Efficient Avian Biodiversity Monitoring through Optimized Knowledge Distillation — Code, Model Weights and Reproduction Material
Year: 2026
Type: software
Abstract: Code, trained model weights and reproduction material for the manuscript Efficient Avian Biodiversity Monitoring through Optimized Knowledge Distillation, submitted to IEEE Latin America Transactions (Computing section). The study addresses fine-grained classification of 134 bird species of the Tolima department, Colombia, under two coupled constraints: discrimination among morphologically near-identical sympatric species, and the absence of the computing infrastructure that high-capacity models require. An EfficientNetB3 teacher (97.62% accuracy) supervises lightweight students through softened logits (T = 4, α = 0.15). Main results. Distillation lifts a MobileNetV2 student from 89.51% — its ceiling under conventional transfer learning — to 94.28% accuracy at 5.46 ms inference latency and 9.83 MB on disk, matching a native edge architecture (MobileNetV3, 94.40%) while being faster and 27.5% smaller. A MobileViT student reaches 96.75% at 4.79 ms, exceeding its own state-of-the-art counterpart (MobileViT-XXS, 96.00% at 4.99 ms) on accuracy and latency simultaneously. Statistical robustness over five independent runs with fixed random seeds: 94.96% ± 0.23%. Every experiment was trained and timed on a single consumer laptop, with no cloud infrastructure. Contents. notebooks-experiments.zip — the full experimental pipeline, from dataset preparation and the VGG16 baseline through the teacher–student distillation core, the state-of-the-art comparison, the five seeded runs and the per-class long-tail analysis. notebooks-explainability.zip — Grad-CAM++ and LIME audits verifying that the distilled network attends to diagnostic morphology (the yellow auricular patch of the endangered Ognorhynchus icterotis, the ocular mask of Dacnis hartlaubi) rather than to habitat context. models-student-weights.zip — the deployed distilled MobileNetV2 student plus the five seeded checkpoints. reproduction-artifacts.zip — stored test-set predictions (6,123 images), class labels, and verify_paper_metrics.py, which recomputes every reported metric in seconds without a GPU, a retraining run, or a dataset download. deployment-docker.zip — EcoCurador Pro, the Streamlit auditing application, containerized. A single docker compose up reproduces the inference environment. figures-paper.zip — figures and the notebooks that generate them. On the image corpus. The raw records originate from the Macaulay Library (eBird, Cornell Lab of Ornithology). Copyright remains with each individual photographer and Creative Commons licensing on eBird is opt-in, so this deposit does not redistribute the images. It ships the code, the weights and the stored predictions, which is sufficient to verify every reported result without redistributing third-party work. Code and weights are released under the MIT License.
Institution: Universidad Cooperativa de Colombia
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204210067