Dataset Masks

Masks remove unwanted source pixels before dataset samples are generated. They can exclude clouds, saturation, sensor errors, water, shadows, or any other pixels that should not contribute to training or evaluation.

Dataset versions apply source image masks by default. Configure each mask's use and effect on the source image in Spectral Explorer, then leave Masks enabled on the dataset. Disable it only when that dataset version should ignore source image masks during finalization.

Masks and labels have different roles: labels identify the samples and classes to include, while masks remove pixels from those samples.