Publication

Abstract : Hyperspectral imaging (HSI) produces high-dimensional data characterized by detailed spectral and spatial information for each pixel, which is instrumental in applications such as object detection and material identification. However, the substantial volume of HSI data renders their processing computationally demanding. Data compression techniques are imperative for efficient storage and analysis to mitigate these challenges. Although traditional compressive sensing methods are effective for low-dimensional data, they often fail to preserve the essential spectral features of HSIs. To overcome these limitations, we propose a structured HSI compression framework that integrates a Branch-and-Bound–based measurement selection strategy, enabling globally optimized and dimension-aware sampling. For reconstruction, we introduce a Deep Tucker Decomposition model enhanced with Spatial–Spectral Collaborative Learning (SSCL), which effectively captures multi-mode correlations and restores fine-grained spectral features from compressed data. The approach was evaluated using benchmark HSI datasets, including Indian Pines and Pavia University datasets. The quality of the reconstruction and the effectiveness of the compression were quantitatively assessed using PSNR, SSIM, and SAM. The proposed Deep Tucker SSCL model not only achieves superior reconstruction quality with average PSNR values above 35 dB and SSIM scores around 0.97, but also demonstrates improved computational efficiency, requiring only 217 MFLOPs and 0.279 sec for Indian Pines and 264 MFLOPs and 0.315 sec for PaviaU. This balance of high accuracy and reduced complexity highlights the model’s robustness and suitability for real-time hyperspectral image reconstruction.