| Green space along road of QuickBird images in Nanjing typical area was studied in this paper,which is on the demand of typical ground objects spectrum analysis and feature extraction,funded by National High Technology Research and Development Program(863 Program)"New Technology of High Spatial Resolution Satellite Images Segmentation"(2008AA12Z106)and the National Natural Science Foundation of China "Multiscale Segmentation Methods of High Resolution Satellite Images Based on Spectrum Feature"(40801166).The spectrum energy distribution features of image information were discussed in this paper,the image recognition marks in frequency domain were analyzed,and the suitable filters in frequency domain were designed for information extraction.Methods in this paper could provide references for feature recognition theories of high-resolution remote sensing imageries and information extraction technologies of urban service oriented ground objects structure detection.Green space along urban road is an important part of urban vegetation,which has been the key factor influencing on urban ecological,microclimate,and living comfort of urban dwellers,the information extraction of urban green cover along road is important for urban construction and management.With the improving of time and spatial resolution of remote sensing images and the lasting needs of detailed datas for urban management,remote sensing of urban green has gradually transited from image pixel based image classification to objects oriented feature detection,and new developments are expected in the object feature bansed image recognition and information detection.This paper was to recognize the spectrum features of urban green cover along road,and extraction the regional information by spectrum technologies,which could enrich the physical interpretation knowledge of remote sensing signal characters,and applied new idea for the urban service oriented classification of urban green.Firstly the image features in spatial domain were analyzed in this paper,and the spectral features,texture features,geometric features and contour features were studied.Then the images were Fourier transformed,the magnitude spectrum of green along road was analyzed,the distribution feature at specific directions and frequencies were in depth exploration,the recognition marks in low frequency and recognition methods were built up.Basing on the frequency domain filtering,tonal features in low frequency and contour features in high frequency were acquired.The supplementary information about roads was obtained by recognition theory and extraction technology in frequency domain.In the end the tonal features,contour features and road information were combined,using spatial structure relations,image segmentation were conducted based on multi-feature-marks in frequency domain to complete the extraction of green space along roads.Main research contents and conclusions are as follows:1.Spatial features of green space image along urban road were analyzed for revealing the relationship between image basic spatial feature and spectrum feature Spectral features were analyzed,the spectral reflection characters were compared between road and green space along road in near-infrared band.Reflection coefficient of green plants is high,and it is 50%for platanus orientalis in July,while lower than 30%for roads,which showed that the two objects had great tonal variance,and could be discriminated from low frequency information in spectrum image.Texture features were analyzed,and GLCM features showed the difficulty to distinguish building from road trees.Corresponding to spectrum image,the spectrum curves in high frequency would include complex ground objects ingredient.The features of periodicity and zonality contribute to spectrum image analysis in frequency and direction.Geometric features were analyzed,relationship between road and green space along road is line and surface,and they were separated,which showed that the green space along road can be explored in spectrum images aided by road feature.Contour features were analyzed,the ribbon pattern feature showed that the spectrum directional feature would be obvious,and the closed contour curve indicated the complexity of spectrum feature.2.Recognition marks of green space along road were built up,providing the basis of features detection in frequency domain.Remote sensing imageries were Fourier transformed,in spectrum images,the low-medium frequency is obvious,bright-band appeared along the direction perpendicular to the roads.The angle energy distribution showed the maximal energy directions were perpendicular to the directions of roads,and the radius energy distribution showed that the spectrum energy dropped rapidly and no obvious humorous appeared with the increasing of frequency.The range energy curves were analyzed in sub-sections,and while the period elements near the direct current center,peak value energy frequency existed at the first basic frequency of the spectrum curve,which mean that the harmonic channels including those frequencies symbolized the local highest energy in low frequency channel,and could be the recognition marks for tonal feature of green space along roads.Based on the peak position of the angular energy distribution curve,the spectral directions of image texture were determined.First order derivative of the max angular energy distribution by hatchings of the spectrum images was analyzed for further study of energy distribution in high frequency channel,in which the frequency with peak value was to be the recognition marks.The optimal range frequency bandwidth was decided by the power spectrum density centered with the high frequency marks.The optimal angular frequency was decided by the variance centered with the peak value in the angular energy distribution curve.The feature recognition marks of green space along roads can be the basis of feature detecting filters.3.Based on frequency domain filtering and frequency domain multi-feature marked watershed transforming,information of green space along road was to be extracted.The tonal features were detected by even Gabor filters in frequency domain combined with recognition marks in low frequency;and the contour features were detected by odd Gabor filters in frequency domain combined with recognition marks in high frequency;the roads aided information was detected centered by the frequencies at 80%energy of the image.Incorporating with the geometric relationship between roads and green space along roads,the tonal features in low frequency were labeled,and the contour images of high frequency were watershed transformed to extract the information of green space along roads.The contour images of high frequency were quantitatively evaluated by Canny rules,∑V is 1.00 and C is 0.91.The information of green space along roads were assessed by criterions of P(completeness),R(correctness)and F-score,and P=0.7605,R=0.7639,F=0.7622.The results of green space along roads had complete and clear boundaries,which could meet the need of feature recognition and information extraction of high resolution remote sensing images.Based on the image basic spatial features of urban green space along roads,the spectrum features were recognized in this paper to establishing the tonal features recognition marks in low frequency and contour features recognition methods in high frequency,which provided a physical interpretation method based on frequency compositional variation for ground objects of high-resolution remote sensing signals.Based on the spectrum feature recognition theories,tonal features in low frequency,contour features in high frequency,and roads aided information were detected by frequency domain filters,fusion features of them were used for extracting green space information along roads by multi-feature marks in frequency domain,in which way,the target of service driven typical urban green information extraction was achieved. |
PDF Full Text Request