4.2 Cloud Classification of Cellphone Photos by Machine Learning

Monday, 8 January 2018: 3:45 PM
Room 7 (ACC) (Austin, Texas)
Kun Yue, Moji Weather, Beijing, China; and J. Zhuang and L. Ding

Handout (31.4 MB)

The cloud observations are still inadequate from various cloud observation instruments and human naked eyes, for cloud structure and classification. However, the structure and classification are very important for short-range weather forecast of severe. This presentation focuses on cloud classification using real-time photos of hundreds of millions of users of MOJI weather APP. Through multiple machine learning models, we can process and analyze features of cloud, and recognize cloud patterns for its classification automatically.

The photo data includes: shooting locations/times and images which captured and uploaded by the user in real time. A training and testing dataset is labeled and classified by forecasters and meteorological experts. The labeled data is divided into two data sets of training and testing, which are available to the machine learning models.

In real time process, a coarse-grained model filters out “NO” images. The fine classification model identify the cloud images and categorize cloud types. According to the category and the amount of training data, we tried various deep learning models and combined with actual training data to achieve the best classification.

Based on the high quality of the output results, we will soon commercialize the algorithm of cloud classification. Users can easily identify the types of the clouds by MOJI Weather APP, and accumulate their scientific curiosity in cloud and meteorology, and keenly observe the weather changes. At the same time, our team plans to use this dataset as a supplemental data for weather analysis and short-term weather prediction, which will ultimately meet the user requirement.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner