Unsupervised Spectral Feature Extraction for Hyperspectral Image Classification

Abstract

AbstractThe modern criterion of remote sensing is the acquisition of enormous dimensions of hyperspectral images. The transformation of such enormous multi-dimensional images, object unmasking, feature drawing, forecasting, and hyperspectral image classification is several progressive affairs in the present scenario. Countless mechanisms have been refined to classify hyperspectral images. Hyperspectral image (HSI) classification has been a very dynamic area in remote sensing and other applications like agriculture, eye care, food processing, mineralogy, surveillance, etc. HSI gathers and summons information cross the electromagnetic spectrum. The objective of HSI is to earn the spectrum for all pixels in the image for the reason of discovering things, analyzing materials, and recognizing processes. It includes varying bands of images. HSI often dispenses with an inseparably nonlinear connection linking the recorded spectral data and the similar materials. Recently, deep learning antiquated as a robust feature extraction tool to effectively disclose irregular problems and extensively utilized in several image processing tasks. In the beginning, fast independent component analysis (FICA) is applied for dimensionality reduction. Convolutional neural network (CNN) is the persistently used deep learning-based technique for observable data processing. Initially, fast independent component analysis is executed to reduce dimensions and then CNN is implemented to the reduced data. The results are tested using different spectral features with CNN for HSI. The overall accuracy of the suggested approach is 99.83 and 100% for Indian Pines, Salinas Scene which shows the reliability of the suggested CNN method for HSI classification.KeywordsClassificationHyperspectralConvolutional neural networksFast-independent component analysis