Models for Exploring the Benefits of using Discrete Wavelet Transformation in HSI

The application of Hyper Spectral Images is increasing day-by-day, with the development of remote sensing technology. With the rapid development of Deep Learning technology, many research work focus on Classification of Hyper Spectral Images based on both spectral and spatial features extraction. It is the task of correctly predicting the class values of different pixel values present in remotely sensed HSI data. In order to achieve the accurate classification of ground features, Feature Extraction is one of the crucial step which increases the accuracy of learned models by extracting relevant features from the input data. As the HSI image consists of hundreds of continuous spectral bands, we need an effective way to extract the spectral features of the HSI images (other than CNN techniques). In this paper, we exploit two different types of Discrete wavelet transformation techniques like Haar, Daubechies (Db4) for spectral feature extraction. This in turn reduces the dimensionality of data. These spectral features are then linked to four layers of 2D CNN to extract the spatial features. The extracted features from the wavelet fusion CNN are provided further for classification. Initially factor Analysis method was used to reduce the dimensions of the HSI input data. Our experimental results conclude the better accuracy method among these, through a comparative analysis with other state-of-the-art methods. We use Overall Accuracy, Kappa Coefficient and Average Accuracy as a Performance measures on 3 benchmark datasets of Indian Pines, Salina Scene, University of Pavia.