Jingyuan Yang

Cloud detection is critical for satellite-based atmospheric observations and subsequent applications. This study presents a deep learning-based cloud mask model, Convolutional Attention Cloud Mask Network Plus Plus (CACM-Net++), developed for both daytime and nighttime cloud detection using multispectral observations from the Advanced Geostationary Radiation Imager (AGRI) onboard the FY-4B geostationary satellite. The model incorporates physically meaningful input features and is independently trained across different satellite zenith angles to better capture radiative characteristics and enhance retrieval accuracy in complex scenes. Cloud labels derived from Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) observations are used for both training and evaluation. CACM-Net++ achieves over 90% detection accuracy under varying illumination conditions, with particularly strong performance over land and ocean at moderate satellite zenith angles. Compared to the operational cloud mask product from the National Satellite Meteorological Center (NSMC), CACM-Net++ demonstrates significant improvements, especially in detecting clouds over nighttime land areas and at high satellite zenith angles. These gains are primarily attributed to the model’s deep learning architecture and its tailored spatiotemporal partitioning strategy. Independent validation against CALIPSO confirms the model’s robust generalization across surface types. Furthermore, annual comparisons with ERA5 reanalysis data reveal that CACM-Net++ produces more consistent cloud distributions, particularly over nighttime land regions such as Australia. Despite its strong overall performance, the model remains challenged during twilight periods due to the scarcity of reliable reference data. Future work will focus on addressing this limitation by developing specialized algorithms for transitional lighting conditions, advancing the model’s all-sky detection capabilities.

Presents corrections to the paper, (Corrections to “Atmospheric Correction Above Clouds for FY-4A AGRI Based on RTTOV”).

Cloud optical thickness (COT) stands as a critical parameter governing the radiative properties of clouds. This study develops a convolutional neural network (CNN) model to retrieve the COT of single-layer non-precipitating clouds during nighttime using FY-4B satellite data. The model integrates multi-channel brightness, temperature, and geographic and temporal features, without relying on auxiliary meteorological data, using the multi-point averaged 532 nm COT from CALIPSO as ground truth for training. Performance evaluation demonstrates robust retrieval accuracy, achieving coefficients of determination (R²) of 0.88 and 0.73 for satellite zenith angles (SAZAs) < 70° and >70°, respectively. Key advancements include the incorporation of temporal features, the Squeeze-and-Excitation (SE) module, and a multi-point averaging technique, each validated through ablation experiments to reduce bias and enhance stability. Meanwhile, a model error analysis experiment was conducted that further clarified the performance boundaries of the model. These findings underscore the model’s capability to retrieve the COT of single-layer non-precipitating clouds during nighttime with high precision.

Developing a rapid and efficient method to mitigate atmospheric interferences above clouds is essential for precise retrieval of cloud physical properties through satellite remote sensing. In this study, Radiative Transfer for TOVS (RTTOV) was employed to generate lookup tables (LUTs) for FY-4A AGRI. These LUTs comprised information on atmospheric transmittance, Rayleigh optical depth, and upwelling emission above clouds. A total of 288 images in 2021 were analyzed. Following correction, notable differences in reflectance or radiance were observed between top-of-atmosphere (TOA) and top-of-cloud (TOC) measurements for bands 1, 6, 7, and 11, with average relative biases ranging from 6.94% to 17.60%. Conversely, bands 2, 3, 5, 12, and 13 showed relatively minor variations, with an average relative bias of less than 3.98%. Solar zenith angle, gas column concentration above clouds, and cloud top temperature were primary factors influencing the variations. Moreover, rapid correction equations were developed utilizing the least squares method. The exponential equation yielded the most precise for gas transmittance above clouds, while linear and exponential equations were found to be optimal for atmospheric upwelling emission in bands 7 and 11, and in bands 12 and 13, respectively. The root mean square error (RMSE) between radiance corrected by the equations and the LUTs, consistently stayed below 0.50, with a mean absolute percentage error (MAPE) below 4.03%, and a coefficient of determination (R ² ) exceeding 0.95. The rapid correction equations, capable of independently, efficiently, and accurately computing TOC reflectance or radiance, are highly suitable for cloud property inversion.

Accurate cloud detection is a crucial initial stage in optical satellite remote sensing. In this study, a daytime cloud mask model is proposed for the Advanced Geostationary Radiation Imager (AGRI) onboard the Fengyun 4A (FY-4A) satellite based on a deep learning approach. The model, named “Convolutional and Attention-based Cloud Mask Net (CACM-Net)”, was trained using the 2021 dataset with CALIPSO data as the truth value. Two CACM-Net models were trained based on a satellite zenith angle (SZA) < 70° and >70°, respectively. The study evaluated the National Satellite Meteorological Center (NSMC) cloud mask product and compared it with the method established in this paper. The results indicate that CACM-Net outperforms the NSMC cloud mask product overall. Specifically, in the SZA < 70° subset, CACM-Net enhances accuracy, precision, and F1 score by 4.8%, 7.3%, and 3.6%, respectively, while reducing the false alarm rate (FAR) by approximately 7.3%. In the SZA > 70° section, improvements of 12.2%, 19.5%, and 8% in accuracy, precision, and F1 score, respectively, were observed, with a 19.5% reduction in FAR compared to NSMC. An independent validation dataset for January–June 2023 further validates the performance of CACM-Net. The results show improvements of 3.5%, 2.2%, and 2.8% in accuracy, precision, and F1 scores for SZA < 70° and 7.8%, 11.3%, and 4.8% for SZA > 70°, respectively, along with reductions in FAR. Cross-comparison with other satellite cloud mask products reveals high levels of agreement, with 88.6% and 86.3% matching results with the MODIS and Himawari-9 products, respectively. These results confirm the reliability of the CACM-Net cloud mask model, which can produce stable and high-quality FY-4A AGRI cloud mask results.

Yajing Sun, Rui Zhang, Jingyuan Yang, Wei Peng. Findings of the Association for Computational Linguistics: ACL 2023. 2023.