Abstract

Computer vision can make important contributions to the analysis of remote sensing satellite or aerial imagery. However, the resolution of early satellite imagery was not sufficient to provide useful spatial features. The situation is changing with the advent of very-high-spatial-resolution (VHR) imaging sensors. This change makes it possible to use computer vision techniques to perform analysis of man-made structures. Meanwhile, the development of multi-view imaging techniques allows the generation of accurate point clouds as ancillary knowledge.

This dissertation aims at developing computer vision and machine learning algorithms for high resolution aerial imagery analysis in the context of application problems including debris detection, building detection and roof condition assessment. High resolution aerial imagery and point clouds were provided by Pictometry International for this study.

Debris detection after natural disasters such as tornadoes, hurricanes or tsunamis, is needed for effective debris removal and allocation of limited resources. Significant advances in aerial image acquisition have greatly enabled the possibilities for rapid and automated detection of debris. In this dissertation, a robust debris detection algorithm is proposed. Large scale aerial images are partitioned into homogeneous regions by interactive segmentation. Debris areas are identified based on extracted texture features.

Robust building detection is another important part of high resolution aerial imagery understanding. This dissertation develops a 3D scene classification algorithm for building detection using point clouds derived from multi-view imagery. Point clouds are divided into point clusters using Euclidean clustering. Individual point clusters are identified based on extracted spectral and 3D structural features.

The inspection of roof condition is an important step in damage claim processing in the insurance industry. Automated roof condition assessment from remotely sensed images is proposed in this dissertation. Initially, texture classification and a bag-of-words model were applied to assess the roof condition using features derived from the whole rooftop. However, considering the complexity of residential rooftop, a more sophisticated method is proposed to divide the task into two stages: 1) roof segmentation, followed by 2) classification of segmented roof regions. Deep learning techniques are investigated for both segmentation and classification. A deep learned feature is proposed and applied in a region merging segmentation algorithm. A fine-tuned deep network is adopted for roof segment classification and found to achieve higher accuracy than traditional methods using hand-crafted features.

Contributions of this study include the development of algorithms for debris detection using 2D images and building detection using 3D point clouds. For roof condition assessment, the solutions to this problem are explored in two directions: features derived from the whole rooftop and features extracted from each roof segments. Through our research, roof segmentation followed by segments classification was found to be a more promising method and the workflow processing developed and tested. Deep learning techniques are also investigated for both roof segmentation and segments classification. More unsupervised feature extraction techniques using deep learning can be explored in future work.

Publication Date

8-2017

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

John P. Kerekes

Advisor/Committee Member

Pengcheng Shi

Advisor/Committee Member

Carl Salvaggio

Campus

RIT – Main Campus

Share

COinS