Deep Learning and Bioengineered Feature Engineering for Automated Taxonomic Identification of Mediterranean and Atlantic Demospongiae

Abstract

Given morphological convergence, cryptic speciation, and severe class imbalance in regional collection records, accurate taxonomic  dentification of marine Demospongiae remains a major challenge in biodiversity monitoring. This study presents a novel computational framework that integrates imbalanced multi-class deep learning with biologically informed feature engineering to automate the hierarchical classification of 503 Demospongiae specimens collected from Mediterranean and Atlantic habitats. We introduce a biologically grounded feature construction pipeline that encodes morphological, ecological, and evolutionary priors derived from sponge ontology structures, generating high-dimensional representations compatible with deep neural architectures. To address strong distributional skew across 7 orders, 42 families, 114 genera, and 230 species, we implement a hybrid imbalance mitigation strategy combining topology-aware synthetic minority oversampling, adaptive class-weighted sampling, and hierarchical focal loss. The proposed architecture further employs a multi-task graph convolutional network to jointly learn taxonomic relationships while preserving hierarchical constraints. Experimental evaluation demonstrates substantial improvements over conventional machine learning baselines, achieving macro-averaged F1-scores of 0.89 at the order level and 0.76 at the species level, with notable gains in recall for underrepresented Atlantic taxa. Ablation analyses further indicate that the incorporation of bioengineered features significantly enhances model generalization.