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

Abstract:

Hyperspectral image classification plays a vital role in remote sensing by enabling precise identification of land cover and material composition across hundreds of spectral bands. However, the high dimensionality of hyperspectral data poses significant computational challenges, particularly on limited-resource systems. This paper presents the DDHC (DNN-DCT Hyperspectral Classification) framework, an efficient classification approach that integrates Discrete Cosine Transform (DCT) based spectral compression with a lightweight Deep Neural Network (DNN) for accurate pixel-wise land cover mapping. The proposed method employs Uniform Band Selection (UBS) to reduce spectral redundancy, followed by subgroup-DCT to extract compact and discriminative spectral features. A multi-scale convolutional neural network with parallel branches of varying kernel sizes captures both local and global spatial patterns from 7×7 spatial patches. Residual connections and batch normalization stabilize training while dropout regularization prevents overfitting. The framework is evaluated on two benchmark datasets, Indian Pines and Pavia University, achieving competitive Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficient scores. Hardware-aware optimizations including GPU acceleration, parallel patch extraction, and mixed precision support ensure practical deployability. Experimental results confirm that DDHC delivers high classification performance while maintaining computational efficiency suitable for resource-constrained environments.

Keywords: Hyperspectral image classification, Discrete Cosine Transform, deep neural network, remote sensing, limited-resource systems