基于K-均值算法模型的区域土壤数值化分类及预测制图

根据封丘县土壤发生学特点遴选质地、有机质、土壤颜色、pH值、电导率和土壤发生层厚度等作为土壤属性向量，运用k-均值算法模型，对研究区40个土壤剖面样本实施数值化分类，并依据《中国土壤系统分类检索》确定算法模型输出的5个中心土壤剖面的系统分类归属。基于40个样本土壤剖面剖面与各中心剖面类型之间的分类距离，应用地统计学手段预测研究区空间任意位置的土壤与各中心剖面分类距离，完成研究区土壤数值化连续分类并在空间上实现可视化表达。在此基础上，运用去模糊化手段，“硬化”连续分类边界，获得可与传统土壤制图互为参比的研究区土壤预测制图，并对输出结果进行了土壤发生学解释。本研究表明，土壤数值化分类手段与地统计学随机模型相结合可以实现区域土壤的空间预测，且预测制图比传统土壤制图蕴含更加丰富的信息。

Numerical soil classification using fuzzy K-means algorithm and predictive soil mapping at regional scale

In the studied region, choosing texture, organic matter content, soil color, pH, electronic conductivity and soil layer thickness as soil variables, numerical soil classification based on 40 soil profiles was conducted by the use of fuzzy k-means algorithm, and the resulting central profiles soil 01, 18, 37, 38 and 40 were allocated into a hierarchical classification according to Key to Chinese Soil Taxonomy. They belonged to Pandian Series, Haplic Uap Ustic Cambisol, Yingju Series, Haplic Uap Ustic Cambisol, Haplic Endorusti-Ustic Cambisol, Haplic Warpic Anthric Entisol and Salinic Warpic Ustic Cambisol respectively. On the basis of the known taxonomic distances (Here,Euclidean distance was employed) amongst the sampled soils and the above central profiles, the taxonomic distance between unknown soils at any sites and the central profiles was figured out using Geostatistic techniques. In this way, continuous soil classification of the studied region was conducted through a predicted relationship of taxonomic distances between soils. For a better visualization, the soft borders between soils from different fuzzy classes were hardened by means of defuzzification defined through the distance thresholds. And the output of predictive soil mapping thus had a same visual appearance with a conventional soil map. From the predictive mapping, it could be clearly seen that the soil cover in the studied region was dominated by Haplic Endorusti-Ustic Cambisols and Salinic Warpic Ustic Cambisols, which accounts for a large percentage of the region's total area (46.3% and 33.2% respectively). By contraries, Haplic Uap Ustic Cambisols and Haplic Warpic Anthric Entisols were discontinuously distributed and totally occupied only around 10% of the total land area of the studied region. It could be concluded that integration of numerical soil classification and geostatistic interpolation would be one of the most recommendable approaches for predictive soil mapping.