Date of Award

Summer 2020

Document Type


Degree Name

Master of Science (MS)


Civil, Construction, and Environmental Engineering

First Advisor

Bai, Yong

Second Advisor

Sheng, Wenhui

Third Advisor

Hernandez, Jaime


This thesis presents the research results of using the Fully Convolutional Network (FCN) model to densely match pixels and accurately determine elevations utilizing low-high ortho-images on a construction site. The research project has three major tasks. First, this researcher developed the FCN model to build up the relations between the reference pixels’ patches from the low-ortho-image and the target pixels’ patches from the high-ortho-image, which are approximate 2:1 scale relations. The developed FCN model has two convolutional blocks and one max-pooling layer in the middle. 286,292 samples of well-matched reference patches and target patches were collected for training the FCN model. The validation results showed that it could match up to 88% features of translating 78×78-pixel reference patches to 39×39-pixel target patches. Second, the researcher used the pixel group-based multiprocessing scheme to improve the time efficiency of the developed dense pixel matching and elevation algorithm. In addition, with the FCN-model generated target patch predictions for the selected reference pixels, the developed dense matching and elevation algorithm has the time efficiency of 298 pixels/second for the 38,809 dense pixel grid case under the conditions that two rounds of pixel matching processes are conducted, and three patch-based correlation calculations are conducted in the three channels of a RGB target patch prediction and a target patch. Third, the researcher evaluated the elevation measurement accuracy of the developed method. The results showed that it had a low elevation measurement error of 5% on the experimental site under the condition that the reference pixels and target pixels were well-matched by the FCN-based target patch predictions. The success of this research project has advanced drone and deep learning technologies in the construction industry. Implementing these technologies will avoid the conflict between the survey operation and other construction operations at project sites and provide the site elevation information to the construction professionals in real-time. As a result, the efficiency of on-site construction will be improved dramatically.

Available for download on Sunday, August 21, 2022

Included in

Engineering Commons