Bayesian classifier applications of airborne hyperspectral imagery processing for forested areas

Pattern recognition problem is outlined in the context of textural and spectral analysis of remote sensing imagery processing. Main attention is paid to Bayesian classifier that can be used to realize the processing procedures based on parallel machine-learning algorithms and high-productive computers. We consider the maximum of the posterior probability principle and the formalism of Markov random fields for the neighborhood description of the pixels for the related classes of objects with the emphasis on forests of different species and ages. The energy category of the selected classes serves to account for the likelihood measure between the registered radiances and the theoretical distribution functions approximating remotely sensed data. Optimization procedures are undertaken to solve the pattern recognition problem of the texture description for the forest classes together with finding thin nuances of their spectral distribution in the feature space. As a result, possible redundancy of the channels for imaging spectrometer due to their correlations is removed. Difficulties are revealed due to different sampling data while separating pixels, which characterize the sunlit tops, shaded space and intermediate cases of the Sun illumination conditions on the hyperspectral images. Such separation of pixels for the forest classes is maintained to enhance the recognition accuracy, but learning ensembles of data need to be agreed for these categories of pixels. We present some results of the Bayesian classifier applicability for recognizing airborne hyperspectral images using the relevant improvements in separating such pixels for the forest classes on a test area of the 4 × 10 km size encompassed by 13 airborne tracks, each forming the images by 500 pixels across the track and from 10,000 to 14,000 pixels along the track. The spatial resolution of each image is near to 1 m from the altitude near to 2 km above the ground level. The results of the hyperspectral imagery processing have shown that for the ensembles of data corresponding to the prevailing pine species and sunlit tree’s tops the young forests (13–26 years old) and mature forests (106–136 years old) on the test area are recognized with high accuracy, but difficulties are essential for intermediate ages (36–96 years old) though some of these ages (for example, 47 and 76 years old) are also recognized with high accuracy. This is due to insufficient sampling of the pixels relating to the sunlit tops for the intermediate ages using the classifier employed. A conclusion is made that a fusion of the passive imaging spectrometer and the active lidar (laser scanner) on the same gyro-stabilized airborne platform is needed to overcome these difficulties.