Hybrid deep learning for hyperspectral cube reconstruction in computed tomography imaging spectrometry via angular spectrum propagation

• Introduces SGD-RODE for high-fidelity CTIS hyperspectral cube reconstruction. • Uses spectral weighting and graph-based ODEs for spatial-spectral consistency. • Applies SPP to correct distortions from diffractive optics and misalignment. • Employs MobileNetV4 for efficient spectral-spatial feature extraction. • Integrates GCAA to optimize loss, speed convergence, and improve stability. The process of reconstructing hyperspectral cubes in Computed Tomography Imaging Spectrometry (CTIS) encounters difficulties which result in low spatial-spectral fidelity and high reconstruction error and limited adaptability to different operational environmental conditions. The research introduces Spectral Gaussian Dynamic-weighted combined arms Reconstruction graph Ordinary Differential Equation (SGD-RODE) as a solution which enables dynamic control of spatial and spectral content while enhancing angular reconstruction capabilities. The shifted pixel preprocessing technique establishes a method to decode the intentionally modulated spectral information encoded by the diffractive optical element. The system utilizes MobileNetV4 for efficient spatial-spectral feature extraction and employs the Gaussian Combined Arms Algorithm (GCAA) for hyper parameter tuning. The model which is designed reached a PSNR value of 39.78 dB, SSIM value of 0.9596, spatial PSNR value of 36.79 dB and a spatial SSIM value of 0.952. The SGD-RODE method provides accurate results which maintain operational dependability at fast processing speeds for application in CTIS systems.