PyTorch domain library for remote sensing
TorchGeo gives researchers and engineers CRS-aware datasets, spatial samplers, multispectral transforms, and pretrained models for satellite and aerial imagery.
pip install torchgeoMIT license.from torchgeo.datasets import EuroSAT
from torchgeo.samplers import RandomGeoSampler
from torchgeo.models import ResNet18_Weights, resnet18
dataset = EuroSAT(root="./data", download=True)
sampler = RandomGeoSampler(dataset, size=256, length=1000)
weights = ResNet18_Weights.SENTINEL2_ALL_MOCO
transforms = weights.transforms()
model = resnet18(weights=weights)
Library surface
TorchGeo stays close to standard PyTorch ergonomics while handling the spatial rules that ordinary vision libraries ignore.
Datasets
Raster and vector benchmarks with coordinates, metadata, and domain-specific conventions already handled.
Samplers
Patch extraction and region-aware iteration that respect extent, scale, and geographic boundaries.
Transforms
Kornia-compatible transforms for arbitrary band counts, not just RGB assumptions carried over from natural images.
Models
Satellite-pretrained backbones and task-ready components that fit directly into PyTorch training loops.
Design principles
The site should feel closer to a strong scientific software project than a startup landing page. Clear claims. Clear entry points. Enough polish to feel current.
Clear APIs, reproducible dataset wrappers, and a narrow learning curve for teams already fluent in PyTorch.
Coordinate reference systems, tiling, geospatial metadata, and multi-band imagery treated as first-class concerns.
Documentation, examples, and library surface arranged for fast orientation instead of startup-style visual noise.
Workflow
The path is simple, but the underlying geospatial details are not. TorchGeo carries them for you.
Open satellite or aerial benchmarks with spatial metadata preserved.
Generate geographic tiles and train-validation splits that match the study area.
Apply augmentation and preprocessing across multispectral tensors, not just three channels.
Fine-tune pretrained models or build new ones inside familiar PyTorch workflows.
Examples
Existing benchmarks and example workflows make the project legible for both first-time users and experienced remote sensing groups.
Start here
Documentation, tutorials, paper, and source code all in one place.
pip install torchgeoOpen documentation