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 implements a variety of model architectures optimized for remote sensing tasks. Models are organized by their primary use case: classification backbones, segmentation, change detection, and foundation models.

​
Classification Backbones

Pre-trained encoders suitable for transfer learning and feature extraction.

​
ResNet

Residual Networks for image classification and feature extraction. Available Variants:
  • resnet18: 18-layer ResNet (11.7M parameters)
  • resnet50: 50-layer ResNet (25.6M parameters)
  • resnet152: 152-layer ResNet (60.2M parameters)
Key Features:
  • Residual connections for training deep networks
  • Multiple pre-trained weights for different sensors
  • Support for arbitrary input channels via in_chans parameter
  • Based on timm implementation
Usage: 
from torchgeo.models import resnet50, ResNet50_Weights

# Load with pre-trained weights
weights = ResNet50_Weights.SENTINEL2_ALL_MOCO
model = resnet50(weights=weights)

# Custom number of channels
model = resnet50(in_chans=4, num_classes=10)
Reference: Deep Residual Learning for Image Recognition

​
Vision Transformer (ViT)

Transformer-based architecture for image classification. Available Variants:
  • vit_small_patch16_224: Small ViT with 16x16 patches (22M parameters)
  • vit_base_patch16_224: Base ViT with 16x16 patches (86M parameters)
  • vit_large_patch16_224: Large ViT with 16x16 patches (304M parameters)
  • vit_huge_patch14_224: Huge ViT with 14x14 patches (632M parameters)
  • vit_small_patch14_dinov2: Small DINOv2 ViT with 14x14 patches
  • vit_base_patch14_dinov2: Base DINOv2 ViT with 14x14 patches
Key Features:
  • Pure transformer architecture without convolutions
  • Self-attention mechanisms for global context
  • Extensive pre-trained weights from SSL4EO-S12 and SSL4EO-L
  • Support for MAE, DINO, MoCo, and other SSL methods
Usage: 
from torchgeo.models import vit_small_patch16_224, ViTSmall16_Weights

# Load with Sentinel-2 MAE weights
weights = ViTSmall16_Weights.SENTINEL2_ALL_MAE
model = vit_small_patch16_224(weights=weights)

# Feature extraction mode
model = vit_small_patch16_224(weights=weights, features_only=True)
Reference: An Image is Worth 16x16 Words

​
Swin Transformer

Hierarchical vision transformer with shifted windows. Available Variants:
  • swin_t: Tiny Swin Transformer
  • swin_s: Small Swin Transformer
  • swin_b: Base Swin Transformer
  • swin_v2_t: Swin Transformer V2 Tiny
  • swin_v2_b: Swin Transformer V2 Base
Key Features:
  • Hierarchical feature maps at multiple scales
  • Shifted window attention for efficiency
  • Pre-trained on SatlasPretrain dataset
  • Support for both RGB and multispectral inputs
Usage: 
from torchgeo.models import swin_v2_t, Swin_V2_T_Weights

# Load with Satlas weights
weights = Swin_V2_T_Weights.SENTINEL2_MI_MS_SATLAS
model = swin_v2_t(weights=weights)
Reference: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

​
Foundation Models

Large-scale models trained on diverse geospatial data with specialized capabilities.

​
DOFA (Dynamic One-For-All)

A dynamic model that adapts to any number of spectral bands via wavelength-conditioned convolutions. Available Variants:
  • dofa_small_patch16_224: Small DOFA (22M parameters)
  • dofa_base_patch16_224: Base DOFA (86M parameters)
  • dofa_large_patch16_224: Large DOFA (304M parameters)
  • dofa_huge_patch14_224: Huge DOFA (632M parameters)
Key Features:
  • Dynamic channel adaptation: Works with any spectral bands by conditioning on wavelengths
  • Trained on SatlasPretrain, Five-Billion-Pixels, and HySpecNet-11k
  • Transformer architecture with dynamic weight generator
  • Pre-trained with MAE (Masked Autoencoding)
Usage: 
from torchgeo.models import dofa_base_patch16_224, DOFABase16_Weights

# Load pre-trained weights
weights = DOFABase16_Weights.DOFA_MAE
model = dofa_base_patch16_224(weights=weights)

# Forward pass with wavelengths (in micrometers)
wavelengths = [0.443, 0.490, 0.560, 0.665, 0.705]  # 5 bands
output = model(x, wavelengths=wavelengths)
Reference: Dynamic One-For-All (DOFA)

