Abstract
In recent years, Unmanned Aerial Vehicle (UAV)-based object detection technology has demonstrated immense potential for forest fire monitoring in complex environments. However, constrained by the drastic multi-scale variations in fire targets, severe background interference, and the limited computational resources of edge devices, existing object detection models struggle to strike a balance between detection accuracy and inference efficiency. To systematically address the aforementioned issues, this paper proposes MVLFireNet, a lightweight and real-time forest fire detection model driven by multi-granularity vision-language enhancement. First, a Multi-scale Spatial-Aware attention (MSA) module is proposed to capture global long-range dependencies while explicitly preserving the high-frequency two-dimensional spatial features of weak fire spots and smoke edges. Second, a Cross-Modulation Fusion (CMF) module is designed to replace the traditional passive feature concatenation with bidirectional nonlinear conditional modulation, thereby achieving active denoising and compensation for both deep high-level semantics and shallow spatial details. Finally, a Multi-granularity Vision-Language Enhancement (MVLE) branch is innovatively introduced to inject robust semantic discriminative capabilities into visual features via a hierarchical text alignment enhancement mechanism covering both global scenes and local targets. Furthermore, FSDataset-VL, the first large-scale multi-granularity text-image dataset tailored for forest fire detection, is constructed. Extensive experiments on FSDataset-VL demonstrate that MVLFireNet, with merely 2.41 million parameters, achieves mAP@0.5 and mAP@0.5:0.95 of 86.3% and 54.8%, respectively, both representing the highest performance among all comparative models. Under the mAP evaluation framework, it attains an optimal balance between detection accuracy and computational efficiency, thereby providing an efficient solution for UAV-based forest fire monitoring in complex environments.