Integrated point-line-network model for road extraction based on tree-shaped point process

Unsupervised road extraction methods based on traditional point processes have long faced challenges such as bottlenecks in processing efficiency and deficiencies in topological connectivity. To improve these drawbacks, this study proposes a new integrated modeling framework. First step is to construct the integrated point-line-network model based on the tree-shaped point process, in which the relationships between point, line and network are supposed to be constrained according to topological structure features like branching, trend, connectivity of road. In second step, integrated point-line-network model is further constrained by spectral Gaussian mixture model and Kullback-Leibler divergence of road, and then extraction model is obtained. Third step is to redesign transfer kernels of Reversible Jump Markov Chain Monte Carlo (RJMCMC) for simulation and optimization of road extraction. Finally, different scales of sub-meter-level remote sensing images are tested, and the results show that efficiency of the proposed method is higher than traditional methods, and the connectivity is well maintained.