A novel quality assessment index for evaluating the spatial uniformity of remote sensing image classification

Accuracy assessment techniques have always been considered essential to evaluate the quality and reliability of image classifications. They are commonly based on metrics extracted from the confusion matrix. Application fields whose source data are organized in a two-dimensional disposition, such as remote sensing imagery, generally present their classification outcomes in the same format. However, the spatial organization of the class objects is typically neglected for accuracy evaluation. We propose an additional quality index for measuring the uniformity ratio (UR) of remote sensing image classifications. When visually examining image classification outcomes, we tend to give lower reliability to results that present higher non-uniformity of class objects, frequently known as classification noise. The basic hypothesis of this paper is that classification models that better delineates the class signature regions, in the available feature space, produce less spatial variation of class assignments to the mapping units (normally pixels). The experimental results indicate that the proposed index can be efficiently incorporated into remote sensing applications to complement standard accuracy metrics.