|
In this paper, we present a novel fusion and classification process for remotely sensed images developed in the framework of possibility theory. Unlike probability theory, possibility theory has the ability to handle both uncertainty and imprecision of classified pixel through a possibility and a necessity measures. Proposed multisource fusion and classification process involves several steps: First of all, the probability distribution of each spectral class from the samples representing thematic classes is estimated. This estimation is based on the histogram analysis of each class. Then, the possibility distribution is estimated from probability distributions using the Klir probability-possibility transformation. Next, once the possibility of samples is determined, we apply the Lagrangian interpolation method to estimate the other observations (grey level/reflectance) of the information sources. Combination operators: conjunctive and disjunctive are then used to combine possibilities of each source. This operation is performed a pixel level. Finally, decision is taken by using a criterion based on maximization of the possibility measures in order to select the optimal class. Experimental results obtained as land cover maps and land change maps indicate by using a statistical assessment (confusion matrix, Kappa coefficient, local spectral analysis method) that the proposed possibilistic fusion and classification process outperform the existing probabilistic fusion and classification process: It is important to emphasize that classical probabilistic data fusion models do not provide an adequate combination tool to detect and label change areas. Thus, possibility theory seems to be the best methodological framework which allows the development of multisource and multitemporal image fusion and classification process.
|
