Abstract
Deep learning-based segmentation of urban digital surface models (DSMs) endures challenges from limited features, class imbalance, and sparse data, which limits the application of DSMs in urban solar energy assessment. In this study we propose a dynamic graph CNN (DGCNN) based segmentation model and address abovementioned problems by adding artificial features, using adaptive-weighted loss function, and introducing a modified spatial transformation module. Our model achieved outstanding performance in predicting the test dataset with an average accuracy of 0.95 and F1 scores of 0.94 after 300 epochs of training. The presented approach inspired several potential applications for solar energy simulation.
Original language | English |
---|---|
Title of host publication | Proceedings of the 2nd 4TU/14UAS Research Day on Digitalization of the Built Environment |
Chapter | Data-Driven Buildings and Urban Transformation |
Pages | 42-46 |
Number of pages | 5 |
ISBN (Electronic) | 978-90-386-5717-2 |
Publication status | Published - 27 Mar 2023 |
Event | 2nd Research Day on Digitalization of the Built Environment - Eindhoven University of Technology, Eindhoven, Netherlands Duration: 27 Mar 2023 → 27 Mar 2023 https://www.4tu.nl/bouw/Events/DigitalizationEvents/2nd%204TU-14UAS%20Research%20Day%20on%20Digitalization%20of%20the%20Built%20Environment/ |
Conference
Conference | 2nd Research Day on Digitalization of the Built Environment |
---|---|
Abbreviated title | 2nd 4TU/14UAS DBE |
Country/Territory | Netherlands |
City | Eindhoven |
Period | 27/03/23 → 27/03/23 |
Internet address |
Keywords
- Solar energy
- Digital surface model
- Machine learning