Destructive wildfires pose a serious threat to ecosystems, economic development, and human life and property safety. If wildfires can be extinguished in a relatively short period of time after they occur, the losses caused by wildfires will be greatly reduced. Although deep learning methods have been shown to have powerful feature extraction capabilities, many current models still have poor generalization performance when faced with complex tasks. To this end, in this study, we considered introducing attention modules both inside and outside the nested U-shaped structure and trained a neural network model based on the U2-Net architecture to enable the model to suppress the activation of irrelevant areas. Compared with baseline models such as U-Net, our model has made great progress on the test set, with an F1 score improvement of at least 2.8%. The experimental results indicate that the model we proposed has certain practicality and can provide a significant scientific basis for forest fire management and emergency decision-making.