[3F1-ES-2] Machine learning: Social application (3)

[3F1-ES-2-05] Identifying the Snowfall Cloud at Syowa Station, Antarctica via a Convolutional Neural Network

〇Kazue Suzuki¹, Masaki Shimomura², Kazuyuki Nakamura², Naohiko Hirasawa³, Hironori Yabuki ³, Takashi Yamanouchi³, Terumasa Tokunaga⁴ (1. Hosei University, 2. Meiji University, 3. National Institute of Polar Research, 4. Kyushu Institute of Technology)

Keywords:CNN, Satellite Observation, Meteorology

This study evaluated snowfall values based on limited observation data to estimate the surface mass balance (SMB) of Antarctica. To accomplish this, we attempted to identify the snowfall cloud at Syowa Station, Antarctica. We constructed a new convolutional neural network (CNN) architecture with multinomial and binary classifications and added National Oceanic and Atmospheric Administration (NOAA)/Advanced Very High Resolution Radiometer (AVHRR) images over five years. The CNN was based on VGG16, and concatenate layers were added as the inception module. We replaced all the convolution layers with global average pooling to reduce the number of parameters. Based on the positive CNN sample result, the multinomial classification emphasized the entire cloud structure, while the binary classification focused on cloud continuity. The results indicated accuracies of 71.00% for binary and 65.37% for multinomial classifications.