Over the past 5 years, Deep Learning has been applied successfully to solve pattern classification problems in computer vision. Recent work by our group has demonstrated that Deep Learning can be re-targeted for solving pattern classification, localization and segmentation problems for climate datasets. DL works best in the context of supervised learning, when labeled datasets are readily available. This talk will present the motivation, design and status of the ClimateNet dataset and training architecture.

ClimateNet is intended to be a schema for augmenting existing NetCDF datasets with ground truth label information. We will describe our efforts in acquiring labels from the climate community, and storing information, which can then in turn be readily ingested by Deep Learning frameworks. We also intend to release trained network architectures, which can be further adapted and customized by the climate science community. We believe that easy access to both a curated dataset, and a trained network architecture will be critical in terms of lowering the barrier of entry for scientists who are interested in incorporating Deep Learning into their existing workflows.