Tessera#

class torchgeo.models.Tessera(embed_dim=128)[source]#

Bases: Module

Tessera pixel time series foundation model.

Tessera is a foundation model for pixel-level time series data from Sentinel-1 and Sentinel-2 satellites. It uses separate transformer encoders for SAR and optical data with temporal-aware pooling.

If you use this model in your research, please cite the following paper:

Added in version 0.9.

__init__(embed_dim=128)[source]#

Initialize a new Tessera instance.

Parameters:

embed_dim (int) – Output embedding dimension.

forward(x)[source]#

Forward pass of the Tessera model.

Parameters:

x (Tensor) – Input tensor of shape (B, seq_len, 14) containing: - Channels 0-9: Sentinel-2 bands (B2-B12) - Channel 10: Sentinel-2 day of year - Channels 11-12: Sentinel-1 VV and VH - Channel 13: Sentinel-1 day of year

Returns:

Fused embedding tensor of shape (B, embed_dim).

Raises:

AssertionError – If input does not have 14 channels.

Return type:

Tensor

torchgeo.models.tessera(weights=None, *args, **kwargs)[source]#

Tessera pixel time series foundation model.

If you use this model in your research, please cite the following paper:

Added in version 0.9.

Parameters:

  • weights (Tessera_Weights | None) – Pre-trained model weights to use. If using encoder-only weights (TESSERA_SENTINEL1_ENCODER or TESSERA_SENTINEL2_ENCODER), returns the respective encoder backbone instead of the full model.

  • *args (Any) – Additional arguments to pass to Tessera.

  • **kwargs (Any) – Additional keyword arguments to pass to Tessera.

Returns:

A Tessera model or encoder backbone.

Return type:

Module

class torchgeo.models.Tessera_Weights(*values)[source]#

Bases: WeightsEnum

Tessera model weights.

Added in version 0.9.

__new__(value)#