耦合SOFM与SVM的生态功能分区方法——以鄂尔多斯市为例

辨析区域主要生态环境问题及其空间异质性,划定生态功能区,对保障区域生态安全和国土开发优化具有重要指导意义。以往自下而上的分区研究多基于行政区或流域开展,难以体现行政区或流域内部的生态功能分异。以鄂尔多斯市为例,基于生态系统服务与生态敏感性构建区域生态功能分区指标体系,耦合自组织特征映射（SOFM）网络与支持向量机（SVM）划定鄂尔多斯市生态功能分区。结果表明,区域内各生态功能分区指标呈现明显的空间分异特征,通过SOFM网络基于栅格进行指标聚类,构建分类效果指数筛选最佳聚类方案,将区域分为7种不同的生态功能类型。最终,利用SVM识别最优分区界线,将鄂尔多斯市分为11个生态功能区。本文构建分类效果指数实现多分类方案优选,使用机器学习算法解决自下而上的自然分区容易弱化要素空间位置属性的问题,完成了从分类到分区的定量转换,有助于提升分区的空间精度与客观性,为生态功能分区与分区边界划定提供了新的方法途径。

An ecological function zoning approach coupling SOFM and SVM: A case study in Ordos

It is of great significance to analyze the main environmental issues, coupled with their spatial heterogeneity, and divide the ecological function zones to achieve ecological security and optimization of territory development in a certain region. In recent years, ecological function zoning, widely concerned by scholars, has played a vital role in regional ecosystem management and sustainable development. There arose a problem that the spatial characteristics of ecological functions were hard to be reflected in the previous studies based on spatial average data over basins or geopolitical regions such as counties, cities and provinces. This paper, using the approaches of coupling self-organizing feature map (SOFM) and support vector machine (SVM), attempts to develop an automatic demarcation and zoning approach, and explores the best possible division of ecological functions in the city of Ordos with the index system of ecological function zoning in mind involving the ecosystem services as well as ecological sensitivity. Ordos, which is located in the transitional zone between temperate grasslands and desserts, has become a key research area of global terrestrial ecosystem. The ecological functional zoning indexes indicate that there appears an obvious spatial heterogeneity of ecosystem functions in Ordos. Accordingly, land grids have been clustered into 7 ecological function types by SOFM with clustering quality index (CQI) in view. Hence, Ordos has been divided into 11 ecological function zones owing to the implementation of SVM to identify the optimal division borders, in which the border demarcation is treated as a classification system in spatial domain. The optimal combination of SVM hyperparameters is determined by grid search method. In this study, machine learning algorithm has been adopted to cope with the situation where the bottom-up physical regionalization might weaken the spatial position attribute of the partition features, and to realize the quantitative conversion from classification to partition. It has turned out that such SOFM-SVM-coupled zoning approach could effectively improve the spatial accuracy of the partition, which can be considered as a new way to realize the automatic ecological zoning.