Comparison of a decision tree and maximum likelihood classifiers: application to SAR image of tropical forest

The authors compare the use of a decision tree (DT) and the maximum likelihood classifier (MLE). The classifiers are applied to the Global Rain Forest Monitoring (GRFM) mosaic data. Two mosaics are available and were acquired in February and October-November 1996 (high and low water of the Congo basin). The GRFM mosaics are composed of JERS-1 images and cover the entire Central Africa at a resolution of 100 m. The mosaics have 59 equivalent number oflooks, thus spatial intensity variation are mainly due to forest intrinsic texture. The authors focus the comparison on a small area of Eastern Cameroon around the city of Bertoua. The area is characterized by a network of gallery forest in the savanna area, degraded forest along the road network and open forest around the city of Bertoua. The open forest has a different tone in the low and high water mosaics which can be attributed to temporal changes in the understory. The authors selected 7 classes for comparison: Grass Savanna, urban, Open Forest, Degraded Forest, Woody Savanna, Flooded Forest, Closed canopy Forest. The Decision Tree method is based on L. Breiman et al. (1984). It is a set of hierarchical rules which are applied on the data to purify a group of samples. The rules are the nodes of the Decision Tree which split the input data group into purer groups