The present inventions generally relate to a leaf node ranking method in decision trees for spatial prediction and its recording medium. The leaf node ranking method in decision trees includes a learning step to form a decision tree having one root node, in which each parent node has multiple child nodes, using training data sets for spatial prediction; and a leaf node ranking step from the decision tree that finishes the learning. In the learning step, each node of the decision tree stores both the number of classes according to class distribution of training data and structures for storing the number. In the leaf node ranking step, a rank of a leaf node is determined using the number of classes according to class distribution, which is stored in each node on a path from the root node to the leaf node.