Retrieval of Cloud Optical Thickness Based on FY-4B Geostationary Satellite Multichannel Data Combined with Machine Learning

xiangqian wei 

Accurate real-time retrieval of cloud optical thickness (COT) is of great significance for meteorological operations and climate research. To address the limitations of traditional physical methods, which rely on prior parameters, exhibit slow response times, and suffer from poor adaptability, this study proposes a COT retrieval method based on multichannel data from the FY-4B geostationary satellite combined with machine learning. Using FY-4B visible (0.65 μm) and near-infrared (3.7 μm) data as inputs and high-precision COT products from the FY-3F polar-orbiting satellite as ground truth, a total of 342 valid samples were constructed through spatiotemporal matching and quality control. Two neural network models were designed to achieve end-to-end retrieval. The results indicate that Model A, which employs a decoupled feature extraction strategy, achieves the best performance, with a correlation coefficient of 0.770 and a root mean square error of 4.532 relative to the measured values, along with good spatiotemporal consistency. This approach overcomes the observational limitations of polar-orbiting satellites and provides a new pathway for intelligent cloud parameter retrieval.