An Insect Swarm Segmentation Algorithm Based on Two-dimensional Spatiotemporal Aggregation Pattern Attention Mechanism

Radar observations have identified a phenomenon where insects exhibit collective migratory behavior, characterized by uniform flight times, altitudes, and directions. Radar-detected insect swarms display aggregation in both spatial and temporal dimensions, with indistinct boundaries and varied shapes. Different swarms often merge into a single entity, making it challenging to accurately extract individual swarms using simple multi-parameter segmentation methods. To address these issues, this paper proposes an insect swarm segmentation method based on a two-dimensional spatiotemporal aggregation pattern attention mechanism. Building upon the Mask R-CNN architecture, this method leverages spatial attention mechanisms and circular soft nonmaximum suppression algorithms to enhance spatial position sensitivity and reduce occlusion effects between adjacent swarms. This approach transforms the two-dimensional clustering problem into an image instance segmentation problem and demonstrates its effectiveness in the automatic segmentation of insect swarms.