Building An Operationally Relevant Dataset From Satellite Imagery

A set of labeled training and testing data is required for most algorithmic approaches to image classification, and it is well known that data collection and curation is one of the biggest roadblocks to developing classification systems for real world applications. The quality of the dataset has a direct impact on the effectiveness of a trained algorithm for a given application. In this work we describe a dataset that we created to train algorithms to classify ships in overhead satellite imagery. Our experience revealed to us many challenges associated with datasets and with classification systems in general, which we outline as a cautionary tale to others embarking on a similar task.