Abstract
Abstract. High spectral dimensionality of hyperspectral images makes them useful data resources for earth observation in many remote sensing applications. In this case, the convolutional neural network (CNN) can help to extract deep and robust features from hyperspectral images. The main goal of this paper is to use deep learning concept to extract deep features from hyperspectral datasets to achieve better classification results. In this study, after pre-processing step, data is fed to a CNN in order to extract deep features. Extracted features are then imported in a multi-layer perceptron (MLP) network as our selected classifier. Obtained classification accuracies, based on training sample size, vary from 94.3 to 97.17% and 92.35 to 98.14% for Salinas and Pavia datasets, respectively. These results expressed more than 10% improvements compared to the classic MLP classification technique.
Highlights
Hyperspectral images mostly contain hundreds of spectral bands that can make a continuous spectral signature for every observed pixel (Chang, 2003)
Chen et al, (2014) employed stacked autoencoder containing five layers as deep architecture to a hyperspectral image transformed by Principal Component Analysis (PCA) in order to extract deep features to be fed in Support Vector Machine (SVM) classifier
Convolutional neural networks were exploited in order to extract deep features from a hyperspectral image
Summary
Hyperspectral images mostly contain hundreds of spectral bands that can make a continuous spectral signature for every observed pixel (Chang, 2003). This property makes them useful for earth observation, and remote sensing tasks such as image classification, anomaly detection, and target detection. High dimensional data needs a high amount of labeled samples to obtain reliable results This problem, known as the Hughes phenomenon, may cause redundancies and disturbances (Hughes, 1968; Yu et al, 2017). Yue et al, (2015) proposed a CNN-based spectral-spatial classifier for hyperspectral data classification after transforming the data using the PCA in order to cancel data redundancy and dimension reduction. In (Hu et al, 2015), the authors trained a feed-forward neural network with five layers including an input layer, one-dimensional (1-D) convolutional layer, a max-pooling layer, a fully connected layer, and an output layer
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
