WUSA-SDZWA-v1

This repository contains open-source models redistributed for easy integration with AddaxAI, hosted by Addax Data Science. Each model retains its original license (see license files) and attribution. We comply with all original license terms. Users must review and comply with individual model licenses before use. See below for detailed model information including original sources, licenses, and attributions.

<strong>Developer & Owner</strong>

San Diego Zoo Wildlife Alliance

<strong>Links</strong>

<ul> <li><a href="https://github.com/conservationtechlab">Learn more</a></li> <li><a href="https://opensource.org/license/mit">License</a></li> </ul>

<strong>Model</strong>

Western USA v1, an EfficientNet V2 Medium classifier trained with animl-py by Kyra Swanson (SDZWA) in 2026. It covers 51 species and groups across the western United States and expands the Southwest USA v3 model to the whole region. Trained on 555,668 images, mostly from LILA datasets, with 94.9% overall accuracy on a 46,238 image evaluation set.

<strong>Files</strong>

<ul> <li><code>sdzwa_westernus_v1_efficientnet.pt</code>: the checkpoint as delivered by the author (state dict plus training stats). The ImageNet weights that Southwest USA v3 ships are not needed: this checkpoint holds every parameter.</li> <li><code>classes.csv</code>: class order as the model emits it. This is the order of the author's evaluation confusion matrix, not the alphabetical list that came with the model; verified on 81 labelled ENA24 crops (85% top-1, all misses plausible confusions). The <code>bird_other</code> label from that list is <code>bird</code> here for the same reason.</li> <li><code>inference.py</code>: AddaxAI inference (299x299 resize, no normalisation, 0 pixel crop buffer, as in animl-py).</li> <li><code>taxonomy.csv</code>: taxonomy ranks per class, resolved against the GBIF backbone.</li> </ul>