Early detection of forest fires is critical for minimizing ecological and economic losses. However, most existing approaches rely on direct feature concatenation or general adaptive fusion mechanisms, which are often insufficient to capture discrete, rapid changes in fire dynamics or filter out environmental sensor noise. To address these algorithmic limitations, we propose MAFF-Net, a multimodal adaptive feature fusion network. It specifically integrates a key timestep identification module to capture temporal saliency, a selective spatial feature extractor to eliminate visual redundancy, and a multimodal feature fusion module to unify heterogeneous representations. This study also constructed a time-synchronized multimodal dataset combining six types of environmental sensor data with image sequences to validate the approach. The experimental results show that MAFF-Net achieves a 3.79%-13.72% higher F1-score than baseline unimodal models and outperforms other adaptive fusion strategies. In wild environments, it maintains an average F1-score of 95.79%, confirming its superior fusion efficiency and discriminative capability. These findings highlight MAFF-Net’s strong generalization and make it a promising framework for intelligent, real-time forest fire early-warning systems. The source code and the constructed dataset are publicly available at https://github.com/Universe-ustc/MAFF-Net .
Liang et al. (Wed,) studied this question.