torchvision is great for natural images. But remote sensing data is different:

• GeoTIFFs, not PNGs — with coordinate reference systems baked in
• Multi-spectral bands — beyond RGB into near-infrared, thermal, SAR
• Massive sizes — a single satellite image can be 10,000×10,000 pixels
• Spatial context matters — random cropping destroys geographic patterns

TorchGeo is PyTorch's official geospatial extension by Microsoft. It provides 50+ remote sensing datasets (one-line download), geo-aware samplers, and seamless integration with torchvision and PyTorch Lightning.

pip install torchgeo rasterio

Loading Your First Dataset

Let's start with EuroSAT — 27,000 Sentinel-2 satellite images across 10 land cover classes:

from torchgeo.datasets import EuroSAT

dataset = EuroSAT(root="./data", download=True)
print(len(dataset))          # 27000
print(dataset.num_classes)   # 10
print(dataset.classes)
# ['AnnualCrop', 'Forest', 'HerbaceousVegetation',
#  'Highway', 'Industrial', 'Pasture', 'PermanentCrop',
#  'Residential', 'River', 'SeaLake']

Each sample is a dict with image (multi-spectral tensor) and label (integer):

sample = dataset[0]
print(sample['image'].shape)  # torch.Size([13, 64, 64]) — 13 Sentinel-2 bands
print(sample['label'])        # 0 → AnnualCrop

Building the Data Pipeline

Remote sensing datasets return dicts, so we need a custom collate_fn:

from torch.utils.data import DataLoader
from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize((64, 64)),
    transforms.Lambda(lambda x: x.float() / 255.0),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

def collate_fn(batch):
    images = torch.stack([transform(b['image'][:3]) for b in batch])
    labels = torch.tensor([b['label'] for b in batch])
    return images, labels

loader = DataLoader(dataset, batch_size=32, collate_fn=collate_fn)

For RGB display, we take image[:3] (B04, B03, B02). For multi-spectral analysis, keep all 13 bands.

Transfer Learning with ResNet18

Replace the final fully-connected layer for our 10 classes:

from torchvision.models import resnet18
import torch.nn as nn

model = resnet18(weights='IMAGENET1K_V1')
model.fc = nn.Linear(512, 10)

With 3 epochs and ImageNet pretrained weights on GPU (RTX 4060, 8GB), this reaches 97.8% training accuracy and 83.7% test accuracy in just 40 seconds:

criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

for epoch in range(3):
    for images, labels in train_loader:
        images, labels = images.to(device), labels.to(device)
        optimizer.zero_grad()
        outputs = model(images)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

Key Datasets at a Glance

Where to Go Next

1. Geo-aware sampling — RandomGeoSampler for tiling massive GeoTIFFs
2. Pre-trained remote sensing models — TorchGeo ships weights pretrained on BigEarthNet
3. Semantic segmentation — LandCoverAI + DeepLabV3 for pixel-level classification
4. Multi-spectral processing — Work with all 13 Sentinel-2 bands
5. Change detection — Compare satellite images across time

Official docs: docs.torchgeo.org

References

• TorchGeo GitHub
• EuroSAT Paper