Abstract: According to the difference of objects in the WorldView-2 imagery in West Lake District of Hangzhou, sub-regions were divided. Within each partition, different multi-scale segmentation was used and a hierarchical structure was built. To make a comprehensive utilization of spectrum, shape and texture features of variables, the CART (classification and regression trees) decision tree classification algorithm was constructed to select the optimal characteristics and thresholds for each sub-region to map the entire green space of West Lake District. To determine the texture window size and optimize the texture features, the method of J-M (Jeffries-Matusita) distance was used. The results showed that with the method of J-M distance, the texture window size of grassland, agricultural land, shrubs and trees was 5 × 5, 11 × 11, 13 × 13, 13 × 13, respectively. It greatly improved the precision and efficiency of information extraction for the selection of texture window size and dimension of texture features. Comparing with the maximum likelihood method classification based on pixel, the overall accuracy was increased from 76.53% to 88.56%, and the kappa coefficient was improved from 0.711 7 to 0.862 3, the average user accuracy of green space was also increased from 72.73% to 84.63%; Comparing with the conventional object-oriented method, the proposed method is more quickly flexible to determine features and thresholds, greatly improving the efficiency and accuracy of classification.