Leveraging Geospatial Data Gateways to Support the Operational Application of Deep Learning Models

Abstract

Geospatial data providers have adopted a variety of science gateways as the primary method for accessing remote geospatial data. Early systems provided little more than a simple file transfer mechanism but over the past decade, advanced features were incorporated to allow users to retrieve data seamlessly without concern for native file formats, data resolution, or even spatial projections. However, the recent growth in Deep Learning models in the geospatial domains has exposed additional requirements for accessing geospatial repositories. In this paper we discussed the major data accessibility challenges faced by the Deep Learning community namely: (1) reproducibility of data preprocessing workflows, (2) optimizing data transfer between gateways and computational environments, and (3) minimizing local storage requirements using on-the-fly augmentation. In this paper, we present our vision of spatial data generators to act as middleware between geospatial data gateways and Deep Learning models. We propose advanced features for spatial data generators and describe how they could satisfy the data accessibility requirements of the geospatial Deep Learning community. Lastly, we argue that satisfying these data accessibility requirements will not only enhance the reproducibility of Deep Learning workflows and speed their development but will also improve the quality of training and prediction of operational Deep Learning models.