The Identification of the snowfall cloud at Syowa Station, Antarctica

Abstract:

Here, we attempt to estimate the surface mass balance (SMB) of Antarctica by accounting of the snowfall values based on the spatial synoptic patterns among some elements (e.g. geopotential height, relative humidity, sea ice concentration, and so on) for several decades.

For this subject, we investigate the relationship among the atmospheric synoptic patterns and cloud patterns from satellite data. The characteristic spatial patterns between atmospheric elements and clouds can be defined based on the observation data at Syowa Station. We did construct new Convolution Neural Network (CNN) Architectures and added images for 5 years. The image data are Ch.4 of NOAA/AVHRR. The ‘positive’ cloud with snowfall were tagged in case of the snow weather condition and high cloud amount based on observation. The ‘semi-positive’ cloud as like as ‘positive’ cloud but they do not have snowfall at Syowa Station or there is the lack of whole structure of the cloud. The CNN is based on VGG16 (Simonyan and Zisserman, 2014) and the concatenate layers have been added as Inception module. There is a clear bias between ‘positive’ and ‘negative’ in image sizes. This might occur the lower accuracy for the learning in multinomial classification. In case of binary classification, we have the almost same amount for each situation.

From the visualization by Grad-CAM (Selvaraju et al. 2016), for the multinomial classification, the classifier highlights the whole structure of the cloud. However, for the binary classification, it does the continuity of the cloud. The accuracies of learning were 87.30% for binary and 68.63% for multinomial classifications. For test, the accuracies were 71.00% for binary and 65.37% for multinomial classifications. There would be a room of improvement so we will add other tagged images to this data set and update.