Integrating spectral variability and spatial distribution for object-based image analysis using curve matching approaches

Abstract

Object-based image analysis (OBIA) has been widely used to classify high spatial resolution (HSR) imagery. In a traditional OBIA, object-level statistical summaries such as mean values are usually used for classification. This implies that the spectral values within objects follow a Gaussian distribution. However, the pixel values in an object do not necessarily conform to a Gaussian distribution because of within object spectral heterogeneity. Consequently, these statistical summaries may misrepresent the features of the object. This shortcoming is addressed in this paper by integrating both the spectral variability and the spatial distribution of the pixels within objects to improve the traditional object-based image classification. The spectral variability is represented by histograms of the pixel values in the object, and the spatial distribution is characterized by the binary spatial covariogram of these pixels. To construct a binary spatial covariogram, a principal component analysis (PCA) is first applied to compress multiple bands into one, and the Otsu thresholding is then performed to generate a binary map reflecting the spatial configuration of the pixels. Spatial covariance is then computed for this binary map and plotted with different lag distances to derive the binary spatial covariogram. Our proposed model utilizing curves composed of the spectral histograms and binary spatial covariogram (referred to as the His-Cov model) are then used for classification based on curve matching approaches. The integration of spectral variability and spatial distribution of the pixels in the object produced superior results to curve matching approaches based on spectral variability alone and to traditional OBIA based on spectral and spatial features of the objects when classifying complex land use types in urban environments.