AraNet: A Two-Stage Convolutional Neural Network for Spider Species Classification

More than 50,000 spider species have been identified worldwide, and efficient species identification is important for applications ranging from habitat monitoring to emergency medical treatment. Prior studies have shown that deep learning can be effective for spider classification on limited datasets. To extend this work, we use recent deep learning techniques that have proven effective in other domains. Our classifier, AraNet, is a two-stage convolutional neural network based on the ConvNeXt-V2 model design. AraNet sequentially combines two ConvNeXt networks: the first performs intermediate fine-tuning on a broad set of species, and the second freezes the early backbone layers and fine-tunes only the later layers for species of particular relevance. The model is trained using a large dataset comprising 1000 spider species and 548,689 images. On a 100-species classification task, AraNet achieved 97.4% accuracy, outperforming prior single-stage approaches. AraNet also achieved 99.5% accuracy when evaluated on medically significant Australian spiders, improving over previous spider classification models.