Optimal Sample Size for SOC Content Prediction for Mapping Using the Random Forest in Cropland in Northern Jiangsu, China.

A soil organic carbon (SOC) map of high accuracy is the basis for taking mitigation measures against crises of food security and global climate change. Predicting SOC based on a limited number of soil samples can reduce the cost and time for laboratory analysis. This study aimed to assess the influence of sample size on the prediction of SOC and to identify the optimal sample size of SOC prediction for cropland in northern Jiangsu, China. A total of 1182 soil samples were randomly split into calibration and validation sets. Ten calibration subsets of samples between 108 and 1064 were selected by using a parent material-based stratified random resampling strategy. The random forest algorithm was used to develop 10 calibration models validated based on the same validation sample set. These 10 models were evaluated through the explained variance (EV) and the root mean square error (RMSE). The results showed that the calibration model based on 960 soil samples had the best performance in SOC prediction. Significantly biased predictions were produced by the calibration models based on more or less than 960 soil samples due to underrepresentation or overrepresentation. Relief and climate were demonstrated to be the predominant factors influencing SOC prediction in this study area. These results may provide theoretical support for studies relevant to SOC mapping. [ABSTRACT FROM AUTHOR]