在基于高分辨率遥感影像的农林地块自动/半自动识别应用中，地块边界提取是地块信息精确识别的基础，其精度决定了地块信息识别的精度，这一过程是通过遥感影像多尺度分割过程实现的。在综合比较几种常用的面向对象分割方法效果的基础上，认为均值漂移算法能够更好地满足农林地块识别与边界提取的需求，并可获得较好的地块边界分割效果。进一步面向实际应用问题，本文对算法的实现进行了以下两方面的改进：从实现方式角度考虑，对其在多尺度实现方式方面进行了改进；从算法效率角度考虑，针对农林地块识别过程中数据量大、计算复杂的特点，对其进行了基于MPI（massage passing interface）集群计算环境及OMP(OpenMP)多核任务处理方式的改进，能够显著提高农林地块数据的分割速度。相应地改进算法在实际应用中取得了较好的应用效果。

During the course of high-resolution remotely sensed images information recognition for agriculture and forestry，image segmentation is an important procedure since its results decide directly the effects of land-parcel auto recognition.After analyzing and comparing the segmentation effects of general methods，this paper draws the conclusion that mean shift algorithm can be applied to agriculture and forestry application and has been applied to multi-scale image segmentation because it can extract exactly the edges of the land-parcel.By analyzing the principle of the mean shift image segmentation，this paper improves the algorithm by using two methods，one is to implement the multi-scale merge of the segmentation parcels，and the other is to implement parallel segmentation in message passing interface (MPI) and OpenMP (OMP) model environments.And the applications in agriculture and forestry show that the precision and speed can meet the need of Agroforestry.