Using Deep Learning to Identify Potential Roof Spaces for Solar Panels

House, D, Lech, M and Stolar, M 2018, 'Using Deep Learning to Identify Potential Roof Spaces for Solar Panels', in 2018 12th International Conference on Signal Processing and Communication Systems (ICSPCS), Cairns, Australia, Australia, 17-19 December 2018, pp. 1-6.


Document type: Conference Paper
Collection: Conference Papers

Title Using Deep Learning to Identify Potential Roof Spaces for Solar Panels
Author(s) House, D
Lech, M
Stolar, M
Year 2018
Conference name 2018 12th International Conference on Signal Processing and Communication Systems (ICSPCS)
Conference location Cairns, Australia, Australia
Conference dates 17-19 December 2018
Proceedings title 2018 12th International Conference on Signal Processing and Communication Systems (ICSPCS)
Publisher IEEE
Place of publication USA
Start page 1
End page 6
Total pages 6
Abstract Solar photovoltaic (PV) installation businesses frequently encounter problems with lead generation. A commonly used approach to identify credible customers involves cold-calling contacts from a purchased database containing very limited information or information that is inaccurate, out of date, and doesn't identify whether the building already has solar PV installed. This process of contacting potential customers, therefore, is often time-consuming, cost-ineffective, and inefficient, which results in increased costs for customers to account for these limitations. The objective of the current research project is to propose a method of automating this industry problem by applying Deep Neural Networks (DNNs). A Semantic Segmentation Network (SegNet) will be utilized, with a database of satellite images and corresponding pixel label images. The SegN et will seek to identify buildings from satellite imagery, and to in turn identify whether buildings have pre-existing solar installations, using a cascaded Convolutional Neural Network (CNN). Transfer learning on the CNN will fine-tune the network to classify roofs of buildings into two categories of having solar PV installed and not having solar PV installed. The CNN will be trained and tested on separate augmented databases to improve the classification accuracy with the output of the system recording a database of buildings without solar PV installed. By automating what was previously a time-consuming manual process, the savings incurred can be passed onto customers. Results of the current project demonstrate successful segmentation of buildings and identification of pre-existing solar PV installations. Implications of results are discussed.
Subjects Signal Processing
Keyword(s) Databases
Training
Buildings
Testing
Deep learning
Image segmentation
Satellites
DOI - identifier 10.1109/ICSPCS.2018.8631725
Copyright notice © 2018 IEEE
ISBN 9781538656020
Versions
Version Filter Type
Altmetric details:
Access Statistics: 11 Abstract Views  -  Detailed Statistics
Created: Mon, 29 Apr 2019, 13:04:00 EST by Catalyst Administrator
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us