Soil is characterized by high spatiotemporal variability due to the combined influence of internal and external factors. The most efficient approach for addressing spatial variability is the use of management zones (MZs). Common approaches for delineating MZs include K-means and fuzzy C-means cluster analysis algorithms. However, these clustering methods have been used to delineate MZs independent of the spatial dependence of soil variables. Thus, the accuracy of the clustering results has been limited. In this study, six soil variables (soil pH, total nitrogen, organic matter, available phosphorus, available potassium, and soil apparent electrical conductivity) were used to characterize the spatial variability within a representative village in Suining County, Jiangsu Province, China. Two variable reduction techniques (PCA, multivariate spatial analysis based on Moran's index; MULTISPATI-PCA) and three different clustering algorithms (fuzzy C-means clustering, iterative self-organizing data analysis techniques algorithm, and Gaussian mixture model; GMM) were used to optimize the MZ delineation. Different clustering model composites were evaluated using yield data collected after the wheat harvest in 2020. The results indicated that the variable reduction technologies in conjunction with clustering algorithms provided better performance in MZ delineation, with average silhouette coefficient (ASC) and variance reduction (VR) of 0.48-0.57, and 13.35-23.13%, respectively. Moreover, the MULTISPATI-PCA approach was more conducive to identifying variables requiring MZ delineation than traditional PCA methods. Combining MULTISPATI-PCA and the GMM algorithm yielded the greatest VR and ASC values in this study. These results can guide the optimization of MZ delineation in intensive agricultural systems, thus enabling more precise nutrient management.
Keywords: Gaussian mixture model; MULTISPATI-PCA; clustering model composites; management zone; soil variable.