Hyperspectral Image Classification using Combined Spectral-Spatial Denoising and Deep Learning Techniques

In this paper we propose a hyperspectral image classification chain integrating also a denoising step. Hyperspectral image denoising is extremely important for any type of application due to the fact that the level and type of noise often vary between different bands and with certain spatial areas. Hyperspectral classification, capable of labeling all the pixels in the scene into one or several classes according to their characteristics, has become one of the most popular topic in the domain of high resolution images. In the past decade a large number of classification algorithms have been proposed, but the most important ones are represented by the artificial intelligence methods, especially by convolutional neural networks that provide very good results in image processing areas. The denoising strategy considered in this paper is based on the adaptive total variation method capable of alleviating the difficulty of removing the noise while preserving textures and edges. This approach, lies in the adaptive regularization terms in both the spatial and spectral dimensions. Experiments using the publicly available hyperspectral images Indian Pines, Salinas and Pavia University were conducted in order to validate the proposed classification of hyperspectral data by the processing chain.