INTERBAND PREDICTION OF HYPERSPECTRAL IMAGES USING GENERALIZED REGRESSION NEURAL NETWORK Kalyan Kumar Halder Manoranjan Paul School of Computing and Mathematics Charles Sturt University Bathurst, NSW 2795, Australia E-mail: {khalder, mpaul}@csu.edu.au ABSTRACT Predicting upcoming bands of hyperspectral images is an im- portant task in modern image compression algorithms. This paper proposes a new algorithm to predict the band-wise cor- relation of hyperspectral images based on a generalized re- gression neural network (GRNN). The proposed algorithm uses the intensity values of the previous bands to train the GRNN and approximates the correlation between them. The next band is then predicted using the trained network and the immediately previous band. This algorithm works on a pixel-by-pixel basis and does not involve any mathematical modeling or any previous knowledge of the images. The per- formance of the proposed algorithm is evaluated by applying it to several Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) reflectance datasets. Simulation results show that the proposed algorithm provides substantial accuracy in the prediction of upcoming bands. Index Terms— Generalized regression neural network, hyperspectral image, image coding, remote sensing, inter- band prediction 1. INTRODUCTION Hyperspectral imaging system acquires and processes a high definition image of a scene where every pixel in the image contains a continuous spectrum. While ordinary color cam- eras mostly acquire three bands (red, green and blue), the hy- perspectral imaging cameras, using the power of digital imag- ing and spectroscopy, provide more precision and detail of the scene by dividing the spectrum into many more bands even beyond the visible range. Figure 1 illustrates a sample hyper- spectral data cube containing approximately 428 bands that is collected by the NEON imaging spectrometer (NIS) within the 380 nm to 2510 nm portions of the electromagnetic spec- trum within bands that are approximately 5 nm in width [1]. Hyperspectral images are used to detect, analyze and char- acterize the objects based on their spectral properties. These images have wide applications in biotechnology, food indus- Fig. 1. A sample hyperspectral data cube [1]. try, environmental monitoring, pharmaceutical industry, me- teorology, and surveillance [2–4]. Predicting the next bands of a hyperspectral image could be advantageous for its application to lossless predictive im- age compression. The predictive coding is quite simple for encoding and decoding as it only act on residuals, thus re- duces the memory requirement and computational complex- ity of the algorithm significantly. Several research has been conducted so far to address the problem of a-priori estima- tion of the next bands of hyperspectral images [5–11]. Mielikainen et al. [5] proposed a linear function for the interband prediction of hyperspectral images. To predict the next band, this algorithm uses a casual dataset of its previous bands, multiplies them with the same number of prediction coefficients, and adds them. The accuracy of this algorithm is highly depended on the coefficients’ values, thus it can pro- vide limited efficiency. Miguel et al. [6] assumed a linear correlation between the bands of a hyperspectral image and proposed an algorithm that uses the previous bands and two prediction coefficients to predict the next band. Once the pre- diction of a band is accomplished, its residual, with respect to the original band, is compressed using a standard compres- sion algorithm. Rizzo et al. [7] proposed another linear pre- dictor for lossless compression for hyperspectral images. This predictor uses the immediate previous value and a small ca- sual data subset of a given pixel to estimate its next possible 978-1-5090-6067-2/17/$31.00 c 2017 IEEE