Classification of Different Vegetation Types Combining Two Information Sources Through a Probabilistic Segmentation Approach

In this work we propose a new probabilistic segmentation model that allows us to combine more than one likelihood. The algorithm is applied to identify vegetation types in images from Landsat 5 satellite. Firstly, we obtain histograms from two information sources: spectral bands and principal components obtained from vegetation indices. Then, given an image, we compute two likelihoods of pixels to belong to each class (vegetation type), one for each source of information. The computed likelihoods are the inputs of the proposed probabilistic segmentation algorithm. This algorithm gives an estimation of the probability of a pixel of belonging to a class. The final segmentation is easily obtained by maximizing the estimated discrete probability for each pixel of the image. Experiments with real data show that the proposed algorithm obtains competitive results compared with state of the art algorithms.