A Knowledge-Anchored Dynamic Fusion Framework with Feature-Wise Linear Modulation for Multimodal Forest Fire Perception

Early forest-fire perception requires fusing heterogeneous observation streams whose relative reliability varies sharply with environmental conditions. Existing multimodal fusion schemes are mostly static: the learned fusion operator is frozen at inference time, and context dependence, if any, is restricted to a few structured metadata variables. We instead cast multimodal fusion as learning a parametric family of fusion operators conditioned on a semantic state, and propose KnowFuse-Fire, in which a frozen vision–language model (VLM) infers the semantic state while a prompt-conditioned adapter parameterizes the fusion operator of a cross-modal Transformer. Conditioning enters through two channels: a softmax gating map assigning modality weights on the 2-simplex, and a feature-wise linear modulation (FiLM) applying a state-dependent affine transformation at every layer. Bounding the modulation gains keeps the network Lipschitz-stable with an explicit constant, and the entropy of the gating weights measures how decisive a given fusion decision is. On a purpose-built dataset of visible-light, thermal-infrared, and meteorological sequences, the framework attains an overall F1-score of 0.881† (95% confidence interval 0.874–0.887) and exceeds the strongest baseline by 4.0 percentage points overall and by 9.3 percentage points on the challenging smoldering subset, a subset-level margin that does not survive correction for multiple comparisons (Section 4.3); the semantic channel alone accounts for a 2.5-point gain. On this dataset, conditioning on an inferred semantic state helps precisely where static operators fail, and the gain arrives together with a stability guarantee rather than at its expense.