<p>Wildfires threaten global ecosystems, resulting in biodiversity decline, soil deterioration, and climate alteration. Effective fire detection algorithms are crucial for prompt recognition and emergency response. Notwithstanding progress, existing techniques continue to falter in detecting minuscule fire sources within intricate settings. This paper presents the YOLOv8-WMD algorithm for forest fire detection, employing frequency-domain multi-scale feature fusion to tackle these challenges. We integrate a WDown frequency-domain downsampling module into the backbone to augment low-frequency information for enhancing detection of diminutive and low-contrast fire sources. We developed the MCSA multiscale channel spatial attention module to improve multiscale identification capabilities by utilising low-frequency information processed in the frequency domain. This facilitates the detection of fires of varying scales.The DySample upsampling module, implemented in the neck network, can dynamically adjust its position according on feature variation to improve information reconstruction efficacy. Experimental findings indicate that our proposed solution surpasses the conventional YOLOv8n algorithm on a custom-built STF dataset. Precision has increased by 1.5%, recall by 4.2%, mAP50 by 2.3%, and mAP95 by 2.9%. This underscores the algorithm’s efficacy in identifying small targets within a multifaceted forest fire setting.</p>