Available Models

While the list of available models is currently small, we intend to add more over the coming months, especially with feedback from our users. Training a single model is quite expensive, and as an open source project, our budget is small, so we want to make sure that we spend our money wisely and create the most useful possible models.

Currently, we support both pre-built models, as well as weights pretrained on specific datasets. We will release more combinations over time.

UNet

ResNet50 • FMOW RGB • View Code

Attribute	Value
Full Name	unet50_fmow_rgb
Architecture	UNet
Backbone	ResNet-50
Data Source	QuickBird-2, GeoEye-1, WorldView-2, WorldView3
Data Format	RGB
Pretraining	Masked Autoencoding

A UNet that has been pretrained on the functional map of the world RGB dataset. The model was trained using masked autoencoding self-supervised learning, meaning that it should be more task agnostic than a model pretrained on a specific target task.

To pre-process data, use fmow_rgb mode for the dataset and unet for the model. This mode expects uint8 input values from 0 - 255 in RGB ordering.

ResNet50 • FMOW Multispectral • View Code

Attribute	Value
Full Name	unet50_fmow_full
Architecture	UNet
Backbone	ResNet-50
Data Source	QuickBird-2, GeoEye-1, WorldView-2, WorldView3
Data Format	4/8 channel multispectral
Pretraining	Masked Autoencoding

A UNet that has been pretrained on the functional map of the world RGB dataset. The model was trained using masked autoencoding self-supervised learning, meaning that it should be more task agnostic than a model pretrained on a specific target task. Compared to the resnet50_fmow_rgb dataset, this dataset uses the multispectral inputs in either 4 or 8 channel format. The model has been trained to deal with the latter 4 missing channels for images that do not have 8 channels of data.

To pre-process data, use fmow_full mode for the dataset and unet for the model. This mode expects uint16 input values from 0 - 65535, as the default in the functional map of the world dataset.

ResNet50 • Sentinel-2 L2A • View Code

Attribute	Value
Full Name	unet50_sentinel2_l2a
Architecture	UNet
Backbone	ResNet-50
Data Source	Sentinel 2 (L2A)
Data Format	12 channel multispectral
Pretraining	Masked Autoencoding

A UNet that has been pretrained on the BigEarthNet dataset. The model was trained using masked autoencoding self-supervised learning, meaning that it should be more task agnostic than a model pretrained on a specific target task.

To pre-process data, use sentinel-l2a mode for the dataset and unet for the model. This mode expects uint16 input values from 0 - 65535, as the default for Sentinel-2 data.