Abstract: Aimed at the problem that the small size of pyrotechnic targets in aerial forest fire images makes them difficult to be detected, an improved forest fire small target detection algorithm (SSF-YOLOv7) is proposed. A high-order spatial interaction module (C3HB) was designed to enhance the model’s ability to capture the global information of pyrotechnic targets. A non-parametric attention mechanism was added to improve the recognizability of pyrotechnic targets, and K-means++ clustering algorithm was used to regenerate suitable anchor boxes. SAHI (Slice Aided Hyper Inference) technology was integrated into the model to solve the problem of insufficient details of pyrotechnic targets. The experimental results show that compared with YOLOv7, SSF-YOLOv7 has an increase of 3.7 percentage points in mAP₅₀, an increase of 5.8 percentage points in the average accuracy of pyrotechnic small targets, and an increase of 19.8 FPS.