Typical Feature Recognition Of High-Resolution Remotely Sensed Imagery By Using Frequency Spectrum Energy

Feature recognition and extraction is an intermediate link for remote sensing image processing, which is an important guarantee of remote sensing application and services aspects. Recent years, with the improvement of spatial resolution of satellite sensors, high resolution remote sensing image has become one of the most important data sources for feature recognition and information extraction. Due to the more structure, texture, shape and edge information were provided in the high resolution remote sensing image, the traditional recognition methods which based on pixel spectral energy could not satisfy current remote sensing application requirements, and it became a bottleneck restricting the sustainable development of large-scale remote sensing application. For this reason, how to change’pixel-based’methodology into’feature-based’methodology to identify and extract information from such high resolution imagery has become a most active research topic in remote sensing image processing field.Under the auspices of national high technology research and development program (’Study on new segmentation technology for high spatial remote sensing imagery’,2008AA12Z106) and national natural science foundation of China (’Study on multi-scale segmentation method of high spatial remote sensing imagery based on frequency-domain feature’,40801166), feature extraction and image identification from QuickBird images within the experimental area were researched on the basis of ’Parseval energy conservation theorem’and frequency spectrum energy analysis techniques. It laid a theoretical foundation for turning’pixel-based’methodology into ’feature-based’methodology to identify and extract information from high resolution imagery. The main contents and conclusions of this research were presented as follows:1) According to’Parseval energy conservation theorem’and’Theory of Tupu’, the feasibility of remotely sensed imagery recognition from’pixel-baesd spectral energy’to’feature-based spectrum energy’was demonstrated in this paper. A certain signal of a specified object feature can be seen as overlay of a number of different harmonic energy waves, which could reconstruct the original signal by overlaying either in the spatial domain or in the frequency domain. The difference was that in the spatial domain the energy is based on the position (x, y) of pixel, while in the frequency domain that is based on the position (u, v) of frequency. Therefore, in addition to considering the spectral characteristics of the energy, we can also start with the frequency spectral chracteristics of the energy, and then analyze the spatial frequency of different features, such as color tone, texture, edge, direction features and so on, making full use of the features’energy information in different positions of the harmonic wave to better analyze, interpret the target objects.2) The spectral energy distribution and the information characterization of the features on the high-resolution remotely sensed imagery was analyzed and discussed. By the percentages the spectral energy which contained by the circles of different radius taking of the total energy, the spectral energy distribution was preliminarily investigated. As a result, about90%energy was around the spectrum center, and this part of energy played an important role in reconstructing the overall contour and color characteristics of imagery. As the spectral radius increased, high-frequency energy was gradually incorporated. Meanwhile, structural features of the image edge information had become increasingly evident. Then by the wedge-shaped and ring-shaped energy spectrum sampling methods, the quantitative analysis of the coarse-grained level, cyclical and directional information of different surface features in the study area was done. The studies showed that the analysis of energy spectrum for the surface features in remotely sensed imagery, the surface feature information representation which cannot be gotten in the spatial domain could be obtained better. This provided a new and effective idea to further identify and extract information from high resolution imagery.3) On the basis of low-frequency spectrum coefficients, recognition methods and recognition marks of target objects’main color tone feature were researched and discussed. As an important feature of remote sensing images, frequency spectrum is a redistributed spectrum of pixel spectral energy in frequency domain. According to low-frequency spectrum coefficients, the different target objects in experimental area were recognized effectively by using SVM (support vector machine) method, and the overall recognition rate was up to88.96%. Taking residential buildings for example, a cross-hatching analysis method was presented to acquire recognition mark of main color tone feature. Through analyzing the overlay process of low-frequency spectrum coefficients, the main fluctuation characteristics of pixel brightness response curve of residential buildings were discussed. Meanwhile, applying variance contribution of harmonic energy and statistical test method, the center frequency of residential buildings’main color tone feature was acquired as0.023cycles/pixel.4) On the basis of high-frequency spectrum coefficients, recognition marks of target objects’local edge and detail feature were discussed. Combining the overlay analysis of high-frequency harmonic energy, the statistical and physical explanation of minor fluctuation characteristics of brightness response curve was interpreted in this paper. The fluctuation characteristics of curve in spatial domain, to some extent, could be refected as the magnitude of different harmonic energy in frequency domain. If the response curve fluctuates violently, then the brightness value has a larger variance, and the high-frequency spectrum coefficients will play a greater role in the process of spectrum energy’s overlay. By analyzing the approximation between the reconstructed curve which overlayed by different harmonic energy and original brightness response curve, the center frequency of residential buildings’ local edge and detail feature were acquired as0.093,0.153,0.197,0.260and0.328cycles/pixel. Finally, joining these five high-frequency value together the foregoing low-frequency value, the reconstructed result was approximate to brightness curve, and the approximation degree is0.976. It indicated that the established recognition marks were feasible.Based on low-frequency recognition marks and high-frequency recognition marks, the reconstruction results of residential buildings’ feature information were analyzed, and the building objects were extracted with the matched Gabor filters which have direction and frequency selectivity. The result of accuracy evaluation showed that the approach presented in this paper basically satisfied the demand of feature recognition and extraction for high-resolution remote sensing image. This research has important theoretical significance and practical value. First, the spectrum energy analysis based on Fourier transform provided a brand-new methodology for the transition of high-resolution remote sensing image classification from traditional classification in pixels to feature recognition. Second, profound analysis of the relationship between the descriptions of target feature in spatial domain and frequency domain has important theoretical value for the discovery of the feature recognition’s internal mechanism. Meanwhile, the cross-hatching analysis method of target feature, the overlay analysis method of profile signal’s harmonic energy and the recognition marks of low-frequency and high-frequency of target feature proposed in the paper offered important references for the development of methodology and application.