Aiming at the problems of missing and false detection encountered in image recognition on the forest fire, a sample entropy discrimination algorithm based on K-Means clustering was proposed. First, the collected forest fire images were transformed into color gamut space, which reduced the influence of visual deviation in the process of image recognition. Then, the K-Means clustering algorithm was adapted to cluster the image subset that was expected by fire through the Euclidean distance criterion of HSV components. On this basis, the weight of the clustered image subset was identified by using the sample entropy, the entropy values of the correlative fire regions and the real fire regions were statistically distinguished. Then it was confirmed whether there was a fire in the subset of images screened by the cluster. The experimental results showed that by using the sample entropy fusion K-Means clustering algorithm, the recognition accuracy can be effectively improved in forest fire image recognition. After the detection of 60 images, the correct identification rate of fire area in all images was improved to 96.67%, and the average identification time was 16.03 s. Due to the strong robustness and convenience of the algorithm, it is able to adapt to the identification of the fire area under complex background and has better detection effects than the traditional K-Means algorithm.