[Draft] Filter images before clustering

[mihow]

Work by @Yuyan-C

[netlify]

❌ Deploy Preview for antenna-preview failed. Why did it fail? →

Name	Link
🔨 Latest commit	f1a52ef
🔍 Latest deploy log	https://app.netlify.com/projects/antenna-preview/deploys/68a6b7334aa3900008012a35

[netlify]

❌ Deploy Preview for antenna-ood failed. Why did it fail? →

Name	Link
🔨 Latest commit	f1a52ef
🔍 Latest deploy log	https://app.netlify.com/projects/antenna-ood/deploys/68a6b731138e830008ace878

[mihow]

I suggest we open a new branch from main and pull in several of these filters for the regular classification task (and keep clustering it's own experimental branch)

Filters:

• Filter crops that are too small
• Filter crops that are too blurry
• Filter crops on the edge

These can be first implemented in a post-processing job, which can be run on existing predictions. It will add a prediction to existing occurrences and update their determination. If the occurrence does not pass the filters, a new prediction should be added "Not identifiable".

Then we can consider when & how it is run in the existing pipeline. I suggest this pattern for all post-processing steps (regional class masking, rank predictions, etc)

After the post-processing implementation, look into adding these filters to the prediction service itself (the AMI Data Companion). And adjust the confidence score or reject the prediction at the source, before sending to Antenna. Do users want to see blurry & small images with low confidence, or do they want them hidden all together? (no classification & not included in response). Or return the unidentifiable image with sore of 0 and taxon "Not identifiable".