DGFRNet: Dual-guided feature refinement network for forest fire detection near transmission lines

Forest fire detection is crucial for protecting high-voltage transmission lines, yet the coexistence of dense smoke and small flames presents significant challenges for accurate localization and efficient inference. To address this, we propose a framework named dual-guided feature refinement network with progressive granularity (DGFRNet). Specifically, we design a frequency-responsive dilated convolution (FR-DConv) that dynamically adjusts dilation rates based on local frequency energy, enabling the network to balance global context aggregation for smoke with fine detail preservation for flames. In parallel, a bipolar entropy discrimination module (BEDM) selectively amplifies informative channels by estimating local entropy, avoiding the computational overhead of conventional attention mechanisms. Following dual-dimensional information interaction, a spectral reconstruction guided representation module (SRGRM) explicitly decouples semantic intensity and positional structure via amplitude-phase decomposition, enhancing the localization of ambiguous fire regions. Furthermore, we introduce a progressive refinement path combining a hierarchical multi-granularity fusion block (HMGFB) and a bidirectional alignment attention (BAA) module, facilitating feature alignment between semantic abstraction and spatial granularity. Extensive experiments on four datasets show DGFRNet’s high accuracy, compact size, and fast inference. A case study on a 220 kV transmission line demonstrates that combining DGFRNet with fire behavior and breakdown models enables early trip risk assessment, highlighting its engineering applicability.