Contents Chapter 1 Introduction 1 1.1 State-of-the-art road extraction methods 4 1.2 Book outline 9 Chapter 2 An Integrated Method for Road Extraction 11 2.1 Introduction 11 2.2 The proposed method 11 2.3 Experimental evaluation 24 2.4 Summary 43 Chapter 3 The Object-Based Method for Road Extraction 45 3.1 Introduction 45 3.2 The proposed methodology 45 3.3 Experimental results 55 3.4 Summary 62 Chapter 4 Fusion of Multiple Road Extraction Results Using Geometric Characteristics 64 4.1 Introduction 64 4.2 The proposed methodology 64 4.3 Experimental results 77 4.4 Summary 86 Chapter 5 Accurate Road Centerline Extraction 89 5.1 Introduction 89 5.2 Feature point based road centerline extraction 90 5.3 Centerline extraction without spatial connection constraints 96 5.4 Experiment and results 101 5.5 Summary 110 Chapter 6 Road Gap Connection 111 6.1 Introduction 111 6.2 The methodology 111 6.3 Experiment and results 115 6.4 Summary 120 Chapter 7 Conclusion and Recommendation 122 7.1 Summary 122 7.2 Future research 123 Bibliography 125 Appendix 134
摘要
Chapter 1 Introduction Urbanization is happening in just a few years in East Asia ， according to the World Bank (2015) ， as shown by the mass movement of people to cities and the emergence of urban settlements.The region will have more decades of urban growth as economies shift from agriculture to manufacturing and services.Within the advent of modem acquisition sensors (i.e.， Ziyuan-3 ， Ikonos ， and QuickBird) ， Very High Resolution (VHR) satellite images have become increasingly available and thus it is possible to monitor urbanization from space.Figure 1.1 presents three satellite images over Xuzhou city ， China ， captured by different sensors. Figure 1.1 Satellite images with different spatia1 resola tions ovr Xuzhou ， China Although we can obtain massive satellite images ， the useful knowledge is still limitβd.This is in part due to it is unable to timely process these images.To meet this challenge ， object extraction plays an important role in processing these satellite images to produce useful information.Among various ect features ， road extraction from satellite images has received much attention in the past decades.This is because that updated road layers in Geographic Information Systems (GIS) are critical for many urbanization issues ， such as urban expansion estimation ， urban planning ， and traffic/population movement monitoring.Despite this is not an easy task ， espely in urban areas ， computer-aided road network extraction from remotely sensed images provides a new opportunity to meet this challenge.Meanwhile ， road extraction with the aid of computer can reduce manual work load and improve efficiency.However ， after years of development ， there is still no compelling evidence that the state-of-the-art can produce reliable and satisfactory results for any situation (i.e.， rural ， urban/rural ， and urban environments) or any spatial resolution.Numerous road extraction methods have been proposed but it remains unclear as to when these methods will be operational ， as there is still r征e commer software regarding this topic.Therefore ， the road extraction problem remains a tremendous obstacle in the field of remote sensing.The is because that a number of factors complicate the road extraction task ， as summarized in the following section. Spectral similarity.The extraction of roads is particularly problematic in urban areas due to the spectral similarity of roads and other impervious surfaces such as buildings ， as illustrated in Figure 1.2.Indeed ， the sp∞tral separability of asphalt road surfaces and bituminous roofs is still not easy since they tend to share similar spectral properties. Figure 1.2 The spectral similarity between road and building roof Material change.In real world scenarios ， roads are made of various construction materials ， such as cement and asphalt.Generally ， different materials lead to different spectral characters.For instance ， cement roads have high intensity values on multispectral images ， while asphalt roads low intensity values [Fige 1.3(a)J.Hence ， the variety of road material threats to the reliability and accuracy of state-ofthe-art methods ， as most of methods can only process one of these two cases (i.e.，high intensity road or low intensity road) . Figure 1.3 Two complicated scenarios for road extraction Occupy.This problem is further exacerbated by issues related to the occlusion of road surfaces by trees ， shadow and the presence of vehicles [Figure 1.3 ( b ) J.The occlusion of trees and shadow commonly leads to discontinuities of the road segments extracted.Meanwhile ， vehicles on the road result in redundant ‘ holes' of the road segments exactβd.This issue is particularly common in dense urban environment ， such as Hong Kong. To overcome aforementioned challenges ， the integrating of spectral and spatial information plays an important role.Given the extreme difficulty of distinguishing road feature from other man -made structures (such as car park and building roof) ， deriving complementinformation (i.e.geometrical feature) 仕om VHR satellite image is a rational option.In any case ， understanding to what extent the combination of spectral and geometrical information improves the road extraction performance in Integrated Spectral and Gecmetrical Information for Road Extraction from VHR Satellite Images i 1.1 State-of-the-art road extraction methods To date ， road extraction from VHR satellite images h.as received a lot of a lot of attention ， and various methods have been proposed.These methods can be roughly classified inωtwo groups: ① automatr.c and ②semi-auωmatic.In general ，automatic methods can produce satisfacω'ry resu1t and thus are much closer to the 吨>erational level an autom.atic methods.The reason is at autom.atic methods generally cann.ot achleve satiry resu1ts inωmplicated scenes (i.e，dense urban environment).In contrast ，the human interacti.on is more robust ese cases.Despite i
