Data comes from several field campaigns in which an aircraft mounted cloud particle imager (CPI) took photos of ice crystals during flight. The CPIVIEW software was used to cut and classify the data. CPIVIEW classification is user specified and usually very poor as compared to manual classification, so it makes sense to build a convolutional neural network to classify these particles by their shape. This is important because the shape determines the scattering properties of the ice crystals, and given a distribution, of the clouds, and ultamitely the Earth's atmosphere. Clouds are the most significant source of uncertainty in climate models for this reason. Finding a fast and accurate classifier can help close this gap in understanding so we can more accuratley predict the climate.
The data was processed at UIUC, choosen and manually classfied. Dr. Junshik Um (former employee of UIUC DAS) put together this data, which is NOT available here. Contact me if you wish to know more.