Road Extraction Based On Unsupervised Classification And Geometric-texture-spectral Features For High-resolution Remote Sensing Images

As an infrastructure with great economic value,roads are the foundation of social and economic development.The high-resolution remote sensing technology provides a convenient and efficient new way for road network information acquisition.At present,the supervised classification is mostly used in high-resolution remote sensing image road extraction,but the manual selection of sample is needed.Thus it is with low automation and low instability.The pixel-based road extraction method is with low integrity and easy to produce salt-and-pepper noise,while the object-oriented method is easy to produce the adhesion problem.Although there are various road extraction methods in the literature,road extraction has still not been well resolved due to many negative factors,such as the complexity of natural scenes,the noise interference in remote sensing images,and the limitation of extraction algorithms.Therefore,road extraction in high-resolution remote sensing images has remained the challenging research subject.In order to improve the integrity,accuracy and automation of road extraction,from the perspective of multi-feature fusion,a multi-feature road extraction method is proposed in this paper.The main works are as following:(1)The classification system of roads in China is analyzed,and the characteristics of roads in high-resolution remote sensing images are summarized.Then,the interference factors of roads under different scenes are illustrated.(2)A road extraction method for high resolution remote sensing images is proposed based on the unsupervised classification and geometric-texture-spectral features.And a complete non-road region filter system is set up.To verify the effectiveness of the proposed method in this paper,six different high-resolution remote sensing images of different satellites and different resolutions for three scene types(city area,rural area and mountain area)are selected in the experiments.And in order to further prove the feasibility and effectiveness of the proposed method,other two typical pixel-based and object-oriented road extraction methods in China and abroad are compared.The results show that our proposed method can effectively reduce the salt-and-pepper noise and adhesion phenomenon while high degree of automation is obtained.(3)To solve the shadow and occlusion problem which are easily produced in high-resolution remote sensing images,we also propose an improved road extraction method based on Frangi linear target enhancement.Four high-resolution images are selected to carry out contrast experiments.The results show that our proposed improved method can effectively extract the linear feature in images and solve the shadow occlusion problem while with high stability and accuracy.