LSCIDMR: Large-Scale Satellite Cloud Image Database for Meteorological Research.

Area of Science:

• Meteorology
• Computer Science
• Artificial Intelligence

Background:

• Clouds are crucial for weather phenomena and can be observed by meteorological satellites.
• Automatic classification of satellite cloud images is essential for meteorological status and future projections.
• Deep learning models require large-scale training data, but a comprehensive cloud image database was lacking.

Purpose of the Study:

• To introduce the Large-scale Cloud Image Database for Meteorological Research (LSCIDMR), a novel benchmark dataset.
• To provide a publicly available resource for training and evaluating deep learning models for satellite cloud image classification.
• To establish baseline performance metrics for deep learning methods on this new dataset.

Main Methods:

• Development of the LSCIDMR, containing 104,390 high-resolution satellite cloud images across 11 classes.
• Implementation of two annotation methods: single-label (LSCIDMR-S) and multiple-label (LSCIDMR-M).
• Evaluation of several representative deep learning methods on the LSCIDMR to establish performance baselines.

Main Results:

• The LSCIDMR is the first publicly available benchmark satellite cloud image database directly linking weather systems with images.
• The database comprises 104,390 images with 40,625 single labels and 414,221 multiple labels.
• Experimental results demonstrate the feasibility of training effective deep learning models using sufficiently large image datasets for cloud image classification.

Conclusions:

• The LSCIDMR provides a valuable resource for advancing meteorological research through deep learning.
• The established baselines offer a starting point for future research in automated cloud classification.
• Sufficiently large datasets are key to developing robust deep learning models for meteorological applications.