Labeled photovoltaic installations for orthographic aerial imagery in Queens, New York

TL;DR

This paper introduces a dataset of manually labeled photovoltaic installations in Queens, New York, to support research on urban solar energy deployment and computer vision models.

Contribution

The dataset provides densely labeled photovoltaic installations in a high-resolution, multi-channel urban aerial image set with potential for time-resolved studies.

Findings

The dataset includes 14,000 polygons covering nearly 380,000 m2 of photovoltaic installations.

The dataset features four-channel imagery (RGB + infrared) from a densely populated urban area.

Periodic re-acquisition of the source imagery allows for longitudinal analysis of solar deployment.

Abstract

Obtaining data about rooftop photovoltaic installations presents a challenge for energy researchers. Some research efforts have attempted to utilize computer vision approaches to identify photovoltaic installations from aerial imagery. This dataset consists of manually labeled locations of photovoltaic installations for publicly available aerial imagery of Queens, New York, USA in 2018. The labels comprise 14,000 polygons corresponding to roughly 5,500 separate installations. The median polygon size is 13 m2, with a total area of close to 380,000 m2. Researchers may be interested in applying this dataset for the training of deep learning models for computer vision or to investigate deployment of photovoltaics in urban areas. While other similar datasets exist, there are several unique aspects of this location that make it attractive for further study: it encompasses a densely populated, urban environment; imagery contains four channels (three colors, plus infrared); and the source dataset is re-acquired periodically by the state of New York, offering the opportunity for these labels to form the basis of a time resolved study of photovoltaic deployment.

Figures