Preprocessing

In order to maximize the benefits of the pretrained models, you must use the same input processing as the training process. The preprocessing functions vary between both expected type of data inputs, as well as model architectures.

Model

Model	Description
unet	Formatting for the UNet series of architectures. These models expect multichannel inputs on a fixed basis.

Dataset

Dataset	Description	Platforms
fmow_rgb	Applies mean/variance standardization to the RGB images from the functional map of the world dataset. Expects a channel format of RGB. Note that while statistics were computed on a specific dataset, we expect this formatting to work generally with RGB satellite images with roughly the distribution of the dataset (i.e. not too much water or clouds).	QuickBird-2, GeoEye-1, WorldView-2, WorldView3
fmow_full	Applies mean/variance standardization to the multispectral 16-bit images from the functional map of the world dataset. Because some of the platforms used are only 4 channels, we train to be robust to zeros in the missing channels.	QuickBird-2, GeoEye-1, WorldView-2, WorldView3

| sentinel2_l2a | Applies mean/variance standardization to the multispectral 16-bit images from the Sentinel-2 L2A product. This product is missing the 10 band from the L1C product because the cirrus layer cannot be top of atmosphere corrected. | Sentinel 2 |