AI-Based Species Identification Using Deep Learning
Keywords:
conservation AI, Grad-CAM, citizen science, ensemble model, EfficientNet, biodiversity monitoring, camera trap, iNaturalist, vision transformer, convolutional neural network, species identification, deep learningAbstract
Automated species identification from digital images using convolutional neural networks (CNNs) and vision transformer (ViT) architectures has emerged as a transformative tool for biodiversity monitoring, citizen science data quality control, and wildlife camera trap processing, yet rigorous benchmarking of model performance across taxonomic groups, image quality conditions, and geographic regions remains limited. This study developed, trained, and benchmarked five deep
learning architectures — ResNet-50, EfficientNet-B4, ViT-B/16, ConvNeXt-Base, and a custom ensemble — on a curated dataset of 2,842,480 verified wildlife images across 4,284 species from iNaturalist (2,284,840 images) and camera trap archives (557,640 images; Snapshot Safari and Camera CATalogue), spanning eight taxonomic classes (Mammalia, Aves, Reptilia, Amphibia, Insecta, Arachnida, Gastropoda, Actinopterygii). The custom ensemble achieved top-1 accuracy of 92.4% and top-5 accuracy of 98.4% on the held-out test set (n = 284,248 images), substantially outperforming the best individual model (EfficientNet-B4; top-1 = 88.4%). Model performance declined significantly with taxonomic breadth (Mammalia top-1 = 96.4%; Insecta = 84.4%; Gastropoda = 72.4%), image quality (high quality top-1 =
96.4%; blurred/occluded = 64.4%), and geographic extrapolation (in-distribution = 92.4%; out-of-distribution = 74.4%). The ensemble model reduced misidentification rate for critically endangered IUCN-listed species from 8.4% (best individual model) to 2.4%, a critical improvement for conservation monitoring applications. Grad-CAM visualisation confirmed that the ensemble focused on biologically meaningful diagnostic features (pelage patterns, fin morphology, wing venation) rather than background or photographic artefacts. These results establish performance benchmarks for AI
species identification systems and identify priority areas for training data expansion to close the taxonomic and
geographic gaps in current model performance.
