Wildfire may cause serious damage to the forest. Because of the complexity of the forest environment video images obtained by CCD cameras of the automatic forest fire surveillance system often contain much noise. It will be one of the most troublesome things for the follow-up image processing procedures. After features of noise in the forest fire images are analyzed, a kind of forest fire image de-noising method based on the wavelet transform theory is introduced and its feasibility to remove noise in forest fire images is discussed in detail. Then several forest fire image de-noising experiments with various threshold decision strategies under the MATALB platform are carried out. At last these experiment results are compared according to SNR and image loss degree and it is showed that the wavelet de-noising method with the Bayes threshold estimation algorithm is one of the most efficient de-noising techniques for the image preprocessing procedure of a forest wildfire.