Explicit Handling of the nodata-Metadata Value

[ChristopherKotthoff]

Problem

When using .tif files for training, particularly with multispectral imagery, nodata values in the source metadata can be misinterpreted as valid pixel values. This is critical because we retain the .tif format rather than converting to formats like PNG that inherently bound values. Mismanagement of nodata values could degrade model performance and skew training results.

Why It Matters

• Data Integrity: nodata values could lead to incorrect model learning.
• Multispectral Complexity: Different bands may have varying no-data values.
• Model Performance: Clean handling of nodata ensures better generalization.

Proposed Solution

• Introduce explicit handling of no-data values based on .tif metadata.
• Mask nodata pixels to exclude them from training.

This will improve the reliability and quality of input data for training, especially for multispectral workflows.