Identification of representative samples from existing samples for digital soil mapping

Existing sample data are important for digital soil mapping. Different sample points possess different representativeness. The representativeness of samples influences the soil mapping result greatly. However, few study focus on assessing the representativeness of single sample. In this paper, we proposed a method to identify representative samples from existing samples collected from multiple resources. The basic idea of the method was to use clusters of environmental covariates to approximate types of soil variations, and check the occupancy of the existing samples in centroids of environmental clusters. Those samples locating in typical locations or centroids of environmental clusters were considered as representative samples. In this paper, the proposed method was used to discern representative samples in 282 soil samples in Anhui Province, China. SOM content was mapped using a similarity based mapping method. Two cases with different training samples (representative samples, non-representative samples, and training samples including representative and non-representative samples) and validation samples were set to compare the mapping results and accuracies. The results showed that the SOM content maps predicted using representative training samples had generally higher accuracy than the results produced using non-representative samples, and comparative accuracies with the results produced using full training samples. To discern representative samples is helpful for understanding the soil-landscape relationships in an area and the proposed method can be used to design supplementary samples for a better soil mapping result. Mapping results and accuracies showed that different training and validation sample sets impacted the mapping results and accuracies greatly, which indicates that researchers should be cautious when using randomization to obtain training and validation subsets for soil mapping.