Description

We want to be able to implement a 3D SISR CNN, so that we can increase spatial and spectral resolution on two different dimensions at the same time. We want to be able to achieve a spatial resolution of 90m (meaning that every pixel in the real world has a size of 90m by 90m) although a lower resolution is better. Right now our original image has a spatial resolution of 150m, which means a simple scale factor of 2 should help us solve this issue.

Here is a research paper I want to us to read and eventually implement what is discussed here:

https://ieeexplore.ieee.org/abstract/document/8974222

Log in using your utorid, you should be able to access it, and if not ping @Nooh

The same people who wrote this paper have also included the model’s code, so we will be using that as a starting point, and using the model they specified here.

https://github.com/NourO93/SISR_Library/tree/main?tab=readme-ov-file

The research paper talks a little about spectral unmixing, so we may need to have science’s input here as well.

Tasks

A list of sequential tasks for the chapter - the chapter is complete when all tasks are complete

• [x] Read and Understand SISR model
• [x] Populate the Notes function with Progress so far/ summaries of what you learnt so far
• [ ] Have discussions with science on the implementation of spectral unimixing
• [x] Collect dataset to train model
• [x] Start creating model and write training code to train the model
• [x] Try to get the psnr and ssim metrics working with the model
• [ ] Start comparing the model to the bicubic model you implemented earlier in terms of psnr and ssim
• [ ] make refinements to the code, and try tuning hyperparamters for model accuracy

Notes

Use this space for research notes, open questions, progress logs

• Hyperspectral Imagery (HSI) has high spectral resolution but suffers from low spatial resolution due to sensor tradeoffs.
• Single Image Super-Resolution (SISR) is a computer vision technique used to enhance the spatial resolution of a low-resolution (LR) image by estimating missing high-frequency details to generate a high-resolution (HR) image.
• Most enhancement techniques introduce spectral distortions in the reconstructed High-Resolution (HR) HSI.
• Traditional SISR methods (interpolation-based) introduce blurriness, making CNN-based methods more effective.