Description
Over the past three decades, flying insect biomass has declined by more than 75% in protected areas, highlighting the urgency of monitoring insect populations. Moths represent approximately 9% of all known species on Earth and are among the most affected insect groups, despite their important roles as pollinators and food sources. Automated nocturnal light recorders are increasingly used to monitor moth populations, but extracting meaningful species information from the resulting images remains challenging due to the fine-grained differences between many species. Addressing this challenge is essential to provide entomologists and conservation researchers with more reliable tools for large-scale moth monitoring. However, automated moth classification pipelines have not kept pace with recent advances in large-scale training data and vision architectures. To address this gap, we build on the pipeline of Korsch et al. (2022), which provides a solid baseline for automated species recognition, and extend it in two ways. First, we treat web-sourced image collections, museum archives, citizen science platforms, and light trap cameras as heterogeneous external knowledge sources and integrate them by removing cross-dataset duplicates and resolving label ambiguity. Second, we investigate the impact of vision-language pretraining, exemplified by BioCLIP, an image-text model pretrained on ecological data, on fine-grained species classification. By aligning visual and taxonomic textual knowledge, this approach represents a step toward AI-based ecological synthesis, and we compare it against purely image-based convolutional and transformer architectures pretrained on generic versus ecological data. Our initial results show that BioCLIP achieves significant gains over the reproduced baseline from Korsch et al. (2022), though evaluation of the benefits of combining datasets is still ongoing.
| Topic | Topic 4: Can knowledge graphs and AI-based synthesis improve ecological synthesis? |
|---|