AAFNet: An Asymmetric Feature Encoding and Active Energy Modulation Network for Multimodal Remote Sensing Forest Fire Detection

Abstract

Forest fires pose severe ecological threats, yet dedicated multimodal architectures for such complex environments remain largely underexplored. Generic multimodal methods struggle to decouple heterogeneous physical properties, resist feature contamination during dynamic fusion, and preserve the fragile boundaries of microfires against deep semantic smoothing. To address these limitations, we propose an asymmetric feature encoding and active energy modulation network (AAFNet) for multimodal remote sensing forest fire detection. First, an asymmetric feature encoding is designed at the source. Within this architecture, the geometry-aware anchor (GAA) module is proposed to adaptively fit the nonrigid boundaries of smoke via decoupled deformable sampling. In parallel, the thermal-aware salient (TAS) module performs hierarchical wide-range perception operators to capture long-range thermal gradients and suppress radiation artifacts. Second, an energy-guided modulation (EGM) mechanism is formulated to dynamically allocate modality weights by quantifying instantaneous energy responses. This active regulation suppresses background interference and improves feature purity during fusion. Third, a decoupled difference reconstruction (DDR) unit is developed to explicitly reconstruct residual differences between adjacent feature hierarchies. This compensation maximizes the preservation of fine-grained textures for early stage fires. Extensive experiments demonstrate the superiority of our approach. The proposed model achieves an mAP@50 of 75.2%, an mAP@50–95 of 52.1% on forest fire datasets with only 2.531 M parameters, which comprehensively outperforms state-of-the-art methods in both detection accuracy and deployment parameters. Moreover, cross-dataset evaluation further demonstrates the generalization capability of AAFNet under diverse environmental distributions.