基于两级选择性聚类集成的农用地整治时空配置分区

针对传统聚类算法易陷入局部最优、不兼顾地理空间的问题，该研究以湖南省怀化市为例，从生态敏感性、用地适宜性与整治迫切性3个维度建立分区评价指标体系，应用一种基于混合距离，同时具有聚类方案质量识别功能的两级选择性聚类集成方法对怀化市农用地整治进行时空配置。通过聚类运算，将怀化市300个聚类单元划分为：近期重点整治区、近期适度整治区、中期重点整治区、中期适度整治区和远期限制整治区5个类型，面积比例分别为9.05%、30.48%、22.58%、7.33%、30.56%，符合怀化现状条件；相较于传统聚类方法，两级选择性聚类集成方法具有聚类质量识别与兼顾地理空间的优势，且更适用聚类单元较多、属性空间复杂的情况。研究结果在客观上为未来土地整治工作与聚类方法的创新提供了可供借鉴的思路。

Spatial-temporal allocation of agricultural land consolidation using two-level selective clustering ensemble

Most case studies have been implemented to explore the agricultural land consolidation in the temporal and spatial dimensions. An automated clustering has also been applied for the agricultural land consolidation zoning in recent years. However, the inherent limitations cannot been overcome in the automated clustering, particularly without considering the geographical space. It is very necessary for the rational zoning of clustering in the time and space dimension, especially for the spatial organization of agricultural land consolidation projects. The traditional clustering can also be tended to fall into the local optimality without the geographic space. In this study, a zoning evaluation index system was established in the Huaihua City, Hunan Province, China, according to three dimensions of the ecological sensitivity, land suitability, and urgency of consolidation. The clustering schemes were then evaluated for the spatial and temporal allocation of agricultural land consolidation. A two-level selective clustering ensemble was selected using mixed distances and three kinds of clustering of the hierarchical clustering, SOFM neural network (Self-Organising Feature Map), and the K-means clustering. The generated solution was identified using NMI (Normalized Mutual Information) and quality index, where the abnormal solutions were rejected to test the similarity of the solutions. Firstly, three algorithms were used to generate a library of solutions. Secondly, several clustering schemes with the better clustering were selected to evaluate the quality index using the NMI. Thirdly, the schemes were selected from a library of solutions for the second level of clustering. Finally, the similar cells of a cluster were treated with the same maximum number of clustering cells as a new one. The process was then repeated until all clusters were included in the new clusters. As such, the clusters were merged with the greater similarity, where the smaller clusters included in the clusters to reduce the dispersion between clusters. More importantly, the geospatial information was considered to avoid the application of low-quality clustering schemes. Correspondingly, the 300 clustering units were classified into five types: short-term focus consolidation areas, short-term mild consolidation areas, medium-term focus areas, medium-term mild consolidation areas, and long-term restricted consolidation areas, with the area proportions of 9.05%, 30.48%, 22.58%, 7.33%, and 30.56%, respectively. The data was in the line with the current conditions in the study areas. The content and quality of different solutions varied significantly, which were required the optimization and integration of solutions for the spatiotemporal configurations. Consequently, the two-level selective clustering ensemble presented the higher quality of cluster identification, especially considering the geographic space, compared with the traditional clustering. The evaluation system can also be widely expected to treat many clustering units and the complex attribute space. Overall, the finding was enriched the evaluation system for the spatial and temporal configuration of agricultural land consolidation, providing a promising idea for the future innovation of land consolidation and clustering.