The paper presents FFYOLO, a lightweight forest fire detection model based on YOLOv8. To address the limitations of traditional machine learning methods and deep learning models in complex forest fire detection scenarios, the authors propose several novel components and improvements: 1. **Channel Prior Dilatation Attention (CPDA)**: This attention module enhances feature extraction by using multi-scale asymmetric dilated convolutions to focus on target features, improving detection accuracy in complex scenes. 2. **Mixed-Classification Detection Head (MCDH)**: Designed to balance accuracy and speed, MCDH combines two convolutional modules for class prediction, reducing parameter redundancy and improving efficiency. 3. **Lightweight GSCov**: Replacing standard convolution with GSCov in the Neck section reduces parameters and computational complexity while maintaining model accuracy. 4. **Knowledge Distillation**: Applied during training to enhance the model's generalization ability and reduce false detections. Experimental results demonstrate that FFYOLO achieves an mAP0.5 of 88.8% on a custom forest fire dataset, outperforming the original YOLOv8 model by 3.4% with 25.3% fewer parameters and 9.3% higher frames per second (FPS). The model's effectiveness is validated through various ablation studies and comparisons with other models, showing superior performance in complex scenarios and real-world applications.The paper presents FFYOLO, a lightweight forest fire detection model based on YOLOv8. To address the limitations of traditional machine learning methods and deep learning models in complex forest fire detection scenarios, the authors propose several novel components and improvements: 1. **Channel Prior Dilatation Attention (CPDA)**: This attention module enhances feature extraction by using multi-scale asymmetric dilated convolutions to focus on target features, improving detection accuracy in complex scenes. 2. **Mixed-Classification Detection Head (MCDH)**: Designed to balance accuracy and speed, MCDH combines two convolutional modules for class prediction, reducing parameter redundancy and improving efficiency. 3. **Lightweight GSCov**: Replacing standard convolution with GSCov in the Neck section reduces parameters and computational complexity while maintaining model accuracy. 4. **Knowledge Distillation**: Applied during training to enhance the model's generalization ability and reduce false detections. Experimental results demonstrate that FFYOLO achieves an mAP0.5 of 88.8% on a custom forest fire dataset, outperforming the original YOLOv8 model by 3.4% with 25.3% fewer parameters and 9.3% higher frames per second (FPS). The model's effectiveness is validated through various ablation studies and comparisons with other models, showing superior performance in complex scenarios and real-world applications.