During the last ten years computer vision techniques have shown a great potential in solving the problem of automatic fire detection. Vision-based fire detection offers many advantages over the conventional methods that use smoke and heat detectors. This paper presents a novel approach for fire detection by modeling the structure of spatial of fire, this structure is considered in terms of the color intensity of fire pixels. Furthers the type-2 fuzzy clustering technique is applied to separate fire-color pixels into some clusters, then these clusters are used to model structure of fire. Experimental results show that our method is capable of detecting fire in early state of fire and weak light-intensity environment; and this method uses only information on a single image so it can be integrated into the surveillance system that used dynamic camera.