JOURNAL OF NATURAL RESOURCES ›› 2017, Vol. 32 ›› Issue (6): 1074-1086.doi: 10.11849/zrzyxb.20160623

• Resource Research Method • Previous Articles    

Soil Organic Matter Prediction Based on Remote Sensing Data and Random Forest Model in Shaanxi Province

QI Yan-bing1, 2, WANG Yin-yin1, CHEN Yang1, LIU Jiao-jiao1, ZHANG Liang-liang1   

  1. 1. College of Natural Resources and Environment, Northwest A & F University, Yangling 712100, China;
    2. Key Laboratory of Plant Nutrition and the Agri-environment in Northwest China, Ministry of Agriculture, Yangling 712100, China
  • Received:2016-06-14 Revised:2016-07-15 Online:2017-06-20 Published:2017-06-20
  • Supported by:

    Special Foundation of National Science and Technology Basic Work Project of China, No. 2014FY110200A08.


There exists deviation of predication of soil organic matter (SOM) with observed data in special local topography units. The accuracy of SOM predication can be improved by combining observed data and remote sensing (RS) data, especially for SOM predication in large scale. In this study, AWIFS (Advanced Wide Field Sensor) and MODIS (Moderate Resolution Imaging Spectroradiometer) data, whose spatial resolution are 56 and 250 meters respectively, were combined with observed sample data to predict the spatial distribution of SOM in Shaanxi Province with RF (Random Forest) model. The spatial distribution of SOM in six types of topographical units were summarized, and the prediction accuracies of SOM based on RF model and OK (Ordinary Kriging) model were compared. The results indicated that the spatial differentiation of SOM is obvious in north-south direction in Shaanxi Province. It is the highest in Qinling and Daba mountain areas with SOM content higher than 25 g·kg-1, and it is medium high in the south of Loess Plateau area with the SOM content 22-30 g·kg-1. The content of SOM is lower in Guanzhong Plain and Hanzhong basin areas with SOM content among 13-25 g·kg-1, while it is the lowest in north Loess Plateau and the blown-sand areas with SOM content less than 10 g·kg-1. The prediction results based on AWIFS data (with higher spatial resolution) were better than those based on MODIS (with lower spatial resolution) data. The acquired data of images has little influence on SOM prediction. It is shown that the predicted value of SOM is a bit lower in autumn than in spring. With different percentages of sampling, the SOM prediction based on RF model is always better than that based on Ordinary Kriging model. The prediction accuracy in this study is reliable, because the mean error in independent validation set is no more than 3 g·kg-1,and the correlation coefficient of the predicted values and the observed values are higher than 0.7. Elevation is the most importance factor influencing SOM prediction in Shaanxi Province. When the spatial resolution of RS data decreases, the importance of geographic location of sampling points increase and the importance of vegetation decrease.

Key words: multi-resolution remote sensing data, random forest algorithm, Shaanxi Province, soil organic matter prediction

CLC Number: 

  • S158