www.elsevier.com/locate/rse A Comparison of Parametric Classification Procedures of Remotely Sensed Data Applied on Different Landscape Units L. Hubert-Moy,* A. Cotonnec,* L. Le Du,* A. Chardin, † and P. Perez † This paper presents an evaluation of several parametric results depends on the selected classification method. classification algorithms to assess their accuracy on vari- Generally, areas with homogeneous spectral responses ous landscapes. Traditionally the maximum likelihood within a field are modeled through multivariate normal classifier is used to obtain thematic maps in land use. In distribution and the assignment of pixels to classes is of- this work different classification algorithms including ten based on the maximum likelihood technique (Rich- contextual classifiers, one of them being original, are ap- ards, 1993). Sometimes contextual classifications that con- plied and compared on sites belonging to landscape units sider the spatial context of a pixel in the image are applied ranging from tiny fields surrounded by hedges to larger on remotely sensed data when a large variety of spectral and more open fields. Confusion matrices and result responses are observed in a same field. Several studies analysis are presented at two observation scales: at the have pointed out the superiority of such contextual tech- catchment area level and at the landscape unit level. We niques as compared to classification techniques that assign show how the choice of a classification technique can sig- a pixel to a class independent of the labeling of neigh- nificantly influence the results of crop inventories and boring pixels (Sharma and Sarkar, 1998). These studies, how the accuracy of classification algorithms vary ac- however, perform comparisons on subscenes that do not cording to the landscape units of the studied area. From exhibit much diversity in their landscape structure (often these results a strategy can be developed for a better composed of large open fields) (Datcu et al., 1998). choice of classification algorithms regarding the consid- In theory, the choice of a classification method should ered landscape structure. Elsevier Science Inc., 2001. be done according to landscape structures, but in prac- All Rights Reserved. tice analysts often apply the same classification algorithm to various areas without considering landscape features. Thus the methods that give good results on large size INTRODUCTION fields with homogeneous spectral responses are also ap- Many researchers have used remotely sensed data to as- plied on areas that include tiny fields with a higher num- sess land cover and land use changes. Crop inventories are ber of mixed pixels. When selecting a classification generally carried out on areas that correspond either to method, analysts should be aware that the accuracy administrative units ranging from local to regional areas greatly depends on the landscape structure of the stud- or to watersheds when the relationship between crops and ied area. Moreover, the same administrative unit or wa- water quality are concerned. The accuracy of mapping tershed can include more than one landscape unit. Therefore, the studied area should be split out in several subunits before applying the most adapted classification * Costel, UMR 6554, Universite ´ de Rennes 2, France † IRISA/INRIA-Rennes, France technique on each of them, or at least the classification Address correspondence to L. Hubert-Moy, De ´ partement de algorithm that gives the best results on the entire studied Ge ´ ographie, Universite ´ de Rennes 2, 6 ave Gaston Berger, 35043 Re- area should be chosen. nnes Cedex, France. E-mail: laurence.hubert@uhb.fr Received 24 January 2000; revised 23 June 2000. Whereas the exclusive choice of supervised classifi- REMOTE SENS. ENVIRON. 75:174–187 (2001) Elsevier Science Inc., 2001. All Rights Reserved. 0034-4257/00/$–see front matter 655 Avenue of the Americas, New York, NY 10010 PII S0034-4257(00)00165-6