Efficient multi-temporal hyperspectral signatures classification using a Gaussian-Bernoulli RBM based approach

Abstract

This paper presents an efficient Gaussian-Bernoulli Restricted Boltzmann Machines (GB-RBM) framework in order to better address the classification challenge of remotely sensed images. The proposed approach relies on generating well-designed features for a new 3D modality of spectral signature. For this purpose, mesh smoothing is introduced to reduce noise while conserving the main geometric features of the multi-temporal spectral signature. Then, we propose the use of an RBM (Restricted Boltzmann Machine) framework as stand-alone non-linear classifier. The adapted framework focuses on a cooperative integrated generative-discriminative objective allowing the integration of modeling input features and their classification process in one-pass algorithm. The main benefit of the proposed approach is the ability to learn more discriminative features. We evaluated our approach within different scenarios and we demonstrated its usefulness for noisy high dimensional hyperspectral images.