Significance-weighted classification by triplet tree

An efficient classification method using a triplet tree is proposed for target land-cover categories with significance weight. The weights are determined by user in the view of importance in actual classification. In the proposed method, a triplet tree classifier for land cover classification is used. The triplet tree classifier has two types of nodes. It generates two nodes for `definite nodes´ and one optional `indefinite node´ at every node segmentation. Tree design procedure uses the weights in the two meanings. Firstly, significant categories are assigned with high priority in the selection of splitting patterns. Categories with higher priority are separated from other categories at the upper nodes. Secondly, a node for heavily weighted categories are designated with little classification error at every decision of boundaries. Experiment about real remotely sensed images was executed to show the performance of the proposed method. The results of classification were compared with the standard Bayesian classifier or other multistep methods. The classification accuracy about heavy weighted categories by this method is higher than a conventional classifier without weights. The computing cost for this method is small because this approach is based on a decision tree method