An Unsupervised Statistical Segmentation Algorithm for Fire and Smoke Regions Extraction

Estimation of the extent and spread of wildland fires is an important application of high spatial resolution multispectral images. This work addresses an unsupervised statistical segmentation algorithm to map fire extent, fire front location, just burned area and smoke region based on a statistical model. The results are useful information for a fire propagation model to predict fire behavior. The finite mixture (FM) model is a widely used model for image segmentation because of it is mathematically simple and tractable. However, it ignores the spatial constraint of images, and works only on well defined images with low level noise. This is an intrinsic limitation of histogrambased segmentation algorithm, such as K-means and EM algorithm. In this paper we propose model the hidden segmentation field as an Markov random field (MRF). The hidden segmentation field can not be observed directly but can be estimated through the observed vector-valued pixels of satellite/airborne multispectral images. The advantage of the MRF model is that it encodes spatial information by considering the mutual influence of neighboring sites. Based on the MRF property of the segmentation field, we propose model the posteriori marginal probability field on the image sites as a multivariate Gaussian Markov random field (MGMRF). And then implement a Maximize Marginal Probability method (MPM) to segment the images. Our algorithm is a generalization of the Expectation Maximization (EM) algorithm to incorporate spatial constraints in the image. The use of statistical method has the added advantage of providing a direct means of deriving a probability value that is required for new approaches to fire propagation modeling. Experimental results obtained by applying this technique to two AVIRIS real images show that the proposed methodology is robust with regard to noise and variation in fire as well as background. The segmentation results of our algorithm are compared with the results of K-means algorithm and EM algorithm. It is shown that the results of our algorithm are consistently better than those of classical histogram based methods.