区域尺度农业管理分区的无监督特征选择与破碎度优化算法

Unsupervised feature selection and fragmentation optimization of agriculture management zones at a regional scale

针对区域尺度管理分区指标筛选与分区破碎问题,提出基于指标相关性聚类的无监督过滤式指标选择方法 FSCC(feature selection based on correlation clustering algorithm,FSCC)与基于一致性和完整性的指标优化方法(consistency and integrityoptimization,CIO)。以中国主要冬小麦种植区为研究区域,气象、土壤、地形等小麦生长相关指标为数据源,研究区域从大到小划分为4个尺度,首先选用最大方差、拉普拉斯得分2种传统过滤式特征选择方法与FSCC分别进行4个尺度的管理分区指标筛选,对比基于3种方法筛选指标集构建的管理分区划分结果,评价FSCC分区指标选择方法;其次,设计指标优化算法,对4个尺度筛选的指标集分别进行一致性与完整性分析与优化。结果表明:相较最大方差法和拉普拉斯得分法,FSCC筛选指标的分区效果具有较好表现,如皋2.5km处,其评价指标模糊性能指数(FPI)、归一化分类熵(NCE)和修正分离熵(MPE)均低于另外2种方法 52.44%、49.45%和49.52%;CIO在如皋、南通尺度下有效剔除分区破碎指标,分区完整性明显,除南通10 km外,CIO比FSCC的指标集,FPI、NCE、MPE分别平均低0.078、0.061、0.082,相对提升了FSCC的分区效果。

Dividing farmland into different zones for facilitating management(management zone) at regional scalescan help improve agricultural production in reforming agricultural technology implementation in China. Improving detailed prescriptionof the management zone division can provide guidance to farming and service optimization at regional scale.Appropriately selecting indexes in management zones can reduce the required data and can thus subsequentlyimprove management.Available index selection usually relies on empirical knowledge of expertsand/or multivariate statistical analysis. However, expert evaluation method could be bias, while the multivariate statistical analysis methodcannot reduce the number of indexes compared to the original index set and thus need to supervisethe data. In addition, most existing work on fragmentation of management zones focused on zone-dividing method rather than from index selection by removing indexes that lead to fragmentation. This paper aims to resolve these limitations with a proposedunsupervised filtering index selection method, based on the index correlation clustering(FSCC) using the concept of feature selection. FSCC reduces the original index set to obtain a subset called FSCC set. FSCC applies the correlation matrix of all indexes to cluster the original indexes set. It then selectsall cluster centers as a representatives to form a new index subset as theFSCC set. The quantity of the indexes in the FSCC setwas reduced,compared to the original index set, and the redundancy of the indices set was descended. To improve practical operability of the management zones, we applied the index optimization algorithm developed based on the consistency and integrity(CIO) to the FSCC set to remove indices which resulted in fragmentation. CIO couples Kappa Coefficient with fragmentation index to generate an optimization strategy for the FSCC sets. CIO screens the indices which lead to the fragmentationwhile, in the meantime, considering the consistency of the management zone results prior to and after the optimization. We applied the method to winter wheat in China, with factors that affect wheat growth, including meteorology, soil and topography, being dividedat fourregional scales. We first usedthe FSCC and the two traditional filter feature selection methods, Variance and Laplacian Score, to select index subsets for thefour scales, and compared the resultant management zones produced from them. The CIO was then applied to the four scales produced by the FSCC. The results showed that the FSCC method preserves the diversity of the features in the original index set. It significantly removed the redundant indices and had a better performance in the management zones. The best performance shows that in Rugao 2.5 km Grain, FSCC less than 52.44%, 49.52%, 49.45% both of Variance and Laplacian Score in FPI, MPE, NCE. The CIO improved the management zones effect of the FSCC index set, which reduced the number of indexes and effectively removed the indexes that led to anumber of isolated units or patches. Compare to FSCC, except Nantong 10 km, CIO has an average decrease in FPI, MPE, NCE of 0.061, 0.078, 0.082. Usingthe fourregional scales, FSCC and CIO presented in this paper were effectivein selecting indices and havepotentialapplication in management zone division.