Carabids hierarchical classification with ensemble model architecture
Gabin Moulie, Marie Beurton‐Aimar, Matthieu Valé · 2024
Specimen classification is vital for ecological research, biodiversity monitoring, and animal conservation. Traditionally labor-intensive, this process faces scalability challenges in our era. However, advances in machine learning and computer vision technologies have transformed classification tasks. Automated wildlife classification systems, exemplified by platforms like Wildlife Insights and MegaDetector, revolutionize ecological research by interpreting visual data from images and videos at unprecedented speeds.Despite the advancements of Machine Learning algorithms, challenges persist, particularly regarding adaptability to specific taxonomic groups and environmental contexts. While mega-classifiers offer extensive coverage across taxa, achieving accuracy comparable to human classifiers often requires models tailored to specific tasks. The interest in classifying Carabids lies in their role as bioindicators of soil functions such as pest regulation or nutrient cycling.To address this problem, a hierarchical model architecture for Convolutional Neural Networks (CNNs) has been developed, incorporating both EfficientNet and ResNet-101 to merge them into a single algorithm. It was trained from scratch to enable accurate classification of Carabids. The image dataset used for training is hosted on the Kaggle website. The method used for classification in two stages is a hierarchical classification method. Hierarchical classification makes sense because biologists classify species by genus and then by species.In the classification of Carabids species, high accuracy levels were achieved, with 94.21% of images classified correctly at the species level and 96.60% at the genus level. Combining all algorithms yielded an overall precision of 91.01%, indicating promising results considering the physical similarities among species.Further optimization avenues include fine-tuning algorithm models, exploring concepts like Mixture of Experts (MoE), and enhancing data preprocessing techniques. Collaboration with domain experts is crucial for refining classification models and continuous evaluation and validation in real-world scenarios are essential for improving effectiveness and reliability. These efforts will advance species-level classification of Carabids and contribute to ecological research and conservation efforts [12] [13].