Attention Subsumed Residual Dense Network for Reconstruction of Block Compressed Hyperspectral Images
Elza George, Aravinth J, Sathishkumar Samiappan · IEEE Access · 2025
Hyperspectral imagery is a rich source of spectral and spatial information and is widely used in domains such as remote sensing, environmental monitoring, and mineral exploration. However, processing and transmitting such high-dimensional data poses significant challenges due to the large volume and computational load. These issues can be alleviated by reducing the dimensionality of data through band selection and its size through compression. Most existing reconstruction methods following compression, exhibit suboptimal evaluation metric values, incur high processing times, and result in poor visual quality. To overcome these limitations, we propose a three-stage methodology. First, redundant spectral bands are eliminated using a Modified Jaya Optimization-based band selection model to reduce dimensionality. Second, the data is compressed using block compressed sensing with varying block sizes across different sampling rates, from which the best-performing configurations are selected for reconstruction. Third, the compressed data is reconstructed using the proposed Attention subsumed Residual Dense Network based Reconstruction algorithm, where the spectral–spatial features of the hyperspectral image are further enhanced through a fine-tuning network. Experiments are performed on Indian pines, Pavia university and Salinas datasets. Performance metrics Peak Signal-to-Noise Ratio and Mean Squared Error were chosen to evaluate compression and Mean structure similarity index measure, Mean Peak Signal-to-Noise Ratio, Mean feature similarity index measure and spectral angle mapper, were adopted for analyzing reconstruction stages. The proposed work on implementation, outperforms the state-of-the-art algorithms that are considered in this work for comparison, with lesser processing time.