Intelligent Identification of Migratory Insects Using Searchlight Trap Statistics and Swin Transformer-Based Radar Echo Analysis

The insect migration significantly impacts agriculture, making accurate identification of migratory insects essential for pest management. However, constructing radar-based insect species identification datasets traditionally requires costly and time-consuming manual measurements from specialized laboratory equipment or field experiments. Existing identification methods rely on precise radar cross section (RCS) calibration or complex statistical analysis of long-term radar data, both of which hinder the long-term automated identification of migratory insects. This study presents an innovative approach by integrating searchlight trap statistics with radar monitoring to construct a radar-based insect species identification dataset. This method overcomes the need for manual data collection and mitigates the reliance on radar calibration, offering an efficient way to generate high-quality datasets. We also introduce a novel deep learning architecture, the contrastive learning (CL)-global context (GC)-Swin Transformer, which combines CL and GC modules to enhance species identification from radar-derived time-frequency maps, which primarily capture insect micromotion features. The approach is applied to a dataset of five migratory insect species, achieving an average identification accuracy of 88.4%, outperforming existing radar-based techniques and classical deep learning models. This research not only improves dataset construction efficiency but also provides a powerful new solution for radar-based insect identification.