High-resolution Building Height Estimation Based on Sentinel Satellite Data

Thumbnail Image
Date
2023-08
Authors
Yan, Xin
Major Professor
Harding, Chris
Advisor
Committee Member
Franz, Kristie
Hornbuckle, Brian
Journal Title
Journal ISSN
Volume Title
Publisher
Authors
Research Projects
Organizational Units
Journal Issue
Is Version Of
Versions
Series
Department
Abstract
High-resolution building height data is an important indicator for scientific research and practical application. I used Machine learning (ML) techniques to estimate building height at a vertical spatial resolution of 10m. I established spatial-spectral-temporal databases with 160 features (variables) by combining SAR data (provided by Sentinel-1) with optical data (provided by Sentinel-2) and building footprint polygons. From these 160 features, through repeated experiments and an expert scoring system, I established a set of 13 best features for ML models. Using 12 large, medium, and small cities in the United States as training data and testing multiple types of ML models on a separate testing data set, I selected Random Forests as the best-performing ML technique for my application. I used a 50*50m moving average window to aggregate the pixels to mitigate the impact of SAR image displacement and building shadows and to eliminate data noise caused by factors such as building materials. To ensure my random forest machine learning model's reliability, I tested the operating efficiency and accuracy of three model collection methods: bagging, boosting, and stacking, and finally chose the bagging method. I optimize the ML model for large-scale applications for fast production of high-resolution building heights. I tested the accuracy of my results by comparing them to Lidar data in a test area and found that the prediction model explained 98.72% of the variance.
Comments
Description
Keywords
Citation
DOI
Source
Copyright
2023