Generalization and Validation of MACHI: Minimal Atmospheric Compensation for Hyperspectral Images

Abstract

Atmospheric correction (AC) is essential for deriving accurate surface reflectance from satellite imagery for ocean color studies. Previous attempts to apply AC to satellite data from the Norwegian small-satellite mission HYPSO (HYPerspectral Smallsat for Ocean observation) have been unsuccessful, with a problem of pro ducing negative blue-band reflectance. To address this, the Minimal Atmospheric Compensation for Hyperspectral Images (MACHI) algorithm was developed as a fast, unsupervised alternative that requires no external atmospheric data or sensor specific tuning. MACHI estimates atmospheric parameters in a simplified atmo spheric model by optimizing for spectral smoothness in the retrieved ground re flectance under physical constraints. This thesis generalizes the theoretical framework of MACHI to support arbitrary smoothing kernels and introduces a new Python implementation capable of efficient batch processing. The algorithm was applied to a selection of scenes from HYPSO-2, the second satellite of the HYPSO mission, and evaluated through a convergence analysis and validation against ground-based measurements from the AERONET OC program, as well as a comparison to retrievals from POLYMER, an established AC algorithm for water applications. Convergence analysis shows fast and stable convergence of MACHI. Validation indi cates that MACHI retrieves physically plausible atmospheric parameters and surface reflectance spectra that are comparable to both ground measurements and estab lished AC methods. This while also avoiding the negative reflectance artifacts ob served in previous attempts. However, the validation results also reveal a systematic overestimation, particularly in the blue region, as well as a residual absorption fea ture near 760 nm, indicating a need for further development and testing of MACHI. Overall, this work demonstrates the potential of smoothness-based optimization for hyperspectral AC and provides a foundation for future enhancements, such as more detailed atmospheric modeling, refined initialization, and testing on other sensors.