DeepSolar: A Machine Learning Framework to Efficiently Construct a Solar Deployment Database in the United States

Abstract

We developed an accurate deep learning framework to automatically localize solar photovoltaic panels from satellite imagery and estimate their sizes. We used it to construct a comprehensive and publicly available solar installation database of the contiguous US. We demonstrated its value by identifying key environmental and socioeconomic factors correlating with solar deployment, such as income and education. We also found that the solar deployment density can be accurately estimated at the microscopic level with these factors using a novel predictive model.

The impact of policies and business models on income equity in rooftop solar adoption

“…As PV prices have declined, PV has become an economical good that yields direct financial benefits [2]. However, high-income households remain more likely to adopt than LMI households [3][4][5]. In 2018, a household earning more than $200,000 per year was about 4 times more likely to adopt PV than a household earning less than $50,000 (based on data defined in Methods).…”

A harmonised, high-coverage, open dataset of solar photovoltaic installations in the UK

“…PV installations: large and small. Yu et al 14 made a distinction between small-scale (residential) and large-scale solar: this distinction is important to bear in mind, for multiple reasons. Firstly, the type (residential/ commercial/utility) implies differences in the size and capacity, but also the geographical siting and the grid connectivity of an installation.…”

“…Initiatives based on automatic detection have in some cases published data derived from their systems. Yu et al 14 publish estimates that are gridded (i.e. low-resolution) covering the whole contiguous USA.…”

“…These are laborious to collect, especially when precise outlines are desired. Note for example that Yu et al 14 employed a workaround using weakly-labelled data since polygon outlines were not available: if there was even a partial source of polygon data, the performance of automatic detectors could be much improved.…”

Locating Photovoltaic Installations with Deep Learning

“…6 In this issue of Joule, Yu et al report on a new database for photovoltaics in the contiguous US that uses a deep learning framework to analyze satellite data and recognize where and in what capacity solar cells are installed across the country. 7 Using their extensive framework, which is termed DeepSolar, there is potential for rich data extraction. As an example, the authors show that at a population density of 1,000 capita/mi 2 , PV installations are at peak deployment.…”