MFPNet: efficient mid-level fusion of visible-infrared features for real-time UAV forest fire detection

Abstract

To address the limitations of unimodal forest fire detection algorithms, which suffer from high false positive rates and insufficient robustness under complex illumination variations, smoke occlusion, and non-fire heat source interference, this paper proposes MFPNet, a functionally decoupled mid-fusion framework for visible-infrared dual-modal UAV forest fire detection. Unlike existing mid-fusion methods that either focus solely on local cross-modal complementarity or global semantic modeling, we explicitly decouple these two functionalities and synergistically stitch them at the P4/P5 feature pyramid levels through a parallel mechanism. Specifically, the feature complementary mapping module is introduced to explicitly model cross-modal complementarity and suppress modality-specific noise via bidirectional channel-spatial interaction, while the Pyramid Sparse Transformer module—integrated into a dual-modal mid-fusion pipeline in this study—performs efficient multi-scale global context modeling through high-level semantics-guided sparse attention with parameter sharing. By combining these two modules via carefully designed parallel stitching at the same mid-fusion stage, MFPNet achieves complementary performance gains beyond simple addition that neither module can deliver independently. This design provides a valuable reference for leveraging bimodal complementary advantages in UAV-based forest fire detection tasks. Extensive experiments on our proprietary forest fire dataset demonstrate that MFPNet improves mAP@0.5 from 78.0% (Mid-Fusion baseline) to 82.4 ± 0.38% while maintaining real-time inference at 108 FPS. Further validation on three public multimodal benchmarks shows consistent performance of the proposed functionally decoupled stitching strategy under the evaluated settings.