The Application Of Marked Point Process In Road Extraction

With the rapid development of space and information techniques, it has becomepossible to receive all-weather, all-time, multi-satellite, multi-resolution and nearlyreal-time geographical data. The automatic processing and interpretation of the data isnow an important issue of remote sensing. Road network is one of the most importantinformation and its extraction has theoretical and practical significances.This thesis studies the extraction algorithm of road network based on markedpoint processes. These models are object-oriented and can make fully use of theproperties of the object. They also benefit from the stochastic framework (robustnessto noise, initialization, etc). To extract objects, a model is first built according to theproperties of the object and the optimization is achieved by a simulated annealingwith a RJMCMC algorithm.The main works include: (1) The theory of marked point processes is studied,including the concept of point processes, the factors that affect the convergence speedof RJMCMC, and the three parameters of simulated annealing algorithm(the initialtemperature, the temperature decreasing speed and the convergence condition. (2)Road network extraction algorithm based on segment models is studied and realized.The influences of initial condition, different prior models and proposals on thealgorithm are analysed. In order to improve the speed of the algorithm, an improvedbirth-and-death proposal based on data preprocessing and an initialization based onlocal Hough transform are put forward. The experiment results show that theimprovement can significantly improve the speed of the algorithm. (3) Road networkextraction algorithm based on polyline models is studied and realised. An improvedbirth-and-death proposal based on data preprocessing and an initialization based onvectorization are put forward which can significantly improve the convergence speed.(4) Hydrographic network extraction using a hierarchical model is studied. Thenetwork is modeled by tree structures, and a two-step algorithm is used to extract thenetwork: First, thick branches of the network are detected by image segmentation based on Markov Random Field. Second, thin branches are extracted using a recursivealgorithm based on marked point processes defined within the neighborhood.