Retrieval of cloud properties from thermal infrared radiometry using convolutional neural network

In this study, a deep learning algorithm is developed to consistently retrieve the daytime and nighttime cloud properties from passive satellite observations without auxiliary atmospheric parameters. The algorithm involves the thermal infrared (TIR) radiances, viewing geometry, and altitude into a convolutional neural network (denoted as TIR-CNN), and retrieves the cloud mask, cloud optical thickness (COT), effective particle radius (CER), and cloud top height (CTH) simultaneously. The TIR-CNN model is trained using daytime Moderate Resolution Imaging Spectroradiometer (MODIS) products during a full year, and the results are validated and evaluated using passive and active products observed in independent years. The evaluation results show that the cloud properties retrieved by the TIR-CNN are well consistent with all available MODIS day-time products (cloud mask, COT, CER, and CTH) and night-time products (cloud mask and CTH). The retrieved COT and CTH also show good agreements with active sensors for both daytime and nighttime, indicating that the algorithm performs stably in the diurnal cycle. •Solar-independent retrieval of cloud properties for passive satellites with CNN.•Lower computational cost compared with traditional algorithms.•Achieving satisfactory accuracy without auxiliary atmospheric parameters.