To accurately identify and segment wildfires, we developed a wildfire segmentation method based on an improved lightweight Yolov9 network with multi-scale fusion. In previous research, wildfires were segmented based on their textures, shapes, and colors, while we used a deep learning method. We trained the model to learn the multi-scale features of wildfires for the pixel-level segmentation of wildfires. Firstly, we replaced the ordinary convolutional module in RepNCSPELAN4 with the lightweight Ghost module in the backbone network based on the yolov9 network structure and added the CBAM module in the neck network to adaptively adjust the channel and spatial distribution of the feature map. This method improved the characterization ability of the model features and reduced the computational complexity of the model at the same time. The developed wildfire segmentation method based on the improved lightweight yolov9 network with multi-scale fusion improved box precision by 0.9% and mask precision by 0.5% compared to the baseline model, while reducing the number of parameters by 357,308. The method is an effective lightweight method for accurately segmenting wildfire regions.