Hyperspectral imaging platforms such as UAVs and small satellites face strict constraints in data transmission and onboard storage. Lossy compression applied directly to raw sensor measurements offers substantial benefits for bandwidth efficiency. However, most existing studies evaluate compression only after radiometric and atmospheric correction, leaving the propagation of compression-induced errors through the complete pre-processing chain poorly understood. This study addresses this gap through a quantitative analysis of how reconstruction errors evolve from compressed raw data to georeferenced surface reflectance.<br />A representative set of state-of-the-art learning-based compression methods, including A1D-CAE, NLPCA, HyCoT, SSCNet and 3D-CAE, was evaluated. All models were trained on UAV-borne HySpex data and tested at a fixed compression rate of <em>c_R</em> = 4. This setting was chosen as a controlled and practically relevant operating point for a consistent comparison across all investigated methods. Original and reconstructed raw test data were processed with identical metadata through radiometric calibration, georeferencing and atmospheric correction. Reconstruction fidelity was assessed at four pre-processing stages using complementary spectral and spatial metrics.<br />The results show that spectral models retain high reconstruction accuracy throughout the workflow, with minimal error accumulation during pre-processing. Spatial and spatio-spectral architectures introduce spectral distortions that persist after atmospheric correction. These findings indicate that, for the investigated UAV-based HySpex dataset and the fixed compression setting of <em>c_R</em> = 4, spectral compression of raw hyperspectral data can preserve high-quality georeferenced reflectance products throughout the considered pre-processing workflow.