Deep Learning for Rare Energy Infrastructure in Satellite Imagery (2020-2021)


Critical information for energy access decision-making and electricity system planning is not universally available. Information on village-level electricity access and reliability as well as the location and characteristics of power system infrastructure is especially scarce. Decision-makers require information to determine the optimal strategies for deploying energy resources, in order to decide where to prioritize development and whether electrification should be accomplished through grid expansion, microgrids or distributed generation.

Past studies have demonstrated that deep learning techniques may be successfully applied to satellite imagery to fill these knowledge gaps. However, these studies have focused on one type of infrastructure or system characteristic and were applied in one specific geographic region and were therefore limited in scope. Additionally, these techniques often require more training data to adapt to new contexts or identify rare objects. This costly process may inhibit applications of these techniques at scale.

Project Description

This project aims to develop deep learning techniques that can automatically and rapidly scan massive volumes of remotely sensed data, such as satellite imagery, to develop detailed maps of energy infrastructure. These deep learning approaches may provide a powerful tool for researchers, policy-makers and governments to collect energy systems information.

In past years, this Bass Connections project team has developed deep learning models that can detect energy infrastructure in satellite imagery, including power transmission lines and off-grid distributed energy generation sources such as solar photovoltaics. However, deep learning models must be trained to identify energy infrastructure using hand-labeled examples of the target objects in the real-world. Due to the rarity of such objects, human annotators must visually inspect large quantities of imagery to identify training examples.

To address this problem, the 2020-2021 project team will generate synthetic imagery from virtual worlds, such as those in modern video games. In such virtual worlds, the precise locations of all objects are known in advance, making costly hand-annotation unnecessary. Team members will create virtual worlds with various types of energy infrastructure and imaging conditions.

Using these scalable methods of data collection, the team will develop deep learning models that can identify an increasingly large variety of energy infrastructure to fill information gaps. The goals of this project are to investigate a new form of training data procurement by using synthetically generated imagery; investigate multiclass energy object identification at scale; and make data for decision and policy-makers in the energy access and system planning space open source and available for two cities.

Anticipated Outputs

Multiclass dataset of rare energy infrastructure objects; tool for synthetic generation of overhead imagery containing rare objects; poster; final presentation; paper for a conference or journal


Fall 2020 – Spring 2021

  • Fall 2020: Complete a short course on data analytics, energy systems and research techniques; define project goals
  • Spring 2021: Implement the project; meet with guest speakers from industry and academia; refine research results; create presentations and posters

Team Outputs to Date

Baseline dataset and supplemental synthetic images

Estimating Solar PV Capacity and Building Energy Use with Satellite Imagery (presentation by Kyle Bradbury, Energy Data Analytics Symposium, Duke University, December 8-9, 2020)

See related teams, Creating Artificial Worlds with AI to Improve Energy Access Data (2021-2022) and A Wider Lens on Energy: Adapting Deep Learning Techniques to Inform Energy Access Decisions (2019-2020), and Data+ summer project, Deep Learning for Rare Energy Infrastructure in Satellite Imagery (2020).


Image from Electric Transmission and Distribution Infrastructure Imagery Dataset by Varun Nair, Tamasha Pathirathna, Xiaolan You, Qiwei Han, Kyle Bradbury


Team Leaders

  • Kyle Bradbury, Pratt School of Engineering-Electrical & Computer Engineering|Energy Initiative
  • Jordan Malof, Pratt School of Engineering-Electrical & Computer Engineering

/graduate Team Members

  • Jose Moscoso, Interdisciplinary Data Science - Masters

/undergraduate Team Members

  • Tyler Feldman
  • Eddy Lin
  • Yanchen Ou
  • Baoyan Ye
  • Wendy Zhang, Computer Science (BS), Economics (BS2)

/yfaculty/staff Team Members

  • Leslie Collins, Pratt School of Engineering-Electrical & Computer Engineering
  • T. Robert Fetter, Nicholas Institute for Environmental Policy Solutions
  • Marc Jeuland, Sanford School of Public Policy
  • Luana Lima, Nicholas School of the Environment-Environmental Sciences and Policy
  • Robyn Meeks, Sanford School of Public Policy