End-to-End Optimized Lossless Hyperspectral Image Compression With Bit Partition

Lossless compression of hyperspectral images (HSI) is crucial in applications such as scientific archiving, geological exploration, and material identification, where spectral fidelity must be strictly preserved. However, the extremely high bit-depth and strong spectral redundancy of HSI make existing learning-based image compression methods mainly designed for RGB images ineffective, as they fail to model high-dimensional dependencies and suffer from inefficiency under high bit-depth sampling. To address these challenges, we propose LHIC-BP, the first end-to-end optimized lossless compression framework specifically tailored for hyperspectral data. The core idea is to optimize bitrate directly through conditional probability modeling, using a discrete logistic mixture model to estimate pixel-level probability distributions. To efficiently handle high-bit-depth hyperspectral inputs, LHIC-BP introduces a bit partition strategy that decomposes the image into Most Significant Plane (MSP) and Least Significant Plane (LSP). The MSP captures dominant intensity structures with low complexity, while the LSP retains fine-grained residual details that require stronger contextual modeling. To further enhance compression efficiency, we design a dedicated entropy model that leverages both spatial and spectral priors to refine conditional probability estimation for MSP and LSP. In the spatial dimension, a Parallel Spatial Condition Aggregation Module is introduced to efficiently eliminate local redundancy. Meanwhile, in the spectral dimension, an RWKV-guided Spectral Condition Extraction Module is used to capture long-range inter-band dependencies with linear complexity. Extensive experiments on the HyspecNet-11K datasets show that LHIC-BP consistently surpasses state-of-the-art methods, achieving superior compression ratios and validating its effectiveness for hyperspectral lossless compression.