Segmentation of cloud patterns from satellite images to improve climate models

This analysis performs cloud classification and segmentation using satellite images. Shallow clouds play a huge role in determining the Earth’s climate but poorly represented in climate models. Classification of different types of cloud organizations can improve the physical understanding of these clouds and be substantial for understanding climate change. Murky boundaries between different forms of clouds lead to obstacles in traditional rule-based algorithms cloud features separation. In this analysis, deep learning algorithms are build to identify regions in satellite images that contain certain cloud formations. The dataset used in the analysis is
prepared by Max Planck and released on the Kaggle platform. It contains four labels: Fish, Flower, Gravel, Sugar. The cloud segmentation and classification are done using deep learning models. These models led to Top 7% solution in the Kaggle competition. Various improvements of the algorithm are described and results are presented. The influence of the lack of computational resources such as the number of GPUs is being discussed.