Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/torchgeo/torchgeo/llms.txt

Use this file to discover all available pages before exploring further.

TorchGeo provides a comprehensive collection of datasets for geospatial machine learning. These datasets are divided into two main categories based on whether they contain geospatial metadata.

​
Dataset Categories

Geospatial Datasets

Datasets with coordinate information that can be spatially indexed and combined

Non-Geospatial Datasets

Benchmark datasets with pre-defined image chips for various computer vision tasks

​
Key Differences

​
Geospatial Datasets

Geospatial datasets (GeoDataset) contain rich geospatial metadata including:
  • Coordinates (latitude, longitude)
  • Coordinate Reference System (CRS)
  • Resolution
  • Temporal information
This metadata enables powerful spatial operations: 
from torchgeo.datasets import Landsat8, CDL

# Load geospatial datasets
landsat = Landsat8(paths='data/landsat')
cdl = CDL(paths='data/cdl')

# Combine using spatial intersection
dataset = landsat & cdl

# Query by bounding box and time
sample = dataset[xmin:xmax:xres, ymin:ymax:yres, tmin:tmax]

​
Non-Geospatial Datasets

Non-geospatial datasets (NonGeoDataset) are pre-chipped benchmark datasets without coordinate information: 
from torchgeo.datasets import EuroSAT
from torch.utils.data import DataLoader

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

# Use with standard PyTorch DataLoader
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

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

​
Common Patterns

​
Combining Geospatial Datasets

TorchGeo provides two operators for combining geospatial datasets: Intersection (&): Samples must exist in both datasets 
# Combine image and labels from same location
dataset = imagery & labels
Union (|): Samples can exist in either dataset 
# Combine data from different sensors or locations
dataset = landsat7 | landsat8

​
Sampling Strategies

For geospatial datasets, use samplers to generate random queries: 
from torchgeo.samplers import RandomGeoSampler
from torch.utils.data import DataLoader

sampler = RandomGeoSampler(dataset, size=256, length=1000)
dataloader = DataLoader(dataset, sampler=sampler, batch_size=4)

​
Transforms

All datasets support transforms for data augmentation: 
import kornia.augmentation as K
from torchgeo.transforms import AugmentationSequential

transforms = AugmentationSequential(
    K.RandomHorizontalFlip(p=0.5),
    K.RandomVerticalFlip(p=0.5),
    data_keys=['image', 'mask']
)

dataset = MyDataset(paths='data', transforms=transforms)

​
Common Parameters

Most datasets share these common parameters:
​
root
str | PathLike
default:"'data'"
Root directory where dataset is stored (for NonGeoDatasets)
​
paths
str | PathLike | Iterable[str | PathLike]
default:"'data'"
One or more root directories to search or files to load (for GeoDatasets)
​
transforms
Callable[[Sample], Sample] | None
default:"None"
Function to transform samples after loading
​
download
bool
default:"False"
If True, download dataset if not found (for benchmark datasets)
​
checksum
bool
default:"False"
If True, verify file integrity using MD5 checksums

​
Sample Format

All datasets return samples as dictionaries with standardized keys: Image datasets: 
{
    'image': Tensor,      # [C, H, W] or [T, C, H, W]
    'bounds': Tensor,     # Spatiotemporal bounds (GeoDatasets only)
    'transform': Tensor   # Affine transform matrix (GeoDatasets only)
}
Segmentation datasets: 
{
    'image': Tensor,  # [C, H, W]
    'mask': Tensor    # [H, W] or [C, H, W]
}
Classification datasets: 
{
    'image': Tensor,  # [C, H, W]
    'label': Tensor   # Scalar or [1]
}
Object detection datasets: 
{
    'image': Tensor,      # [C, H, W]
    'bbox_xyxy': Tensor,  # [N, 4]
    'label': Tensor       # [N]
}

​
Next Steps

Geospatial Datasets

Explore RasterDataset, VectorDataset, and other geospatial base classes

Non-Geospatial Datasets

Browse benchmark datasets for classification, segmentation, and detection

Build docs developers (and LLMs) love

Get started for freeTalk to us