Blog
February 21, 2020
AI & Industry 4.0

Understanding Clouds Organisation from Satellite Images

Understanding Clouds Organisation from Satellite Images

Climate change has been at the top of our minds and the forefront of important political decision-making for many years. Shallow clouds play a huge role in determining the Earth’s climate. By classifying different types of cloud organisation, researchers at the Max Planck Institute for Meteorology hope to improve the physical understanding of these clouds, which in turn will help build better climate models.

There are many ways in which clouds can organise, but the boundaries between different forms of organisation are unclear. This makes it challenging to build traditional rule-based algorithms to separate cloud features. The human eye, however, is really good at detecting features—such as clouds that resemble flowers.

It is essential to remove the haze from climate models and bring clarity to cloud identification.

Figure 1: (a-d) Examples of the four cloud organisation patterns. (e) World map showing the three regions selected for the Zooniverse project. Pie charts show the area fractions of the human classifications for the regions and seasons.

The Challenge
To build a multiclass segmentation model for cloud organisation patterns from satellite images to help scientists better understand how clouds will shape Earth’s future climate. This research will guide the development of next-generation models that could reduce uncertainties in climate projections.

Duration of Competition: 3 months

The Solution
As the task was very domain-specific, it was necessary to build the whole architecture of the model that can take image specifics into consideration and learn from noisy labels. This competition was evaluated using the mean Dice coefficient. The Dice coefficient can be used to compare the pixel-wise agreement between a predicted segmentation and its corresponding ground truth. To build the final model, the Ciklum AI team applied multiple state-of-the-art deep learning models for segmentation and classification. 

The Result
The final result was evaluated using the mean Dice metric. The goal of the machine learning model is to maximise this value. A perfect model would have an accuracy of 1. The Ciklum team developed a model that identified cloud type with a Dice coefficient of 0.647.

Read also: