IUCN Red List Threatened Bird Classification from Citizen-Science Images using Fine-Tuned CNN Features and a Hierarchical Lightweight Classifier

<jats:p>To prevent bird strikes at wind farms, an approaching protected bird must be identified so that the turbine can be halted before a collision. We present a bird-species classification module intended for that real-time, low-power edge setting: a hierarchical system assigning IUCN Red List categories to in-the-wild citizen-science bird images, combining a ResNet50 feature extractor fine-tuned on a bird-domain corpus with lightweight LightGBM classifiers in a five-stage cascade following the taxonomic hierarchy. We evaluate on 84 species x 10 images (840) plus 840 held-out non-bird images, so that every stage is scored against its own denominator, and all paired tests resample species rather than images. Fine-tuning improves Stage 4 top-1 accuracy and macro recall from 0.273 to 0.404 (+13.1 pt; Holm-adjusted p = 0.015), whereas neither the top-5 gain nor the end-to-end IUCN gain survives correction. Training a flat 71-class Stage 4 and a genus-level hierarchical variant to verified convergence under identical caps, we detect no difference, and a pre-specified 5 pt non-inferiority test is inconclusive. Collapsing the order and species stages into a single 74-class classifier likewise changes no accuracy metric significantly, while raising alert precision from 0.765 to 0.929 at the same per-object cost, so the two stages should be merged and four stages suffice. A controlled down-sampling study finds the cascade degraded but usable at 128x128 and broken at 32x32, where an apparent rise in protected-species alert recall is an artefact of precision collapsing to 0.10. That alert, the operational output, is limited at native resolution too (recall 0.25-0.33). On the target device, a Jetson Orin Nano, accuracy is unchanged (p = 1.000) at 226 ms per object; the cost lies in the Stage 4 classifier and JPEG decoding rather than the backbone, and tuning both - Stage 4 truncated to 500 rounds at equal accuracy, the crop supplied in memory as the camera does - gives 112 ms on the CPU and 13 ms with the GPU, a usable speed for the application.</jats:p>