Cloud Type Identification Using Data Fusion and Ensemble Learning

Abstract

Cloud type classification is a complex multi-class problem where total sky images are analysed to determine their category such as Stratus or Cirrus, among others. However, many properties of this domain make high classification accuracy difficult to achieve. In this paper, we design a novel fusion approach, showing that recent image classification architectures based on deep learning, such as Convolutional Neural Networks, can be improved using statistical features directly calculated from images. In this research, three powerful CNNs have been trained on a comprehensive dataset: VGG-19, Inception-ResNet V2 and Inception V3. Simultaneously, a pool of standard machine learning classifiers have been trained on 14 different statistical characteristics on each colour channel. The results evidence that a fusion approach of the predictions of an image-trained CNN and a feature-trained Random Forest classifier improves the classification ability of both methods individually, reaching 95.05% macro average weighted precision over 12 classes in a complex highly imbalanced dataset with noisy examples.