<p>In the field of wildfire smoke detection using tower-mounted camera systems, conventional algorithms rely on static visual features such as color, texture, and shape for recognition. However, these methods are prone to false positives when encountering visually similar objects such as clouds or sky, largely due to their inability to model dynamic temporal evolution. To overcome this limitation, we propose SmokeFormer—a novel spatiotemporal smoke detection framework that incorporates a hybrid attention-based spatiotemporal encoder. This encoder dynamically captures both local discriminative details and global contextual information through parallel local-global mechanisms. The architecture is further strengthened by a dual prediction branch that simultaneously learns fine-grained local features and multi-scale global patterns. During inference, an optimized decision fusion mechanism integrates predictions from both branches to ensure robust detection performance. Extensive experiments on the large-scale SmokeData forest fire smoke dataset demonstrate the superiority of SmokeFormer. On the test set, it achieves an accuracy (ACC) of 89.62% and an F1 score of 90.07%, comparable to those of the Swin-small model (ACC 89.78%, F1 90.05%) while reducing computational cost by 60.4%. Notably, in challenging scenarios such as terrain-blended smoke and dissipating smoke, SmokeFormer significantly reduces false detections, effectively balancing detection accuracy and computational efficiency.</p>