J. J. Miller

Deep learning for phenomena-based classification of Earth science images

Abstract. Automated classification of images across image archives requires reducing the semantic gap between high-level features perceived by humans and low-level features encoded in images. Due to rapidly growing image archives in the Earth science domain, it is critical to automatically classify images for efficient sorting and discovery. In particular, classifying images based on the presence of Earth science phenomena allows users to perform climatology studies and investigate case studies. We present applications of deep learning-based classification of Earth science images.

Detection of transverse cirrus bands in satellite imagery using deep learning

Abstract We demonstrate the viability of using a convolutional neural network (CNN) for facial recognition of meteorological phenomena in satellite imagery. Transfer learning was used to fine tune the widely used VGG-16 network architecture and allow the network to successfully detect (94% accuracy) the presence of transverse cirrus bands (TCBs) in NASA MODIS and VIIRS satellite browse imagery. The CNN exhibited better performance compared to a random forest classifier (84% accuracy) and was further validated by applying it to NASA satellite browse imagery in order to create a multi-year (2013–2015) global heat map of TCB occurrence. The annual heat map shows spatial patterns that are consistent with known mechanisms for the generation of TCBs, providing confidence in the CNN classifications. Our study shows that CNNs are well suited for meteorological phenomena detection due to their generalization capabilities and strong performance. An immediate application of our work intends to enable phenomena-based search of big satellite imagery databases. With additional modifications, the CNN could be utilized for other applications such as providing situational awareness to operational forecasters or developing phenomena specific climatologies.