A Hierarchical Approach to Identify Snake Species with Resemble Visual Features

Indonesia is home to over 349 snake species, representing approximately 10% of global snake diversity, this diversity shaped by its tropical climate and unique ecosystems. However, accurate species identification is often confounded by Batesian mimicry, where non-venomous species resemble venomous ones, posing challenges for both research and public safety. This study focuses on 10 native snake species, equally divided between venomous and non-venomous groups, and leverages a systematically curated SnakeCLEF 2023 dataset. Refinements addressed mislabeling, irrelevant content, and class imbalance, enhancing its reliability for model training and evaluation. A hierarchical multistage classification framework, utilizing PoolFormer, was developed to first classify specimens into mimicry-informed groups before performing species-level differentiation. Experimental evaluations indicate that the proposed model surpasses leading architectures, including ConvNeXt and Vision Transformer, achieving an accuracy of 0.84 and an F1-score of 0.82 on the curated dataset. These findings highlight the importance of rigorous dataset preprocessing and validate the hierarchical multistage paradigm in addressing interspecies classification ambiguities. Future research could incorporate generative modeling techniques to enhance dataset diversity and further improve classification outcomes, supporting broader applications in conservation and biodiversity monitoring.