A recurrent curve matching classification method integrating within-object spectral variability and between-object spatial association

Abstract

Object-based image analysis (OBIA), which has been commonly used for land cover and land use classification, may encounter challenges when satellite images’ spatial resolution achieves at the sub-meter level. An image object may exhibit spectral heterogeneity, causing traditional object-level statistical measures such as mean values of the pixels in an object not suited to represent the feature of the object. Additionally, an image object may have strong spatial association with its surroundings. Traditional OBIA only considers spatial features of individual object, but ignoring spatial arrangement or spatial association between objects. This paper proposes a new OBIA method by integrating within-object spectral variability and between-object spatial association. The within-object spectral variability is captured by the histograms of the pixels in an object across multispectral bands to reflect the heterogeneity of their pixel values. Based on this, the initial classification result is obtained using non-parametric curve matching methods. Then, the between-object spatial association is represented by curves derived from the frequency of pairwise classes in four main directions, also in the form of curves. The curves now composed of both the histograms of spectral feature and the class pair frequency of spatial feature are then fused for another curve matching based classification. This recurrent process is repeated and the spatial association is recaptured from the previous classification result at each iteration until a stopping criterion is satisfied. The curve matching classification method based on histograms of spectral feature is superior to traditional OBIA based on only object-level statistical measures since it fully characterizes spectral variability in the objects. The between-object spatial association works as a spatial filter that considers spatial arrangement of classes in a neighborhood. The developed method is especially suitable for classifying high spatial resolution (HSR) images with land cover/land use classes in typical urban areas.