Retrieval of Cloud Properties from Thermal Infrared Radiometry Using Convolutional Neural Network

Abstract

Utilizing solar-independent thermal infrared (TIR) radiances, a convolutional neural network (CNN)-based framework (TIR-CNN) is developed to consistently retrieve cloud properties from passive satellite observations during both daytime and nighttime conditions. This framework enables the retrieval of diverse cloud properties, including cloud mask, cloud optical thickness (COT), cloud effective radius (CER), cloud top height (CTH), cloud base height (CBH), column cloud phase, and the identification of single/multi-layer clouds. The TIR-CNN framework primarily consists of two branches. In the first branch, the inputs include TIR radiances, viewing geometry, and altitude, producing outputs such as cloud mask, COT, CER and CTH. The network is trained using daytime Moderate Resolution Imaging Spectroradiometer (MODIS) products over a full year, and the results are validated and evaluated using passive and active products in an independent year. The evaluation results demonstrate that the retrieved cloud properties are well consistent with available MODIS daytime (cloud mask, COT, CER, and CTH) and nighttime (cloud mask and CTH) products. The retrieved COT and CTH also show robust agreements with active sensors during both daytime and nighttime, indicating that the algorithm performs stably across the diurnal cycle. The second branch of the TIR-CNN framework receives inputs including TIR radiances, altitude, landcover, lifting condensation level, and the retrieved cloud products from the first branch. It generates outputs such as CBH, cloud phase, and single/multi-layer cloud identifications. The comprehensive training, validation, and testing procedures are conducted using radar-lidar products from CloudSat/CALIPSO. The estimation of global CBH results in root-mean-square errors of 1.19 km and 1.91 km for single- and multi-layer clouds, respectively. The cloud classifier achieves total accuracies of 82% for single-layer clouds and 85% for multi-layer clouds. In addition, the model has remarkable accuracy in identifying cloud phase within each pixel's vertical column, particularly in distinguishing mixed-phase clouds with an ice cloud top.