Adverse weather conditions significantly degrade the visual quality of images and affect system performance. While existing research has achieved satisfactory results in single image restoration tasks under single adverse weather conditions, challenges remain in developing unified image restoration techniques for multiple adverse weather conditions, including low model efficiency, task conflicts, and insufficient generalization capabilities. To address these issues, this paper proposes an optimized composite weather image restoration technique based on two-stage training strategy. In the first stage, cross-weather universal features are extracted, and in the second stage, composite weather-specific parameters are adaptively expanded. By combining a dynamic weather combination identifier based on a marked vector, the technique achieves collaborative suppression of composite degradation. Experiments show that this method achieves a significant improvement in PSNR compared to traditional methods on a self-built composite weather dataset, especially in rain-haze composite scenes, where it significantly improves detail retention capabilities.