HISNET-FF: Hierarchical identification of species using a network with fused cranial and dental features

Accurate taxonomic identification based on mammalian craniodental features remains critical for evolutionary, ecological, and paleontological research, yet conventional approaches are time-intensive and demand expert input. To overcome these limitations, a deep learning framework, HISNET-FF, was developed with a dual-stream architecture that integrates global cranial morphology with local diagnostic signals from teeth and auditory bullae. The model operates within a hierarchical classification pipeline, processing from genus-level discrimination to species-level resolution. Evaluation on an extensive image dataset encompassing 51 species across 18 genera of Talpidae achieved exceptional accuracy at both the genus (99.6%±0.4%) and species (96.5%±1.3%) levels. This species-level performance substantially exceeded that of single- stream models employing either flat (91.2%±2.3%) or hierarchical (93.9%±2.1%) strategies. To support end-to-end automation, a YOLO-based annotation module was implemented to localize key morphological traits with 97.8% recall, 97.9% precision, and 81.5% mean average precision (mAP@[.50:.95]). Incorporating this module incurred only a marginal reduction of 1.9% in identification accuracy. Thus, HISNET-FF offers a robust and accurate framework that accelerates morphology-based species identification and enables automated taxonomic classification, with strong potential for broader implementation across diverse biological research domains.