Deep learning for genera-level bark beetle taxonomic classification

With the ongoing climate change and increasingly frequent extreme weather events, there has been a concerning rise in bark beetle (Scolytinae) outbreaks in temperate forests, posing a major threat to biodiversity and the forest economy. Rapid and accurate species identification is crucial, but remains challenging due to their tiny size, strong morphological similarities, and the limited availability of high-quality reference images. By combining entomological expertise with artificial intelligence, this work proposes an innovative tool to tackle the urgent risks of forest health and resilience in climate variability crisis. We have developed classification benchmark for genus-level bark beetle, utilizing about 1600 images varying in qualities and settings across 13 classes. Several deep learning models were evaluated, including AlexNet, SqueezeNet, VGG, and existing ResNet architectures. Best results are obtained from ResNet-50, yielding remarkable 96 % of validation accuracy. Along with other measurement standards of 97 % precision, 96 % recall and F1-score, all these performances are also in agreement with multi-class confusion matrices. These findings emphasize highly reliable classification accomplishment, with high degree of the confidence between our prediction and reality matching. Our automated identification system has helped professionals in forestry, ecology, and biology by minimizing labor-intensive tasks, facilitating large-scale taxonomic work, and later enabling real-time analysis. Beyond taxonomy, our system provides a practical strategy for tracking pest population dynamics and movement across regions, anticipating potential outbreaks, and supporting more adaptive forest management. The preventive use of this new application can contribute to maintaining ecosystem stability and conserving natural forests in the future. • Currently no identification of Scolytinae with machine learning techniques. • Scolytinae are very difficult to distinguish even by the experts. • Deep neural networks were trained to recognize 13 genera using 1600 images. • 96 % of validation accuracy is obtained for this automatic classification.