Probabilistic Retrieval of All-Day Overlapping Cloud Microphysical Properties

Approximately 25% of clouds globally are overlapping clouds, critical for the Earth’s radiation budget and weather system evolution. However, traditional physical methods fail to retrieve all-day cloud microphysical properties from passive remote sensing satellites due to their complex vertical structure, which remains challenging. To address this, we propose Diff-IR, a probabilistic deep learning model to retrieve overlapping cloud microphysical properties from Aqua satellite’s thermal infrared channels. Evaluation shows Diff-IR excels in cloud phase classification with 88.18% overall accuracy and 76.08% multi-layer cloud precision rate during the daytime, the model retrieves upper-layer ice clouds with RMSEs of 6.66 µm for cloud effective radius (CER) and 2.78 for cloud optical thickness (COT), and lower-layer water clouds with RMSEs of 19.60 µm (CER) and 11.75 (COT). Diff-IR outperforms the deterministic model with the same input during the daytime, particularly in capturing probabilistic distributions. Additionally, generating diverse ensemble members helps the model estimate uncertainty, enhancing retrieval reliability.