This Bass Connections project team furthered the mission of Duke’s Energy Data Analytics Lab to develop and apply tools to transform diverse energy data into innovative solutions for increasing the reliability, security, resilience and environmental sustainability of energy systems. Building on previous teams’ work, the 2016-17 team focused on automated building energy consumption from aerial imagery.

Detailed building-level energy consumption data for cities is rare due to the prohibitive collection costs, but could be used to identify and plan infrastructure and policy developments. Autonomous detection of objects such as buildings, roads, power lines and pipelines can be useful for policymakers to map infrastructure, track development patterns over time, find indicators of economic activity or quickly assess environmental damages.

Recent advances in computation for big data and image processing make it possible to learn about energy use in a fast and automated manner using machine learning. This project aimed to estimate building-level energy consumption from high-resolution aerial imagery by identifying buildings and extracting their properties (size, perimeter, etc.) and inputting these properties into an energy consumption estimation model. The model was designed using existing energy consumption data from the Department of Energy. To find an effective building detection technique, the team implemented both a traditional and a state-of-the-art deep learning classifier. Team members then applied their workflow to Gainesville, FL, to assess its effectiveness.

The team was able to demonstrate an approach to estimate building-level energy consumption given only high-resolution aerial orthoimagery. Overall energy estimation results for this study resulted in 7% error for a 2.25 km2 region. The team’s building detection approach identified over 80% of building pixels with fewer than 10% false detections.

Timing

Fall 2016 – Spring 2017

Team Outcomes

Artem Streltsov, Kyle Bradbury, Jordan Malof. “Automated Building Energy Consumption Estimation from Aerial Imagery.” 2018. International Geoscience and Remote Sensing Symposium 1676-1679.

Automated Building Energy Consumption Estimation from Aerial Imagery (Mitchell Kim, Sebastian Lin, Sophia Park, Eric Peshkin, Nikhil Vanderklaauw, Yue Xi, Samit Sura, Hoël Wiesner, Kyle Bradbury, Leslie Collins, Timothy Johnson; Bass Connections Poster Award, Audience Choice)

Comprehensive code for building detection from high-resolution aerial images (random-forest classifier approach)

Comprehensive code for building detection from high-resolution aerial images (convolutional neural network approach

Aerial imagery object identification dataset for building and road detection, and building height estimation (dataset produced by Data+ team members)

Automated Building Energy Consumption Estimation from Aerial Imagery (final team presentation, April 21, 2017)

Automated Building Energy Consumption Estimation from Aerial Imagery (presentation by Eric Peshkin and Hoël Wiesner, Bass Connections Showcase, April 24, 2017)

Presentation to North Carolina legislators at the State Capitol, Graduate Education Day, May 16, 2017, Raleigh, NC (Hoël Wiesner)

This Team in the News

Fueling the Future

2017: A Year of #Winning for Energy Students at Duke

Duke University Energy Initiative’s Annual Impact Report - FY 2017

Duke Undergraduates Use Machine Learning Techniques to Evaluate Electricity Access in Developing Countries

Duke Student Shares Energy Data Analytics Lab Research with North Carolina Legislators

Student Laurels and Honors for 2017

Students Present Their Research and Learn from Each Other at the Bass Connections Showcase

Students Share Research Journeys at Bass Connections Showcase

Bass Connections Poster Awards Highlight Neurosurgery and Energy Research

Real Problems, Real Teamwork

Reflections on Mentoring from Bass Connections Graduate Students

See related teams, Energy Data Analytics Lab (2017-2018), Electricity Access in Developing Countries from Aerial Imagery (Data+ 2017) and Energy Data Analytics Lab (2015-2016).