Deep learning-based body length estimation in soil-dwelling arthropods

Abstract

Abstract Body length is a fundamental functional trait in soil ecology used to estimate biomass and metabolic rates, but manual microscopic measurement is a major high-throughput bottleneck. Here, we introduce a device-independent Deep Learning (DL)-based regression framework for automated body length quantification of soil-dwelling arthropods from top-view digital images. Using a robust MaxViT-T backbone combined with image aspect-ratio metrics, the framework was validated across three distinct laboratory and field experiments without requiring manual taxonomic pre-sorting. In high-end laboratory stereomicroscopy (Test 1), the model achieved a global R 2 of 0.94 and a Mean Absolute Error (MAE) of 0.059 mm. To test cross-platform robustness, an independent external blind test was conducted on an unseen stereomicroscope-camera setup (Test 2), where the model maintained high predictive performance (R 2 = 0.96, MAE = 0.054 mm). For automated field extraction systems across 16 macro- and mesofauna groups (Test 3, N  = 1,807), the pipeline achieved an overall R 2 of 0.98 and a global MAE of 0.039 mm. Compared to conventional contour-based edge detection, which systematically introduced 3-fold higher errors due to organism curvature, the DL model maintained geometric precision across complex taxonomic body plans. These results demonstrate that deep learning computer vision provides an accurate, reproducible, and scalable framework for high-throughput trait-based ecological and biomass assessments.