Scene-Level Meteorological Predictors Are Insufficient for AOD550 Estimation: A Baseline Study Toward Hyperspectral Data Fusion

This study investigates whether coarse scene-level meteorological variables alone are sufficient for estimating aerosol optical depth at 550 nm (AOD550), an important indicator of atmospheric aerosol burden widely used in climate, air quality, and environmental monitoring research. Although meteorological variables are frequently incorporated into aerosol-related machine learning frameworks, their standalone explanatory capability remains insufficiently quantified, particularly in simplified scene-level representations. This study addresses this gap by evaluating the predictive limits of meteorology-only modelling and establishing a baseline for future hyperspectral data fusion approaches.

The methodology was based on the integration of three data sources: HYPSO hyperspectral satellite scene metadata, CAMS AOD550 products, and ERA5 meteorological reanalysis variables. A dataset containing 33 satellite scenes was constructed, where each scene was represented by aggregated meteorological descriptors, including air temperature, surface pressure, wind components, and total column water vapour. CAMS-derived scene-level mean AOD550 values were used as the regression target. Two baseline regression approaches, Linear Regression and Random Forest, were applied to evaluate whether the selected meteorological predictors could explain aerosol optical variability.

The results demonstrated that meteorological predictors contained only weak to moderate relationships with AOD550, with total column water vapour and meridional wind showing the strongest associations. Both regression models exhibited limited predictive capability, substantial regression-to-the-mean behaviour, and low or negative coefficients of determination. The findings indicate that scene-level meteorological descriptors do not provide sufficient physical information for reliable AOD550 estimation when used in isolation. Consequently, the study confirms the necessity of incorporating richer spectral and optical information, particularly hyperspectral observations, into future aerosol retrieval and data fusion frameworks.