Models Overview

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.

​

Key Features

​

Pre-trained Weights for Multispectral Data

• Sentinel-1 (SAR): 2 channels (VV, VH)
• Sentinel-2 (Optical): Up to 13 spectral bands
• Landsat: 6-11 channels depending on sensor (TM, ETM+, OLI/TIRS)
• Multi-sensor: Combined modalities for robust feature learning

​

Self-Supervised Learning Methods

• MoCo: Momentum Contrast for unsupervised learning
• SimCLR: Simple framework for contrastive learning
• MAE: Masked Autoencoders for self-supervised learning
• DINO: Self-distillation with no labels
• SeCo: Seasonal contrast for temporal understanding
• DeCUR: Decoupled contrastive learning
• SoftCon: Soft contrastive learning

​

Integration with timm

• Consistent API across all models
• Easy weight loading and model creation
• Support for feature extraction and fine-tuning
• Flexible input channel configuration

​

Model Categories

​

Classification Backbones

• ResNet (18, 50, 152)
• Vision Transformer (Small, Base, Large, Huge)
• Swin Transformer (Tiny, Small, Base)
• DOFA (Dynamic One-For-All)
• Presto (Pretrained Remote Sensing Transformer)

​

Segmentation Models

• U-Net with various encoders
• FCN (Fully Convolutional Network)
• FarSeg (Foreground-Aware Relation Network)

​

Change Detection Models

• ChangeStar
• ChangeStarFarSeg
• FCSiamDiff, FCSiamConc

​

Foundation Models

• DOFA: Works with any spectral bands via dynamic convolutions
• Presto: Temporal transformer for Sentinel-1/2 time series
• CopernicusFM: Copernicus Foundation Model
• ScaleMAE: Scale-aware masked autoencoder

​

Quick Start

​

Loading Pre-trained Weights

from torchgeo.models import resnet50, ResNet50_Weights

# Load model with Sentinel-2 weights (13 channels)
weights = ResNet50_Weights.SENTINEL2_ALL_MOCO
model = resnet50(weights=weights)

# Model is ready for fine-tuning
for param in model.parameters():
    param.requires_grad = True

​

Using the Weights API

from torchgeo.models import get_weight, get_model_weights

# Get a specific weight by string name
weight = get_weight('ResNet50_Weights.SENTINEL2_ALL_MOCO')

# List all available weights for a model
weights = get_model_weights('resnet50')
for w in weights:
    print(f"{w}: {w.meta['dataset']} - {w.meta.get('ssl_method', 'N/A')}")

​

Creating Models with Custom Channels

from torchgeo.models import vit_small_patch16_224

# Create model without weights (random initialization)
model = vit_small_patch16_224(in_chans=4)  # 4 channels for RGB + NIR

# Load weights and modify first layer for different channels
weights = ViTSmall16_Weights.SENTINEL2_ALL_MOCO
model = vit_small_patch16_224(weights=weights)  # Loads with 13 channels

​

Training Datasets

• SSL4EO-S12: 250k+ Sentinel-1/2 image pairs
• SSL4EO-L: 500k+ Landsat images (TM, ETM+, OLI)
• SatlasPretrain: Global coverage Sentinel-2 imagery
• fMoW: Functional Map of the World satellite imagery
• Five-Billion-Pixels: Large-scale hyperspectral dataset