Test regression model for preprocessing

[hpatel0816]

No description provided.

[hpatel0816]

Here is a quick summary of the key decisions:

• To handle class imbalance, I kept all the positive pixels and downsampled to 1:50 ratio of positives to negatives
• Split the train/test dataset by tile, not by pixel
• Used streaming training since computer couldn't handle loading in all the image pixels at once
• Since we had extreme class imbalance and our goal is to reduce image size while retaining most of the plume regions, I tried to evaluate based on 1) plume coverage and 2) the fraction of region of interest relative to entire image instead of using raw accuracy.

The core results were:

The best coverage preserving config:

• Avg plume coverage = ~97%
• 10th percentile coverage = ~94%
• Mean ROI fraction = ~45-50%

The best ROI reducing config:

• Avg plume coverage = ~83%
• 10th percentile coverage = ~45%
• Mean ROI fraction = ~10%