Road Extraction Of High-resolution Remote Sensing Images Based On The Time-frequency Feature And SVM

As an important national infrastructure, the number of roads increases yearly under the raising government investment. It is obvious that updating roads information in real time is of crucial importance in vehicle navigation, traffic management, emergency reaction, urban planning, smart city and some other correlative studies. In addition, remote sensing technology, providing massive near-real-time high-resolution images with richer spectral information, texture information and spatial structure information, makes it possible to extract roads from remote sensing images. Road recognition and extraction from high-resolution remote sensing images has become a significant research subject in image processing, computer vision and image interpretation. It not only enriches fundamental geographic information but also has been widely applied to national defense, urban planning, environmental protection, geographic information updating and agricultural research.Collecting data from satellite images and aerial photos, we combine time-frequency features extraction methods with support vector machine classification to extract roads from high-resolution remote sensing images. The main research contents and results are as follows:1. Described the purpose and meaning of extracting roads from satellite images and aerial photos. The present research situation at home and abroad as well as several major factors influencing the road extraction were analyzed.2. Introduced the basic theoretical knowledge of remote sensing image preprocessing, including geometric correction, radiometric correction, image enhancement and frequency domain enhancement, and remote sensing image fusion, to provide a theoretical basis for remote sensing image preprocessing.3. Studied a road extraction algorithm based on texture features extracted by the geostatistic method and support vector machine classification from remote sensing images. First, this part analyzed the features of road images with different resolution and introduced the basic concepts of the variogram algorithm of geostatistical techniques and support vector machine model (SVM). Then it extracted primary roads based on texture features obtained by the variogram algorithm and support vector machine classification. Furthermore, it processed road areas based on the mathematical morphology combining with road morphology characteristics. Last it extracted central lines to build the road network. Multi-groups experimental results show that this method can extract roads from complex remote sensing images well.4. Proposed a road extraction algorithm based on time-frequency features extraction methods and domain adaptive support vector machine classification from high-spatial-resolution remote sensing images. First of all, this part combined the time domain textural feature extracted by the geostatistic method with the frequency domain spectral feature extracted by three dimensional wavelet transform to model features of the road. Second, it used extracted features to train the domain adaptive transfer learning support vector machine to complete the coarse classification of the image. Last but not least, it used the mathematical morphology method combining with road morphology characteristics to post process the image to extract roads completely. Simulation results show that the proposed method can enhance the precision and accuracy by reducing disturbance factors in the high-resolution aerial remote sensing images.