[Research] Develop a CNN-based Classifier for PSR vs. Sunlit Patch Extraction

[TechieSamosa]

🛑 Research Track Issue

Please read RESEARCH.md before proceeding. Do not submit a Pull Request for this issue without first commenting a brief proposal and receiving approval from the core team.

The Problem:
Our scripts/extract_patches.py currently uses a hardcoded, heuristic photometric threshold to decide if a 64x64 lunar patch is psr (dark) or sunlit (bright). This is brittle and occasionally misclassifies penumbra regions.

The Research Goal:
We want to replace this heuristic math with a lightweight, trained Convolutional Neural Network (CNN) classifier.

What you would do:

1. Build a very small, fast PyTorch CNN (e.g., MobileNetV3-small or a custom 3-layer CNN).
2. Train it on our existing dataset of extracted patches to output a binary classification (0 = PSR, 1 = Sunlit).
3. Integrate the model into the extraction pipeline so it dynamically sorts the 1.2 GB image slices.

To claim this: Comment below with your proposed lightweight architecture and how you plan to handle the severe class imbalance (there are far more sunlit patches than true PSR patches).