Congratulations to Yuanming Shu who successfully defended his PhD thesis entitled “Deep Convolutional Neural Networks for Object Extraction from High Spatial Resolution Remotely Sensed Imagery” on December 18, 2014. The committee includes his supervisor, Professor Jonathan Li, the external examiner Professor Yun Zhang from the Department of Geodesy and Geomatics Engineering, University of New Brunswick; Internal/External examiner, Professor Alex Wong from the Department of Systems Design Engineering; and other two members who are Professor Michael Chapman from the Department of Civil Engineering, Ryerson University and Professor Richard Kelly from the Department of Geography and Environmental Management. The abstract of Yuanming's PhD thesis is as follows:
Abstract
Developing methods to automatically extract objects from high spatial resolution (HSR) remotely sensed imagery on a large scale is crucial for supporting real-world applications with HSR imagery. However, this task is notoriously challenging. Deep learning, a recent breakthrough in machine learning, has shed light on this problem. The goal of this thesis is to develop a deep insight into the use of deep learning to develop reliable automated object extraction methods for applications with HSR imagery. The thesis starts by re-examining the knowledge the remote sensing community has achieved on the problem, but in the context of deep learning. Attention is given to objectbased image analysis (OBIA) methods, which are currently considered to be the prevailing framework for this problem and have had a far-reaching impact on the history of remote sensing. In contrast to common beliefs, experiments show that object-based methods suffer seriously from ill-defined image segmentation. Further, they are significantly less effective at leveraging the power of the features learned by deep convolutional neural networks (CNNs) than conventionally patch-based methods. This thesis then also studies ways to further improve the accuracy of object extraction with deep CNNs. Given that vector maps are required as the final format in many applications, the focus is on addressing the issues of generating high-quality vector maps with deep CNNs. A method combining bottom-up deep CNN prediction with top-down object modeling is proposed for building extraction. This method also exhibits the ii potential to extend to objects of interest. Experiments show that implementing the proposed method on a single GPU results in the capability of processing 750 km2 of 12 cm aerial images in about 20 hours. By post-editing on top of the resulting automated extraction, high-quality building vector maps can be produced about 4-times faster than conventional manual digitization methods.