All-day cloud property and occurrence probability dataset based on satellite remote sensing data.

Area of Science:

• Meteorology and Climate Science
• Remote Sensing
• Artificial Intelligence in Earth Observation

Background:

• Accurate cloud property datasets are vital for meteorological research, climate studies, and applications.
• Existing satellite cloud products often lack comprehensive nighttime data and diverse cloud property information.
• CldNet Version 2.0 builds upon previous work, extending cloud recognition capabilities to a full day-night cycle.

Purpose of the Study:

• To introduce CldNet Version 2.0 (CldNetV2) for enhanced cloud type classification and property prediction.
• To generate comprehensive, all-day datasets of cloud properties and occurrence probabilities.
• To address limitations in current Himawari cloud products, particularly for nighttime conditions.

Main Methods:

• Leveraging transfer learning and model parameter sharing techniques from the foundational CldNet.
• Developing a deep learning model for classifying cloud types and predicting multiple cloud properties.
• Statistically analyzing cloud type occurrence probabilities across annual, seasonal, and monthly time scales, distinguishing between all day, daytime, and nighttime.

Main Results:

• CldNetV2 successfully classifies cloud types and predicts additional cloud properties, creating valuable nighttime datasets.
• Generated datasets include cloud type occurrence probabilities across various time scales and conditions (all day, daytime, nighttime).
• Independent validation using CALIPSO, ERA5, and visualization confirms the reliability of the CldNetV2 cloud products.

Conclusions:

• CldNetV2 significantly advances the capability for all-day cloud property and occurrence probability assessment.
• The publicly released dataset provides a valuable resource for meteorological environment assessment and climate research.
• This work contributes to improved understanding and modeling of atmospheric processes through comprehensive cloud data.