What Size Window for Image Classification? A Cognitive Perspective

Windows are commonly used in digital image classification studies to define the local information content around a single pixel using a per-pixel classifier. Other studies use windows for characterizing the information content of a region, or group of pixels, in an area-based classifier. Research on identifying window size and shape properties, such as minimum, maximum, or optimum size of a window, is almost exclusively based on the results from automated classifications. Under the notably different hypothesis about optimum sizes of windows in automated classifications and approaches for determining such optimum window size, this article presents a cognitive approach for evaluating the functional relationship between window size and classification accuracy. Using human subjects, a randomized experimental design, and a continuum of window sizes, portions of digital aerial photographs were classified into urban land-use classes. Unlike ihe-findings from purely automated approaches, classification accuracv from visual analvsis increased in a monotonic form with &;easing window &ze for the urban land-use classes investigated. A minimum window size of 40 by 40 pixels (60-m by 60-m ground areal was required for classifying Level II urban land use using 1.5-m by 1.5-m resolution data (2 75 percent accuracy). This finding is dramatically different from the "ideal" window size range (i.e., 3 by 3 to 9 by 9) and functional relation between window size and classification accuracy found in automated per-pixel classifications. A theoretical curve depicting the relationship between classification accuracy and window size, spatial resolution, and classification specificity is presented.