Inversion for spatial distribution of soil organic matter content based on multivariate geostatistics

The classical statistical method is always used to construct quantitative remote sensing retrieval model. However, the method doesn't take into account the spatial relations between data, which will severely affect the retrieval accuracy. In order to improve the spatial predictive accuracy of soil organic matter, a multivariate geospatial method for making retrieval model was presented in this paper. Considering the spatial distribution characteristic of regression error, a multivariate geostatistical method called ordinary Kriging with varying local means (OKLM) was presented, which was used to construct remote sensing retrieval model. The method was illustrated using soil organic matter (SOM) content in Southwest Sichuan province, and was compared with other method, such as ordinary Kriging, ordinary remote sensing retrieval method, and remote sensing retrieval model based on regression Kriging. The results showedthe proposed method improved the predictive accuracy effectively among these methods, because the proposed method was based on relations between SOM sampling data and TM images using spatial statistics, taking fully into account the spatial relations among the data, and obtained more accurate retrieval model. Compared with regression Kriging, OKLM assumed that the means of regression errors cannot always be zero in local neighborhood, which was more in line with the actual situation. The proposed method provides a scientific basis for the farmland nutrient management and sustainable development of the regional agricultural.