ИСТИНА |
Войти в систему Регистрация |
|
ИПМех РАН |
||
Detecting polar mesocyclones massively in high-resolution satellite data is a challenging, time-consuming task for a human expert. Current algorithms for automatic detection are developed based on conventional field expertise and defined with some empirical thresholds of geophysical variables. At the same time, several detection methods may produce different results on identical data. This research presents a principally new method with the use of deep convolutional neural networks (CNNs) for polar mesocyclones detection. Train dataset is formed based on the reference database of satellite-derived polar mesocyclones trajectories in the Southern Hemisphere (Verezemskaya et al., 2017). We have tested six different architectures of CNN. Particularly CNN from-the-scratch, CNN based on VGG16 pre-trained weights (Simonyan, Zisserman, 2014) with the transfer learning technique (Pan, Yang, 2010) used, and CNN based on VGG16 with fine-tuning technique applied. Each of these networks was used with both IR and IR+WV satellite data. The best classification quality - 97% in terms of binary classification accuracy score - was reached by a CNN based on VGG16 pre-trained weights with transfer learning and fine-tuning techniques applied sequentially. The presented detection method could be easily extended to other atmospheric phenomena presented in satellite data with a distinct signature.