Aerial Biological Target Classification Based on Time–Frequency Multi-Scale Feature Fusion Network

Migrating insects and birds are the primary biological targets in the aerial ecosystem. Radar is a powerful tool for monitoring and studying aerial animals. However, accurately identifying insects and birds based on radar observations has remained an unsolved problem. To address this research gap, this paper proposed an intelligent classification method based on a novel multi-scale time–frequency deep feature fusion network (MSTFF-Net). A comprehensive radar dataset of aerial biological targets was established. The analysis revealed that radar cross section (RCS) features are insufficient to support insect and bird classification tasks, as aerial biological targets may be detected in radar sidelobes, leading to uncertainty in RCS values. Additionally, the motion characteristics of insects and birds are complex, with diverse motion patterns observed during limited observation periods. Simple feature extraction and classification algorithms struggle to achieve accurate classification of insects and birds, making aerial biological target classification a challenging task. Based on the analysis of insect and bird features, the designed MSTFF-Net consists of the following three modules. The first module is the amplitude sequence extraction module, which abandons traditional RCS features and instead extracts the dynamic variation features of the echo amplitude. The second module is the time–frequency feature extraction module, which extracts multi-scale time–frequency features to address the complex motion characteristics of biological targets. The third module is the adaptive feature fusion attention module, which captures the correlation between features to adjust feature weights and achieve the fusion of different feature types with varying representations. The reliability of the classification algorithm was finally verified using a manually selected dataset, which includes typical bird, insect, and other unknown targets. The algorithm proposed in this paper achieved a classification accuracy of 94.0% for insect and bird targets.