Hyperspectral Image Classification using Trilateral Filter and Deep Learning

The high spectral variability and resolution make hyperspectral image classification difficult. Most of the previous researches focused on solving its high dimensionality problem; very few types of researches have focused on using edge information of the spectral bands to carry out classification. This paper aims to compare two edge filtering methods-bilateral filter and trilateral filter to extract edge information of hyperspectral images and then use those features for further classification. Minimum Noise Fraction (MNF) and Principal Component Analysis (PCA) were separately used for dimensionality reduction. The images obtained from both the methods were stacked, and the two edge filters were applied to them individually. The final feature images were classified using Convolutional Neural Network (CNN). The proposed methodology was applied to the Indian Pines dataset. The results obtained are satisfactory and have been compared to state of the art (SOTA) techniques.

I. Introduction

Hyperspectral imaging (HSI) deals with the imagery of narrow spectral bands over a continuous spectral range. Over the past decade, HSI has been a rapidly developing field. Due to hyperspectral images discriminating ability, it has found applications in a variety of areas, such as agriculture, eye care, food processing, surveillance, chemical imaging, environment management, and land cover mapping [26]. Unlike color images with 3 bands, generally, hyperspectral images have more than 100 bands that capture both the spectral and spatial information of different objects in the image. A pixel in a hyperspectral image is a high dimensional vector where the spectral reflectance of the captured image at a particular wavelength is stored. Since HSI can detect subtle spectral differences, the classification of hyperspectral images is an important task, where the aim is to classify each pixel vector in the image into one of the various classes present in the image; as a result, pixel classification in hyperspectral images has attracted a lot of interest, [10], [14].