SVM for hyperspectral images classification based on 3D spectral signature

Hyperspectral imaging sensors acquire images in hundreds of continuous narrow spectral bands spanning the visible to infrared spectrum which led to obtain hyperspectral image with high spectral resolution. Thus each object presented in the image can be identified from their spectral response. The classification of multi-temporal hyperspectral image is a challenge task due to the problem of spectral variation over the time. In fact, many factors can affect the spectral signature of object like weather and climatic effects, so two images taken on the same area but at different times and under different conditions can lead to different spectral signatures for the same objects. This observation has fostered the idea of adopting 3D representation of spectral signature to classify multi-temporal hyperspectral image. The main objective of this representation is to have for each object a compact model which illustrate their spectral variation over the time, it represent the variation of reflectance as a function of time and spectral waveband. In this paper, we propose a new approach for multi-temporal hyperspectral image classification based on 3D spectral signature to solve the problem of spectral variation. This approach consist, foremost, to represent each pixel of classified image by a 3D spectral signature after the application of the powerful 3D modeling method "Non-uniform Rational Basis Spline" (NURBS), after, to apply local shape descriptor "spherical harmonic decomposition" to extract spectral features from each 3D spectral signature and, finally, to classify the image by means of supervised classifier "Support Vector Machines (SVMs)" with Radial Basis Function (RBF) kernel. To evaluate this approach, we used a series of multi-temporal hyperspectral "hyperion" images.