A Comprehensive Study of Estimating Atmospheric Cloud Microphysical Properties Using Deep Learning Techniques

Cloud properties such as cloud optical thickness (COT) and cloud effective radius (CER) are essential for weather forecasting, climate monitoring, and Earth’s energy budget estimation. Traditional physics-based retrievals using independent pixel approximation (IPA) often incur biases due to three-dimensional radiative effects. While existing deep learning approaches reduce these biases, they demand large annotated datasets and high computational cost. This study frames cloud property retrieval as an information-limited learning problem (limited spectral information and limited training samples) and incorporates CloudUNet with Attention Module (CAM), a compact deep learning model with channel attention for joint estimation of COT, CER, and cloud mask from bi-spectral radiance observations. Using synthetic datasets from large-eddy simulation (LES) cloud fields, CAM outperforms state-of-the-art models in both direct radiance-based retrieval and IPA correction, achieving 38% better performance in terms of mean absolute errors (MAE) and higher correlation with true properties. Ablation studies demonstrate that CAM-based IPA correction achieves 73% and 80% MAE reduction relative to the IPA baseline when using no radiance input and single-band radiance, respectively. Including cloud mask information as input improves COT retrieval across deep learning models (except CAM) but degrades CER retrieval for all models except CAM, which shows a slight 3% MAE improvement. These findings highlight the advantage of joint retrievals of multiple cloud properties and IPA correction models under limited labeled data constraints.