A hybrid approach to cloud image classification
A.I. Elizarov1, A.V. Shaleev1,2, I.I. Galtsev2
1V.E. Zuev Institute of Atmospheric Optics of Siberian Branch of the Russian Academy of Science, Tomsk, Russia
2National Research Tomsk State University, Tomsk, Russia
Keywords: image classification, texture characteristics, image processing, neural network
Abstract
This paper considers the problem of classifying cloud images, which are complex texture structures with heterogeneous characteristics. Traditional image analysis methods do not always adequately classify such images, and modern deep learning methods require large amount of data and computational resources. The research focuses on evaluating the feasibility of developing a hybrid method combining traditional statistical approaches to texture description and state-of-the-art deep learning techniques. It was hypothesised that the high-level features extracted by a neural network during training can be insufficiently sensitive to subtle local differences in cloud formations. The hybrid approach was implemented and analysed; low-level texture features were extracted from the images before being analysed by the neural network. However, the test results showed that this technique did not improve the classification quality and turned out to be less effective in terms of accuracy compared to the use of unprocessed images. The results of this work can be of interest to specialists in of Earth remote sensing data analysis, meteorology, and development of new texture image analysis methods.
|
