
Multi-label classification (19 classes)
120x120m patches
Europe coverage

SEN12MS

Multi-modal semantic segmentation.

from torchgeo.datasets import SEN12MS

dataset = SEN12MS(
    root='data/sen12ms',
    split='train',
    sensors=['s1', 's2']  # Sentinel-1, Sentinel-2
)

Properties:

180,662 triplets (Sentinel-1, Sentinel-2, Landsat 8)
Global coverage
Land cover classification
Multi-modal learning

BioMassters

Biomass estimation from satellite imagery.

from torchgeo.datasets import BioMassters

dataset = BioMassters(
    root='data/biomassters',
    split='train'
)

Properties:

Sentinel-1 & Sentinel-2 time series
Above-ground biomass regression
Competition dataset

Temporal/Time Series Datasets

Datasets with temporal sequences.

PASTIS

Panoptic segmentation of satellite time series.

from torchgeo.datasets import PASTIS

dataset = PASTIS(
    root='data/pastis',
    split='train',
    download=True
)

Properties:

Sentinel-2 time series
2,433 parcels
18 crop classes
Temporal semantic segmentation

CropHarvest

Global crop type mapping.

from torchgeo.datasets import CropHarvest

dataset = CropHarvest(
    root='data/cropharvest',
    download=True
)

Properties:

Satellite time series
Global coverage
Binary crop/non-crop classification

SustainBench Crop Yield

Crop yield prediction.

from torchgeo.datasets import SustainBenchCropYield

dataset = SustainBenchCropYield(
    root='data/sustainbench',
    split='train'
)

Properties:

Satellite imagery + weather data
Yield regression task
US coverage

Embedding Datasets

Pre-computed foundation model embeddings.

from torchgeo.datasets import (
    ClayEmbeddings,           # Clay foundation model
    PrestoEmbeddings,         # Presto foundation model  
    SatlasPretrain,           # Satlas pretraining
    SSLEOLBenchmark,          # SSL4EO benchmark
    TesseraEmbeddings,        # Tessera embeddings
    MajorTOMEmbeddings,       # MajorTOM embeddings
    EarthEmbeddings,          # Earth observation embeddings
    GoogleSatelliteEmbedding, # Google embeddings
)

dataset = PrestoEmbeddings(
    root='data/presto',
    download=True
)

These datasets provide pre-computed embeddings from foundation models, useful for transfer learning and fine-tuning.

Usage Examples

Basic Training Loop

from torchgeo.datasets import EuroSAT
from torch.utils.data import DataLoader
import torch.nn as nn
import torch.optim as optim

# Load dataset
train_dataset = EuroSAT(root='data', split='train', download=True)
val_dataset = EuroSAT(root='data', split='val')

# Create dataloaders
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=32)

# Training
model = MyModel()
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters())

for epoch in range(num_epochs):
    for sample in train_loader:
        images = sample['image']
        labels = sample['label']
        
        outputs = model(images)
        loss = criterion(outputs, labels)
        
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

With Transforms

import kornia.augmentation as K
from torchgeo.transforms import AugmentationSequential

transforms = AugmentationSequential(
    K.RandomHorizontalFlip(p=0.5),
    K.RandomVerticalFlip(p=0.5),
    K.RandomRotation(degrees=90),
    K.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    data_keys=['image']
)

dataset = EuroSAT(
    root='data',
    split='train',
    transforms=transforms
)

Segmentation Example

from torchgeo.datasets import DeepGlobeLandCover

dataset = DeepGlobeLandCover(
    root='data/deepglobe',
    split='train',
    transforms=transforms
)

for sample in dataset:
    image = sample['image']  # [C, H, W]
    mask = sample['mask']    # [H, W] with class indices
    
    # Your segmentation model
    pred = model(image.unsqueeze(0))
    loss = criterion(pred, mask.unsqueeze(0))

Datasets

Models

Components

Non-Geospatial Datasets

Base Classes

NonGeoDataset

NonGeoClassificationDataset

Common Parameters

Classification Datasets

EuroSAT

RESISC45

UC Merced

Additional Classification Datasets

Segmentation Datasets

Inria Aerial Image Labeling

DeepGlobe Land Cover

LoveDA

Potsdam2D

Vaihingen2D

Additional Segmentation Datasets

Object Detection Datasets

VHR-10

DIOR

DOTA

Additional Detection Datasets

Multi-Task Datasets

BigEarthNet

SEN12MS

BioMassters

Temporal/Time Series Datasets

PASTIS

CropHarvest

SustainBench Crop Yield

Embedding Datasets

Usage Examples

Basic Training Loop

With Transforms

Segmentation Example

See Also

Build docs developers (and LLMs) love

Datasets

Models

Components

Documentation Index

​Base Classes

​NonGeoDataset

​NonGeoClassificationDataset

​Common Parameters

​Classification Datasets

​EuroSAT

​RESISC45

​UC Merced

​Additional Classification Datasets

​Segmentation Datasets

​Inria Aerial Image Labeling

​DeepGlobe Land Cover

​LoveDA

​Potsdam2D

​Vaihingen2D

​Additional Segmentation Datasets

​Object Detection Datasets

​VHR-10

​DIOR

​DOTA

​Additional Detection Datasets

​Multi-Task Datasets

​BigEarthNet

​SEN12MS

​BioMassters

​Temporal/Time Series Datasets

​PASTIS

​CropHarvest

​SustainBench Crop Yield

​Embedding Datasets

​Usage Examples

​Basic Training Loop

​With Transforms

​Segmentation Example

​See Also

Build docs developers (and LLMs) love

Base Classes

NonGeoDataset

NonGeoClassificationDataset

Common Parameters

Classification Datasets

EuroSAT

RESISC45

UC Merced

Additional Classification Datasets

Segmentation Datasets

Inria Aerial Image Labeling

DeepGlobe Land Cover

LoveDA

Potsdam2D

Vaihingen2D

Additional Segmentation Datasets

Object Detection Datasets

VHR-10

DIOR

DOTA

Additional Detection Datasets

Multi-Task Datasets

BigEarthNet

SEN12MS

BioMassters

Temporal/Time Series Datasets

PASTIS

CropHarvest

SustainBench Crop Yield

Embedding Datasets

Usage Examples

Basic Training Loop

With Transforms

Segmentation Example

See Also