Lightweight Band-Adaptive Hyperspectral Image Compression With Feature Decouple and Recurrent Model

Advanced deep-learning methodologies have led to notable improvements in hyperspectral image compression. While most existing deep learning approaches primarily concentrate on reducing spatial redundancy, the challenge of addressing spectral redundancy remains unresolved. Furthermore, the implementation of current models in resource-limited settings is often impeded by their high parameter counts and computational demands. To address these challenges, we propose a lightweight band-adaptive hyperspectral image compression model (LBA-HIM) aimed at enhancing compression efficiency while ensuring low computational overhead. LBA-HIM incorporates a feature decoupling mechanism that effectively separates hyperspectral images into fundamental and detailed features, thereby facilitating the reduction of spatial redundancy while preserving critical image details. Additionally, a recurrent structure is integrated into the band encoding and decoding processes, enabling the utilization of prior information from previously processed bands in subsequent operations. This strategy contributes to more efficient data compression by minimizing spectral redundancy. An adaptive weighted fusion mechanism is also employed to optimize the integration of multi-level features. Evaluation across six hyperspectral datasets indicates that the LBA-HIM model significantly enhances both compression efficiency and image quality while simultaneously lowering computational costs. At a compression rate of 0.25 bits per pixel per band (bpppb), LBA-HIM achieves an average Peak Signal-to-Noise Ratio (PSNR) of 38.25 dB, which represents an improvement of approximately 2.5 dB over the current state-of-the-art (SOTA) techniques.