Abstract:
Objective At the county level, the relationship between soil and soil forming environment is often complex and highly heterogeneous. How to accurately characterize the spatial variation of soil is a problem that needs to be studied.
Method This article selected Yizheng City as the research area, and based on 420 surface (0-20 cm) soil sampling points (of which 15% were randomly selected as independent validation points), compared the spatial prediction effects of soil attributes (total nitrogen, total phosphorus, total potassium, organic matter) using five methods: inverse distance weight, ordinary kriging, pan kriging, random forest, and random forest regression kriging.
Result The spatial distribution pattern of the four soil attributes predicted by 5 methods is basically the same, and the accuracy is different. Among them, the R2 of the random forest regression Krieger model is higher than other models in the prediction of all nitrogen, organic matter and phosphorus. In the prediction of total potassium, R2 is higher than other statistical models and random forest models.
Conclusion After comparing multiple models comprehensively, it was found that the random forest regression Kriging model has significant advantages in both accuracy and the detail of the predicted results, recommended as a mapping model for various agricultural districts and counties in the Third Soil Survey. There is a significant positive correlation between the four attributes.
下载: