Spatial Accuracy Assessment for Coral Reef Classifications

Abstract. This paper illustrates methods of assessment of a supervised classification using the Maximum Likelihood classifier performed for coral reef classifications that explicitly include the spatial pattern of classification errors, and which presents the user with a visual indication of the reliability of the pixel label assignments. Two of Ikonos images covering two coral reef test sites in Thailand were used in this study. Non-spatial statistics such as: overall accuracy, Kappa coefficient, variance, and Z statistics; were computed from the error matrix and thematic images were also generated. None of these statistics explicitly considers the spatial distribution of misclassified pixels. Results show that regardless of statistical accuracy measurements which present outstanding outcome (values), the accuracy is not spatially similar for every pixel through the image (homogeneous). Thus, a considerable amount of research and development needs to be accomplished before the spatial characterization of thematic accuracy associated with remote sensing products can be adequately reported in standardized format and legends