Abstract: Eliminating impulse noise of high-quality remote sensing images is of great significance for applied research. It has always been a challenge to eliminate high-density impulse noise while remaining detailed information on edges in original remote sensing images. This study concluded that uncertain changes will appear when a remote sensing image is corrupted by impulse noise. Given this, an uncertainty model based on the evidence theory was constructed using multiple features of impulse noise. The BJS divergence and the reliability entropy were fused into the model to obtain new weights and a new probability assignment. Then, the classification between noise and signals was given according to fusion rules and probability transformation, thus effectively reducing the possibility of high-level conflicts. The experimental results show that the classification method proposed in this study is effective even when the noise density is up to over 90% and can well maintain detailed information on different ground objects in the denoised remote sensing images.