预测性土壤有机质制图中模糊聚类参数的优选

Optimization of clustering parameters in predictive mapping of soil organic matter

对数字高程模型(DEM)中的地形特征进行模糊c均值聚类,分别采用3种方法来选择最优模糊度和分类数组合及分类结果;在最优分类结果上,用多元线性回归方法建立土壤A层有机质含量与地形景观之间的定量关系,并应用该关系进行土壤制图应用。结果表明:3种方法选择的最优模糊度比较接近,主要为1.5,还包括1.4和1.6,但3种方法选择的最优分类数却有很大差别;尽管依据回归模型r2选择的分类结果较多地解释了土壤A层有机质含量的变异,但基于这种分类结果的制图偏差较大,与实测值相比较的结果也说明基于这种分类结果的制图精度较低;用内部判据选择的分类结果在制图过程中产生的偏差较小,制图精度也较高。

After fuzzy c-means clustering of topographic attributes derived from a digital elevation model （DEM） of the study area, three measures had been used to identify optimal combinations of clustering parameters, i.e. clustering exponent and number of classes, and its corresponding clustering results. Based on the identified optimal clustering results, soil-landscape models were built through multiple linear regressing the relationships between the content of organic matter in soil A layer and clustering memberships. Furthermore, the soil-landscape models were applied into predictive mapping of organic matter in soil A layer over the study area. Results show that, the three measures have different powers in choosing optimal number of clusters while they chose the similar optimal fuzziness exponents, 1.5 and/or near to 1.5, i.e. 1.4 and 1.6. And the predicted maps that were based on the optimal clustering identified by regression r^2, show greater deviations from normal than the other two maps. Moreover, comparisons of values on predicted maps and determined in lab indicated that greater accuracies existed in the predictive maps which were based on the fuzzy clustering identified by internal criterion.