Efficient Hyperspectral Image Classification using Discrete Cosine Transform on Limited-Resource Systems

Deep learning based approaches to hyperspectral image analysis have attracted large attention and exhibited high performance in image classification tasks. However, deployment of deep-learning-based hyperspectral image analysis systems is challenging due to the computational complexity of deep learning and the large amount of data involved in hyperspectral images. To address this problem, this paper introduces a novel framework that integrates deep neural network (DNN) based image analysis by learning the network from a spectral frequency representation in discrete cosine transform (DCT) coefficients for hyperspectral image classification. The framework allows designers to derive diverse implementation configurations using a variable number of DCT coefficients for training. These configurations can be used to flexibly trade off classification accuracy and computational cost (e.g., based on specific characteristics of the device being used or based on time-varying operating requirements). In this work, we demonstrate the feasibility of our proposed technique for deploying DNN-based hyperspectral image classification through extensive experiments by using public remote sensing datasets and a resource-constrained Android platform. This work enables designers to make effective choices between processing speed and image classification accuracy based on operational constraints and requirements.