An algorithm for spatially constrained classification of categorical and continuous soil properties

The creation of spatially contiguous clusters from multiple sources of information is an important objective in many scientific disciplines. We describe a supervised spatial classification algorithm in which the primary goal is to stratify a landscape or other spatial data into a set of discrete patches based on available categorical and continuous data. The proposed method consists of four major steps: (1) data pre-processing, (2) initial allocation through spatially constrained cluster analysis, (3) spatial aggregation and peripheral re-allocation of individuals, and (4) an optional hierarchical or non-hierarchical classification to summarize individual patches in few classes. Parameters needed are the number of spatial clusters, the minimum size of a cluster and the weight assigned to the categorical variable(s). The proposed method was sensitive to the spatial association between continuous and categorical variables as well as to the spatial correlation structure of the continuous variables. Both determine the optimum weight assigned to the categorical information and, therefore, the effectiveness of the classification. The proposed clustering method was robust in terms of finding natural clusters, allows using categorical and continuous variables, and offers a reasonable compromise between geographic contiguity and class homogeneity. Possibilities for further modifications are outlined.