​
Presto

Pretrained Remote Sensing Transformer for Sentinel-1/2 time series. Key Features:
  • Temporal transformer for satellite image time series
  • Encoder-decoder architecture with masked token prediction
  • Multi-modal: Sentinel-1 SAR + Sentinel-2 optical
  • Includes auxiliary inputs: Dynamic World labels, lat/lon, month
  • Pre-trained on LEM (Presto pretraining dataset)
Band Groups:
  • S1: Sentinel-1 (VV, VH)
  • S2_RGB, S2_Red_Edge, S2_NIR, S2_SWIR: Sentinel-2 bands
  • ERA5: Climate reanalysis data
  • SRTM: Elevation data
  • NDVI: Vegetation index
Usage: 
from torchgeo.models import presto, Presto_Weights
import torch

# Load pre-trained model
weights = Presto_Weights.PRESTO
model = presto(weights=weights)

# Input format: [batch, timesteps, channels]
x = torch.randn(2, 12, 17)  # 2 samples, 12 timesteps, 17 channels
dynamic_world = torch.randint(0, 9, (2, 12))  # Land cover labels
latlons = torch.randn(2, 2)  # Latitude/longitude

# Forward pass
reconstructed, dw_output = model(x, dynamic_world, latlons)
Reference: Presto: A Foundation Model for Remote Sensing

​
CopernicusFM

Copernicus Foundation Model for multi-temporal satellite imagery. Available Variants:
  • copernicusfm_base: Base CopernicusFM model
Key Features:
  • Multi-temporal Sentinel-2 processing
  • Foundation model trained on Copernicus data
  • Supports various downstream tasks
Usage: 
from torchgeo.models import copernicusfm_base, CopernicusFM_Base_Weights

model = copernicusfm_base()

​
ScaleMAE

Scale-aware Masked Autoencoder for multi-resolution imagery. Available Variants:
  • scalemae_large_patch16: Large ScaleMAE with patch size 16
Key Features:
  • Multi-scale masked autoencoding
  • Handles images at different spatial resolutions
  • Vision transformer backbone
Usage: 
from torchgeo.models import scalemae_large_patch16, ScaleMAELarge16_Weights

model = scalemae_large_patch16()

​
Other Foundation Models

CROMA (Contrastive Multi-modal Alignment):
  • croma_base: Base CROMA model
  • croma_large: Large CROMA model
  • Multi-modal contrastive learning
Aurora:
  • aurora_swin_unet: Swin-UNet for weather forecasting
Panopticon:
  • panopticon_vitb14: Vision transformer for global monitoring
EarthLoc:
  • earthloc: Model for geographic location prediction
Tessera:
  • tessera: Tessellated earth observation model
TileNet:
  • tilenet: Tile-based representation learning

​
Segmentation Models

Dense prediction models for pixel-wise classification tasks.

​
U-Net

U-shaped encoder-decoder architecture for semantic segmentation. Key Features:
  • Encoder-decoder with skip connections
  • Multiple encoder backbones (EfficientNet, ResNet, etc.)
  • Pre-trained weights for field boundary detection
  • Based on segmentation_models_pytorch (smp)
Usage: 
from torchgeo.models import unet, Unet_Weights

# Load with pre-trained weights
weights = Unet_Weights.SENTINEL2_2CLASS_FTW
model = unet(weights=weights)

# Custom architecture
import segmentation_models_pytorch as smp
model = smp.Unet(
    encoder_name='resnet50',
    encoder_weights=None,
    in_channels=13,
    classes=10
)
Available Pre-trained Weights:
  • Field boundary detection (2-class and 3-class)
  • Various EfficientNet encoders (B3, B5, B7)
  • Commercial and non-commercial licenses
Reference: U-Net: Convolutional Networks for Biomedical Image Segmentation

​
FCN (Fully Convolutional Network)

Simple fully convolutional architecture for semantic segmentation. Key Features:
  • 5-layer fully convolutional architecture
  • LeakyReLU activations
  • Lightweight and fast
  • Customizable number of filters
Usage: 
from torchgeo.models import FCN

# Create FCN model
model = FCN(
    in_channels=13,      # Sentinel-2 bands
    classes=10,          # Number of output classes
    num_filters=64       # Filters per conv layer
)

​
FarSeg

Foreground-Aware Relation Network for object segmentation. Key Features:
  • ResNet backbone with FPN (Feature Pyramid Network)
  • Foreground-scene relation module
  • Designed for building, road, ship segmentation
  • Can be extended for change detection
Usage: 
from torchgeo.models import FarSeg
from torchgeo.models import ResNet50_Weights

# Create FarSeg with pre-trained backbone
backbone_weights = ResNet50_Weights.SENTINEL2_ALL_MOCO
model = FarSeg(
    backbone='resnet50',
    classes=2,  # Binary segmentation
    backbone_weights=backbone_weights
)
Reference: Foreground-Aware Relation Network for Geospatial Object Segmentation

​
Change Detection Models

Models specialized for detecting changes between bi-temporal images.

