Deep Learning-Based Methods for Automated Estimation of Insect Length, Volume, and Biomass

Accurate information on insect biomass and size is fundamental for studying insect behavior, ecology, and decline. However, most methods are labor-intensive, often invasive, and impede large-scale studies. Here, we introduce two novel, non-invasive, deep learning-based methods to automatically measure key insect traits from images. First, we introduce a general approach using Oriented Bounding Boxes (OBB) designed for broad applicability across diverse insect taxa. By adjusting for specimen orientation, this method measures length accurately with a mean absolute error (MAE) of 0.211 mm compared to conventional photomicroscope measurements and provides initial biomass estimates using established length-weight relationships. Second, we tested whether a specialized segmentation model requiring taxon-specific training improves biomass predictions. We show that a model for Tachinidae (Diptera, Calyptratae) can effectively delineate key body parts (head, thorax, abdomen) and provide accurate curvilinear body length, volume, and biomass estimates, correlating strongly with both wet (R = 0.937) and dry (R = 0.907) weight. Validation experiments demonstrate that our methods are accurate and offer advantages over traditional techniques by reducing handling and improving scalability. This is important because developing similar segmentation models for just the 20 most dominant flying insect families collected in bulk samples like Malaise traps could cover around 50% of specimens, and the modular design of our segmentation pipeline allows for seamless updates to include additional families. Together, these two new approaches for morphometric analysis provide an efficient, scalable framework for automated insect data acquisition, with broad implications for diverse ecological and evolutionary studies, ranging from functional trait ecology and population-level assessments to large-scale biodiversity monitoring and conservation initiatives.