Topologically Correct Construction of a Road Graph Based on Skeletonization Using Neural Network Methods

The paper presents an improved method for automatic construction of topologically correct road graph based on skeletonization using neural network technologies. The main contribution of the study is the combination of multi-level texture analysis and deep learning to improve the accuracy and robustness of the algorithm. The method includes three key stages: statistical texture analysis at several levels (Mahalanobis distance, modified Bhattacharyya distance, texture cubes and co-occurrence matrices), post-processing and vectorization using mathematical morphology and the U-net neural network, as well as geometric and topological adjustment using polynomial approximation and splines. The experimental results demonstrate a significant improvement in the accuracy of road network extraction (up to 87.2% IoU) compared to traditional methods. The developed approach is successfully applied to processing satellite and aerial images, which confirms its practical value for updating geoinformation systems.