Abstract: To effectively control forest fires, it is crucial to predict the spread of the fire after obtaining information on the ground vegetation, topography, and terrain of the affected area, and to develop firefighting strategies accordingly. However, remote sensing images of the ground are often obscured by uneven cloud and fog due to weather conditions, which impairs the observation of ground vegetation information and subsequently affects predictions of fire spread. Dehazeformer, a deep learning-based dehazing method, has demonstrated some effectiveness in haze removal. However, the algorithm struggles to meet real-time performance requirements in time-sensitive tasks. To address the limitations of large model parameters and inadequate handling of haze details, this study introduces improvements to the model, focusing on lightweighting and enhancing dehazing performance. The improved model's test results indicate certain improvements in PSNR and SSIM metrics. In practical fire monitoring scenarios, the model significantly enhances the discernibility of ground information surrounding fire-affected areas through dehazing, thereby aiding in the prediction of fire spread trends.

Key words: image dehazing, remote sensing, deep learning, fire spread, forest fire