Remotely sensed imagery has increased dramatically in quantity and public availability. However, automated, large-scale analysis of such imagery is hindered by a lack of the annotations necessary to train and test machine learning algorithms. In this study, we address this shortcoming with respect to above-ground storage tanks (ASTs) that are used in a wide variety of industries. We annotated available high-resolution, remotely sensed imagery to develop an original, publicly available multi-class dataset of ASTs. This dataset includes geospatial coordinates, border vertices, diameters, and orthorectified imagery for over 130,000 ASTs from five labeled classes (external floating roof tanks, closed roof tanks, spherical pressure tanks, sedimentation tanks, and water towers) across the contiguous United States. This dataset can be used directly or to train machine learning algorithms for large-scale risk and hazard assessment, production and capacity estimation, and infrastructure evaluation.
Read full abstract