​
ChangeStar

Change detection model combining segmentation and change prediction. Key Features:
  • Combines semantic segmentation with change detection
  • ChangeMixin module for binary change prediction
  • Bi-directional change detection (t1→t2 and t2→t1)
  • Architecture reusability: works with any segmentation backbone
Usage: 
from torchgeo.models import ChangeStar
from torchgeo.models import FarSeg

# Create base segmentation model
seg_model = FarSeg(backbone='resnet50', classes=10)

# Wrap with ChangeStar for change detection
model = ChangeStar(
    seg_model=seg_model,
    in_channels=128,  # Feature channels from seg_model
    inner_channels=16,
    num_convs=4
)
Reference: Change is Everywhere: Single-Temporal Supervised Object Change Detection

​
ChangeStarFarSeg

Pre-configured ChangeStar with FarSeg backbone. Usage: 
from torchgeo.models import ChangeStarFarSeg

model = ChangeStarFarSeg(
    backbone='resnet50',
    classes=10
)

​
FCSiamDiff / FCSiamConc

Siamese fully convolutional networks for change detection. Variants:
  • FCSiamDiff: Difference-based fusion
  • FCSiamConc: Concatenation-based fusion
Key Features:
  • Siamese architecture with shared weights
  • Process bi-temporal images
  • Lightweight and efficient
Usage: 
from torchgeo.models import FCSiamDiff, FCSiamConc

# Difference-based
model_diff = FCSiamDiff(in_channels=13, classes=2)

# Concatenation-based  
model_concat = FCSiamConc(in_channels=13, classes=2)

​
ChangeViT

Vision transformer-based change detection model. Usage: 
from torchgeo.models import ChangeViT

model = ChangeViT()

​
Time Series Models

Models for temporal sequence processing.

​
ConvLSTM

Convolutional LSTM for spatiotemporal sequence modeling. Key Features:
  • Combines CNN and LSTM for spatiotemporal patterns
  • Processes sequences of images
  • Maintains spatial structure through convolutions
Usage: 
from torchgeo.models import ConvLSTM

model = ConvLSTM(
    input_dim=13,
    hidden_dim=64,
    kernel_size=(3, 3),
    num_layers=2
)

​
LTAE

Lightweight Temporal Attention Encoder. Key Features:
  • Attention-based temporal encoding
  • Lightweight and efficient
  • Designed for satellite time series
Usage: 
from torchgeo.models import LTAE

model = LTAE(
    in_channels=13,
    n_head=4,
    d_k=16,
    d_model=64
)

​
Other Models

​
RCF / MOSAIKS

Random Convolutional Features for large-scale geospatial analysis. Usage: 
from torchgeo.models import RCF, MOSAIKS

# Random convolutional features
rcf = RCF(num_features=512)

# MOSAIKS variant
mosaiks = MOSAIKS(num_features=512)
Reference: MOSAIKS: A measure of satellite imagery usability

​
BTC

Behavioral cloning model for trajectory prediction. Usage: 
from torchgeo.models import BTC

model = BTC()

​
Model Selection Guide

​
For Image Classification

Small datasets (under 10k images):
  • ResNet18/50 with pre-trained weights
  • ViT Small with MAE pre-training
Large datasets (over 100k images):
  • ResNet152 or ViT Large/Huge
  • DOFA for multi-sensor/multi-spectral
Time series classification:
  • Presto for Sentinel-1/2 time series
  • ConvLSTM or LTAE

​
For Semantic Segmentation

General segmentation:
  • U-Net with EfficientNet encoder
  • FarSeg for object-aware segmentation
Field boundaries:
  • U-Net with FTW pre-trained weights
Large-scale mapping:
  • Swin Transformer with Satlas weights

​
For Change Detection

Binary change:
  • ChangeStar with FarSeg
  • FCSiamDiff/FCSiamConc
Semantic change:
  • ChangeStar for multi-class change
  • ChangeViT for transformer-based

​
For Multi-Sensor Fusion

Any spectral bands:
  • DOFA (wavelength-conditioned)
Sentinel-1 + Sentinel-2:
  • Presto (temporal)
  • Separate encoders with late fusion

​
For Transfer Learning

Best pre-trained weights:
  • ResNet50: SENTINEL2_ALL_MOCO or SENTINEL2_ALL_DINO
  • ViT: SENTINEL2_ALL_MAE or SENTINEL2_ALL_FGMAE
  • DOFA: DOFA_MAE for any bands
Domain-specific:
  • FTW weights for agricultural fields
  • Satlas weights for global mapping
  • SSL4EO-L for Landsat applications

Build docs developers (and LLMs) love

Get started for freeTalk to us