Machine Learning-Based Retrieval of Cloud Droplet Number Concentration and Liquid Water Path From Satellite Spectral Data

Accurate estimation of cloud microphysical properties, particularly the cloud droplet number concentration ($N_{\mathrm{d}}$) and liquid water path ($L$), is essential for improving our understanding of aerosol-cloud interactions (ACI). Traditional satellite retrievals of these variables depend on assumptions that often lead to systematic errors. In this study, we present a machine learning (ML) framework that directly predicts $N_{\mathrm{d}}$ and $L$ from satellite spectral reflectance and radiance data, thereby circumventing conventional assumptions in retrieval algorithms. We use data from ICOsahedral nonhydrostatic large Eddy simulations simulations and moderate resolution imaging spectroradiometer-like spectral channels to evaluate the relevance of spectral features using traditional statistical techniques and ML interpretability methods. Our results demonstrate that, using a neural network model, $L$ can be accurately predicted using three spectral channels, achieving a coefficient of determination ($R^{2}$) of 0.93 and a normalized mean absolute error (nMAE) of approximately 16% . The prediction of $N_{\mathrm{d}}$ requires seven channels, achieving an $R^{2}$ of 0.76 and an nMAE of approximately 26% . As expected, $N_{\mathrm{d}}$ requires a richer spectral representation than $L$. Our ML approach enables a more direct and flexible estimation of cloud properties by avoiding assumptions linked to intermediate retrieval variables. This framework offers new insights into spectral sensitivities and supports an alternative and potentially more robust assessment of ACI from satellite observations, potentially leading to improvements in climate model constraints.