Cloud Classification via Satellite Images using Residual Network

Identifying cloud shapes and predicting the weather based on these shapes is an important phenomenon for many people and businesses. Making a suitable definition for the clouds obtained from satellite images and revealing the cloud shapes will make things easier in weather forecasting. Although the prediction of cloud shapes using artificial intelligence has become widespread recently, the problem has still not been fully resolved due to complex clouds structure. This study addresses a problem that involves multi-label features in a training dataset of 5546 samples using a deep learning method. Clouds have four different types: fish, flowers, sugar, and gravel. In this multi-label classification pattern Resnet50 model, we studied images with 256x256 pixels, using and comparing two different loss functions, namely Binary Cross-Entropy and Weighted Binary Cross-Entropy. With the applied method, a success score of 69.1% was achieved and better results were obtained than the methods at the state-of-the-art level.