Extracting Urban River From High-Resolution Remotely Sensed Imagery Based On Features In Frequency Domain

The thesis was supported the National High Technology Research and Development Program of China "Research on the novel technique of image segementation based on High-Resolution Remotely Sensed Imagery", and the National Natural Science Foundation of China "Research on the algorithm of multi-scale segmentation based on frequency feature from High-Resolution Remotely Sensed Imagery"As the carrier of water resource in city, urban rivers have important functions in landscape ecology, flood control and Mitigating heat island effect. Thus, it is significant to recognize and extract the urban river. Because of the spatial resolution of remotely sensed imagery becoming higher, the traditional technology of classification based on pixels for low and medium resolution remotely sensed imagery can not satisfy the requirment in the processing of remotely sensed imagery, and is instead of the object-oriented method of features recognition and extracting gradually. Using object-oriented algorithm, land feature in city, such as road, building, river, green land, can be extracted readily. According to bilinearity of urban river in the high-resolution remotely sensed imagery, the urban river was divided into two parts, namely edge features and low-frequency information, in the research of this thesis.After the spectrum analysis of the urban river in IKONOS panchromatic image, the magnitude recognition mark of edge features and the low-frequency information of urban river can be decided. Subsequently, the corresponding two-dimension frequency Butterworth filters were designed to extract edge features and the low-frequency information of urban river. Finally, according to the edge features and low frequency information of urban river from IKONOS image, urban river was extracted. The content of this thesis includes:1) Using the radius distribution and angle distribution, the frequency spectrum of urban river in high-resolution remotely sensed imagery was analyzed, and spectrum signatures of different types of urban river were described and the connection of urban river between spatial domain and frequency domain was discussed. The magnitude spectrum was retrieved after the IKONOS image being transformed by fast Fourier transform, and it reflects the frequency distribution of different land features in frequency domain. For the urban river, there are three main types urban river including linear type, intersectional nonlinear type and curve nonlinear type. All types urban river have different features of spectral response in radius distribution and angle distribution of magnitude spectrum. Radius distribution and angle distribution are the accumulation of magnitude along the radial direction and angle direction respectively, thus, they can be used to analyzing spectral response globally for further spectral analysis.2) Edge features and low-frequency information of urban river were analyzed in frequency domain respectively. Meanwhile, the magnitude recognition mark of edge features and low-frequency information of urban river were decided respectively. Urban river in frequency domain can not be described by adius distribution and angle distribution in details. Thus, the distribution rule of dominant frequency and frequency doubling were obtained by analyzing the spectral response of edge of one-dimension signal in frequency domain. Then, it was retrieved that the relationship between the distance of adjacent Spectral peak and the distance of adjacent edge in spatial domain, and between the number of spectral peaks and the distance of adjacent edge in spatial domain respectively. Subsequently, the relationship was extended to the two-dimension image to build the magnitude recognition mark of edge features of urban river.As to the low frequency information of urban river in remotely sensed imagery, this thesis highlighted the information of urban river using Perceval theory, and then decided the magnitude recognition mark by analyzing the energy distribution in different radius of magnitude spectrum.3) Based on the magnitude recognition mark of edge features and low-frequency of urban river, the corresponding two-dimension Butterworth filters in frequency domain, were designed to extract edge feature and low frequency of urban river. In the extraction of edge features of urban river, since the urban rivers had different types in the research area, two-dimension Butterworth filters were designed according to the spectral response from edge features of different type’s urban river in thesis. Using the one-dimension Butterworth filter, this thesis designed the two-dimension band-pass Butterworth filter to extract edge features of linear river; designed the two-dimension arc ring band-pass log Butterworth filter to extract edge features of curve nonlinear river; designed two-dimension band-pass Butterworth filter bank to extract the edge of intersectional nonlinear type river. Subsequently, parameters of filter were set to retrieve edge features of urban river according to the magnitude recognition mark of edge features of urban river. In the process of extracting the low frequency information, the centre of two-dimension low-pass Butterworth filter locates at the DC component (direct-current component).4) Based on the extraction of edge features and low-frequency information, urban river was extracted. And then, a qualitative assessment of the result of extraction was presented. Firstly, the binarization of extraction result of low frequency information was retrieved by setting appropriate threshold. River channel was intercepted by bridges, thus, an operation of buffering was carried out to maintain the continuity of the urban river in binary image. And area with small size was deleted in binary image. Secondly, in order to decrease the calculation and influence from other land object’s edge features, a function of overlaying was operated to obtain the region for extracting urban river. Thirdly, the edge was determined. Finally, the river information was filled into the boundary of river.After the extraction of urban river, this thesis used the error criterion and the description of precision and recall to complete the evaluation. The result of evaluation certify that the proposed algorithm can satisfied the needing of urban river recognition and extracting of high-resolution remotely sensed imagery.In conclusion, this thesis built the magnitude recognition mark of urban river in frequency domain, by analyzing the magnitude spectrum of high-resolution remotely sensed imagery. Meanwhile, two-dimension band-pass log Butterworth filter was desgined to extract edge features of urban river, and two-dimension low-pass Butterworth filter was designed to extract low frequency information of urban river. Based on these two extraction results, the extraction of urban river is completed. This approach of extracting urban river had an important theoretical value. Further more, this approach presented by this thesis can extend the urban river to all bilinear land objects, so it is significant for features extraction in high-resolution remotely sensed imagery.