Research On Road Extraction And Change Detection From Remote Sensing Images With Vector Data

As an important kind of ground objects,the automatic recognition and extraction of road has always been a focused and difficult problem in the field of photogrammetry and remote sensing.With a small number of road vectors,training samples for the recognition of roads on remote sensing images can be obtained,thus improving the automatic level of road extraction.In return,roads extracted from the new remote sensing images can be used for the change detection of the old road vector data to evaluate if it can reflect the current situation.Under this background,this dissertation focuses on the technology of road extraction and change detection from remote sensing images with vector data.The major work and innovations of this dissertation are listed as follow:1.An automatic segmentation method for remote sensing images assisted by vector data has been proposed.The main idea of this method is using the road vector data to obtain a number of seed points,and select effective road sample points via checking the shapes of homogeneous regions(regions with similar spectral or grayscale features)in the corresponding positions at the image.As a result,classified image can be obtained.Two different operating processes have been designed to deal with images with different resolutions.Experiment results prove the effectiveness of this method.2.A road refine method for classified images based on Direction Consistency Segmentation is proposed.This method supposes that road regions have similar local directions.Therefore,pixels with similar main directions are merged into regions,among which roads can be recognized and preserved via geometric measurements such as LFI(Linear Feature Measurement)and area.After groups of contrast experiments with several existing methods,the proposed method is proved to have advantages in both accuracy,computational effectiveness and stability,thus can be used on images with different resolution and road distribution characteristics.Classified images with massive interference information can be handled as well.3.A neighborhood centroid voting algorithm is proposed to extract road centerlines from classified remote sensing images.This algorithm detects road centerlines via analyzing neighborhood shapes and computing the centroids of neighborhood polygons.In this process,non-road parts can be excluded as well.Contrast experiments with the angular texture method for binary images has been conducted.The experiment results prove the advantage of this algorithm in terms of computational effectiveness,stability and the adaptation to roads with different width.4.An integral strategy for automatic road extraction from remote sensing images assistedby road vector data has been proposed.For high-resolution images,spectral information is used to segment road regions,after which road centerlines are extracted and the vectorization and rectification process are performed.For low or moderate resolution images,grayscale or spectral information is used for the segmentation,after which the road refinement process is performed.The final result can be obtained after morphologic thinning,vectorization and rectification.The experiment results prove the effectiveness of the proposed method for automatic road extraction from remote sensing images assisted by road vector data.5.A change detection method between road vector data has been designed based on the idea of buffer region.An accuracy evaluation method for change detection of old vector data using extracted roads has been designed inspired by the accuracy evaluation method for the change detection of remote sensing images.The change detection result has been compared to the manual change detection result to test the effectiveness of the proposed method.