Automated Hyperspectral Image Classification Using Spatial-Spectral Features

Hyperspectral imaging has transpired as a compelling tool in various fields like geology, mining, agriculture, etc with applications ranging from object detection to quality inspection. Feature extraction, as well as the methodology used for feature extraction, plays an indispensable role in increasing the accuracy of the classification of hyperspectral imaging (HSI). This paper proposes an algorithm for automated hyperspectral image classification using nine spatial-spectral features, which includes linear predictive coefficients, wavelet coefficients, standard deviation, average energy, mean, fractal dimension, entropy, Rényi entropy and Kraskov entropy. These features are further used for classification using the quadratic support vector machine (SVM). The elaborated scheme exercises 10-fold cross-validation. The collective effect of the excerpted features is determined and the accuracy trends for the various number of features is ascertained. Appreciable overall accuracies (OA) for all the three publicly available data sets are acquired as follows: Salinas-A data set $(\mathbf{OA} = 99.60\%)$, Salinas data set $(\mathbf{OA}=92.4\%)$ and Botswana data set $(\mathbf{OA} =89.5\%)$.