LAUSR

Abstract Geographical object-based image analysis (GEOBIA) is widely used for the processing of fine spatial resolution images, with increased research on contextual modelling and classification related to GEOBIA. Specifically, a previously developed object-based image classification method, known as geostatistically weighted k-NN (gk-NN) method, has shown advantages in increasing classification accuracy. The gk-NN method incorporates spatial weighting into the k-NN classifier through modelling spatial covariance of underlying area classes. However, change-of-support problem (COSP) due to different geometries of image objects is not considered therein. In this paper, we propose a method based on geostatistical de-regularization and regularization for quantifying spatial dependence in area-class occurrences and accounting for scale discrepancy in image objects and pixels. In this new modelling approach, an area-weighted (AW) distance measure is applied for modelling spatial covariance pertaining to sample image objects. The covariance model fitted with image objects sample data is de-regularized to a point-support one, so the spatial covariance over unsampled image objects can then be computed through regularization of a point-support model (RP). Unlike the previous modelling approach in the object-based gk-NN classification, whereby spatial dependence modelling is based on centroids of image objects, this method accounts for change of support and incorporates the geometry of image objects in modelling. The new modelling method was tested on three remote-sensing image subsets with different environments, using regular and irregular segmentation methods at hierarchical scales. It was confirmed that the RP method leads to largely significant increases in classification accuracies (with an average increase of 38.09% in classification accuracy with eighteen cases), compared with that by geostatistical modelling with image object centroids. The proposed method can be used for the modelling of spatial dependence in block-support data which are common in many geospatial applications.