With the increasing application of hyperspectral imaging in remote sensing, agriculture, and environmental monitoring, hyperspectral data exhibit great potential due to their fine spectral and spatial details. However, the high dimensionality and redundancy of hyperspectral images present significant challenges for efficient storage and transmission. Traditional methods based on spectral transforms, prediction coding, or tensor decomposition often struggle to achieve a satisfactory balance between compression ratio and reconstruction quality. Learning-based image compression (LIC) techniques, particularly those relying on variational autoencoders (VAEs), have demonstrated remarkable success in natural image compression, yet their effectiveness in hyperspectral scenarios remains limited because of the absence of explicit frequency modeling and the restricted capacity of CNN-based context modeling to capture long-range dependencies. To address these limitations, we introduce two enhancements to the Causal Context Adjustment (CCA) framework. First, a Frequency-Enhanced Selective State Space Module (FE-SSM) performs multi-scale frequency decomposition and selective state updates for adaptive modeling of frequency components and improved feature expressiveness. Second, a Frequency-Aware Adaptive Loss Weighting (FA-ALW) mechanism dynamically adjusts the CCA loss according to the spectral energy distribution of input data, enabling content-adaptive optimization. Experiments on the widely used Indian Pines dataset demonstrate that our method achieves substantial reductions in bit rate and consistent improvements in PSNR and MS-SSIM, validating the effectiveness of integrating frequency-domain modeling with adaptive loss reweighting for hyperspectral image compression.