Objective  Sampling density is closely related to the spatial prediction accuracy of soil organic matter (SOM) in cultivated land. Determining the reasonable sampling density of SOM spatial prediction in cultivated land is beneficial to scientifically guide soil sampling and save work cost.

  Method  In this study, taking Yueyang County of Hunan Province as an example, the conditional Latin Hypercube Sampling (cLHS) scheme was designed with R language. Eight training sets with different sampling densities (10.01, 7.41, 3.70, 1.85, 0.93, 0.46, 0.23 and 0.12 samples per km²) were independently extracted from 7399 cultivated land soil samples (the sampling density was 14.82 samples per km²). In order to take into account the sample feature space and geographical space, the topographic position, slope, soil parent material, soil type, township, longitude and latitude and other information were added to the cLHS. Combined with ordinary Kriging method, the spatial prediction effect of SOM in cultivated land with different sampling densities was analyzed and discussed.

  Result  The results showed that the average of SOM in the training set with different sampling densities was higher than the average level of Hunan Province. Each training set had strong representativeness on the whole. Semi-variance function models were exponential models, with good semi-variance structure (structural ratio: 87.5%-94.5%) and strong spatial correlation. The range was positively correlated with the goodness of fit (correlation coefficient r = 0.96), and negatively correlated with the structural ratio (correlation coefficient r = −0.79). When the sampling density was 3.70 samples per km², the structural components in the detected variation structure of SOM was the most complete, and the accuracy was the best. Further increasing the sampling density was not necessarily conducive to the identification of structural continuous components. When the sampling density was lower than 0.46 samples per km², the semi-variance function model could not be effectively inferred, and the accuracy was poor. When the sampling density reached more than 1.85 samples per km², the spatial structure characteristics could be revealed more robustly. Continuously increasing sampling density would not significantly improve prediction accuracy.

  Conclusion  Considering the requirements of prediction accuracy and working cost, in areas similar to the natural and geographical conditions of the study area, the expected effect can be obtained by controlling the soil sampling density of cultivated land above 1.85 samples per km².