What specific augmentations for satellite images (rotation, multi-scale)?

For augmenting satellite images, specific techniques that consider the features of geospatial data are used:

• Rotations: often 90°, 180°, 270° rotations are used, as satellite images can have different orientations. It is important to preserve georeferencing during rotations.
• Multi-scale: changing the scale of the image to train models that are robust to different resolutions. This may include resampling while maintaining proportions.
• Flips: horizontal and vertical flips, if they do not violate geographic correctness.
• Brightness and contrast adjustments: to simulate different lighting and atmospheric conditions.
• Adding noise and artifacts: to mimic interference, cloud cover, or other distortions.
• Cropping and shifting: for training models on parts of images and increasing robustness to displacements.

It is important to note that some augmentations can distort geospatial information, so their application should be deliberate and aligned with the task.