HISNET-FF: Hierarchical Identification of Species using a Network with Fused Cranial and Dental Features

Accurate species identification from mammalian craniodental features is essential but traditionally slow and requires specialized expertise. We address this by developing HISNET-FF, a deep learning framework featuring a dual-branch architecture to fuse global features from the cranium and local features from the teeth and auditory bullae. The network employs a hierarchical pipeline, first classifying to genus and then to species. Tested on a comprehensive image dataset of the Family Talpidae (18 genera, 51 species), HISNET-FF achieved exceptional accuracy at both the genus (99.6 ± 0.4%) and species (96.5 ± 1.3%) levels. This species-level accuracy significantly outperforms single-modality approaches, including both flat (up to 91.2 ± 2.3% accuracy) and hierarchical (up to 93.9 ± 2.1% accuracy) strategies. To enable a fully automated workflow, we also developed a YOLO-based tool that annotates diagnostic features with high performance, achieving 97.8% recall, 97.9% precision, and 81.5% mean average precision (mAP@.50:.95). This automation resulted in a minor drop in final identification accuracy of 1.9%. HISNET-FF thus provides a robust and highly accurate framework that can accelerate morphology-based research, with strong potential for broader application.