Abstract: [Objective] This paper studies the method of determining the segmentation parameters for multiresolution segmentation algorithm in the process of object-oriented tree species classification. This paper seeks to replace the traditional optimal segmentation scale evaluation method that relies on the reference polygons, and quantitatively evaluates the suitability of the optimal segmentation scales proposed by the ESP2 tool for tree species classification.[Method] The segmentation experiment of object-oriented classification is carried out, taking the Huapiqiangzi forest farm in Yichun city, Heilongjiang Province as the study area, with the GF-2 remote sensing image used as the experimental data. Based on the idea of local variance(LV)of object heterogeneity within a scene reflects the optimal segmentation scale, scales are found corresponding to obvious peaks of the homogenous local variance variation rate in the specific scale range(100-400, step size is 1)generated by ESP2, which is defined as the optimal segmentation scale range. Finally, the multiresolution segmentation at each scale in the optimal segmentation scale range is performed with the optimal composition of homogeneity criterion parameters. The distribution of the tree species sample points in each segmentation results is counted, and the segmentation time is recorded. The optimal segmentation scale is determined by comparing the ratio of the sample points of tree species to the correct distribution and the segmentation time.[Result] The segmentation experiment under the same scale parameter of multiresolution segmentation algorithm shows that when the composition of homogeneity criterion parameters are shape=0.5 and compactness=0.3, the segmentation result is relatively the best. The segmentation evaluation method based on tree species sample point pairs shows that among all the segmentation results of the experiment, the scale parameter corresponding to the largest ratio of the sample points of tree species to the correct distribution is 259. Totally, 203 of 210 pairs adjacent tree species sample points fall into the adjacent segmentation objects. The vector distance index and the ED3_modified in the optimal segmentation scale range are calculated. The results show that the evaluation results are consistent with the evaluation results based on the pair of tree species. The vector distance index and the ED3_modified of segmentation result for each scale in the optimal segmentation scale range are calculated. The results shows that the evaluation result are consistent with the evaluation result based on the tree species sample points.[Conclusion] The influences of different composition of homogeneity criterion parameters on the segmentation results are significantly different. It is necessary to design an efficient experimental scheme to find this combination.The evaluation method based on tree species samples points makes full use of the tree species survey data, and simplifies the reference polygon samples commonly used in previous evaluations into point samples, which avoids the complicated workload of manually delineating real object boundaries.Compared with the object matching method or the similarity principle of the area principle,the point index of optimal segmentation scale evaluation method considering the segmentation efficiency could improve the comprehensiveness of the segmentation factors.

Key words: optimal segmentation parameters, object-oriented classification, tree species classification, GF-2