Repository navigation
Support for "geo-fencing" approach to improve prediction accuracy for regional species list #797
Description
Summary
As an alternative to training regional models, we should be able to use the Taxa List feature in a post-processing step to provide much more accurate predictions.
Detailed Description
TBD
Implementation Details
Filter the global model predictions (30k) to a regional list and re-weight the softmax scores
Related Issues
Additional Context
Needed for CFIA moth monitoring project (ports of entry)
Activity
[-]Support for "geo-fencing" approach to improve prediction accuracy of global model[/-][+]Support for "geo-fencing" approach to improve prediction accuracy for regional species list[/+]jbowden18 commentedon Apr 3, 2025
if there was a user friendly approach to filtering species to create ones own 'regional' pipeline would be great --also species come and go
mihow commentedon Jul 17, 2025
Notes from call about implementation:
Receive all raw classifications from the processing backend (e.g the global species classifier)
All occurrences have a determination, which is a Taxon from the global classifier
Import the regional list of species that exist for a given deployment
Run post-processing step in Antenna
Each "Classification" has a prediction for all 20k species in the global list (every item in the model's category map). Each entry is associated with a Taxon instance in the Antenna DB
The regional taxa list has a subset of species names that were matched to Taxon entries in the Antenna DB when imported
Create a new CategoryMap with the subset of labels
Create a new Algorithm entry (called "World Moths (filtered for Pennsylvania)")
A,5,0.9; B,2,0.1; C;1,0.1;
A,5,0.9; C,1,0.2;
The masking function will remove all items from the labels, scores & logits fields that have the same index - which is mapped to a species which is not in the regional list
The scores will be recalculated so they add up to 1 (softmax)
Some occurrences will have a new determination
"Zeroing" entries instead of deleting them is another approach to consider. (set scores of unwanted species to zero, rather than deleting them) - Needs to be a very low negative number since these are logits (zero could be center).
mihow commentedon Jul 17, 2025
Usage in our test branch
feat/logistic-binningmihow commentedon Sep 3, 2026
This is now being pursued as #1364 / #1367 (build a project taxa list from a region, and track which species a model can predict), building on class masking (#999) and the post-processing framework (#1289). Leaving open as the original geo-fencing proposal, since #1364/#1367 cover the 'generate the list' half rather than every idea here, but linking so the two threads don't diverge.