[Python] How to Normalize 16-bit GeoTIFF Images and Stack Them for RGB Creation Without Losing...

I am working with a set of remote sensing images in GeoTIFF format for a machine learning project. These images include:

• Red band, Green band, and Blue band (optical bands) in the 16-bit resolution range.
• VH band from Synthetic Aperture Radar (SAR) data.

The pixel values in these images range from 0 to 65535, representing the intensity of each pixel. I need to preprocess these images to:

• Normalize the pixel values without losing any data.
• Stack the individual bands (Red, Green, Blue, and VH) to create a composite RGB image that is suitable for training a model.

Key Points: Normalization Goal: I need to normalize the images so that the pixel values fit within a specific range, but I do not want to lose any of the original data during the process. I’m aware that typical normalization scales pixel values to a range like [0, 1] or [0, 255], but I’m concerned about losing valuable information due to the 16-bit depth.

Stacking Bands for RGB: Once normalized, I need to stack the Red, Green, Blue, and VH bands to create an RGB composite image. However, since SAR images (VH band) are not inherently in RGB format, I’m unsure about how to combine them effectively.

I've tried standard min-max normalization for scaling the pixel values, but I fear that this method could lead to loss of resolution in the higher or lower range of the pixel values.

I’ve also considered simply clipping the values to a range like [0, 255] to fit standard RGB format, but this could cause information loss for high dynamic range pixels (especially in SAR data).

my objectives: Best Approach to Normalize 16-bit Images: How can I properly normalize 16-bit images so that I retain all valuable information, especially from the SAR band (VH)? Should I normalize them to a range like [0, 1] or [0, 255], or use a different approach?

Proper Method for Stacking Bands for RGB: How do I effectively combine the Red, Green, Blue, and VH bands into a single image for training a model without introducing distortions or losing data? Should I apply some form of scaling for the SAR band before stacking?

I’m using Python for preprocessing, and libraries like rasterio, numpy, and matplotlib for image manipulation. I am open to suggestions on advanced methods such as histogram matching, quantization, or other advanced preprocessing techniques to retain the high dynamic range of these images.

Any help or guidance on how to approach this preprocessing pipeline would be greatly appreciated. Thanks in advance!

link for the images: text

Continue reading...