COMPUTATIONAL STATISTICS &DATAANALYSlS ELSEVIER Computational Statistics & Data Analysis 30 (1999) 443455 Contextual classification in image analysis: an assessment of accuracy of ICM Giuseppe Arbia”, *, Roberto Benedettib, Giuseppe Espa” “ Department of Quantitative Methods and Economic Theory, University of Pescara “G. d’ilnnunzio” Viak Pindaro, 42; 65127 Pescara, Italy “ISTAT, Agricultural Service, Via Rav&, 150; 00142 Roma, Italy ‘Institute of Statistics and Operational Research, University of Trento, Via Inama, 5, 38100 Trento, Italy Received 1 May 1996; received in revised form 1 October 1998; accepted 30 November 1998 Abstract This paper considers the performances of the ICM image classification technique contrasted with the maximum likelihood ordinary discriminant analysis (ML). The latter technique is the most widely used in an applied context by space agencies and remote sensing units. The two methods are compared in terms of the global accuracy produced and in terms of the spatial continuity properties of classification errors. ICM outperforms ML in most experimental cases in terms of the global accuracy produced. However, in some instances, it has a more marked tendency to produce classification errors that are short-distance correlated. @ 1999 Elsevier Science B.V. All rights reserved. Keywords: Autologistic model; Bayesian classification; Global accuracy; Image classification; Markov random fields; Random fields simulation; Spatial pattern of classification errors 1. Introduction Image classification is the stage of image analysis in which the multivariate quanti- tative measurement associated with each pixel (usually expressed in 256-grey levels on two or more optical bands) is translated into a label from a pre-defined set (e.g. land use categories). The current practice in an applied context is that of * Corresponding author. 0167-9473/99/$ - see front matter @ 1999 Elsevier Science B.V. All rights reserved. PII: S 0167-9473(98)00104-2