FireMamba: An efficient visual representation learning framework for fire detection

Transformers dominate sequence modeling, but their quadratic self-attention makes high-resolution vision computationally expensive. State-space models such as Mamba offer linear time token mixing, yet have not been fully realized as vision backbones because images demand both positional sensitivity and global context. We present FireMamba, a bidirectional state space network that removes self attention while preserving rich spatial reasoning. First, the input image is processed by a multiscale spatial attention module (MSAM), which aggregates evidence across multiple receptive fields and pooling paths to produce a spatial attention map that suppresses clutter and emphasizes fire related cues. The refined features are then forwarded to FireMamba, which uses a squeeze-expand channel pathway and bidirectional state-space mixing to model long range dependencies efficiently with linear time complexity. This front end attention followed by linear mixing focuses computation where it matters while maintaining high throughput. Experiments over DVF, Foggia, and DF datasets shown that FireMamba outperforms strong convolutional and transformer baselines, delivering state-of-the-art accuracy for applications such as early wildfire monitoring, industrial safety inspection, UAV and robotic inspection, and urban video surveillance. • An efficient and memory-optimized fire detection model based on a bidirectional SSM backbone. • MSAM fuses multiscale cues to emphasize fire regions in diverse scenes. • Ablation study highlights contributions of SSM, Fire blocks, and MSAM modules. • Achieves highest accuracy on DVF, Foggia, and DF fire detection benchmarks.