Method of parameterization and discrimination of spectral profile shape

• Published in: 2013 Second International Conference on Agro-Geoinformatics (Agro-Geoinformatics) pp. 143 - 146
• Main Authors: Quanfang Wang, Yuanyuan Chen, Huaying Wu, Tao He
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-08-2013
• Subjects: matching arithmetic with wildcard strings; MATLAB; parameterization; Sensors; Shape; spectral profile shape; Vectors
• DOI: 10.1109/Argo-Geoinformatics.2013.6621897

Every object has unique spectral profile which often shows different shape from other covers. The information of spectral curve shape could be applied to discriminate objects using remote sensing data. However, it was seldom used in the existing studies on remote sensing image classification, which mainly resulted from the spectral shape not easy to be quantified. In this paper, a parameterization method of spectral profile's shape was proposed. Its core idea is in coding the spectral profile's shape based on symbolic representation. In generally, the spectral profile is composed of extreme values and some segments including ascending segment, descending segment and smooth segment. Therefore, it's able to parameterize the object's spectral profile's shape only using several code numbers. Based on the coding method and using a scene of hyper-spectral AVIRIS image, the spectral profile shapes of some typical land covers were described in Boulder, Colorado, USA. For discriminating each pixel of other land cover, a matching arithmetic with wildcard strings was applied. Finally, a land cover classification map was derived from the AVIRIS image. According to an assessment of the map, the overall accuracy and Kappa coefficient are 83% and 0.788, respectively. The result shows the proposed method is better at land-cover classification using the objects' differences in spectral profile shape. The reason mainly lies in the following two aspects. The first is that the proposed classification method does more emphasis on the basic similarity in spectral profile shape while other remote image classification method generally adopts the similar spectral value. Secondly, the proposed matching arithmetic with wildcard strings allows a little spectral shape difference in the same kind of objects, which solves the within-class dissimilarity problem of the spectral angel mapper (SAM) method and other image interpretation methods, etc. Moreover, the proposed methods of parameterization and discrimination of spectral profile shape not only is suitable to the hyper-spectral images but also can be applied to the objects identified based on multi-temporal remote sensing images, such as time-series VEGETATION INDICE images.