County Level Digital Soil Mapping Using Decision Tree Model

Soil is the foundation of human agricultural production. It plays an important role in global.biogeochemical cycle. Soil information is the basic parameters in the model of land, environmental and agricultural management. It is beneficial for environmental protection, agricultural production and rational use of land resources when soil information is with thorough knowledge or understanding.Soil digital mapping is a result of rapid developing in computer and "3S" technology. It meets the demand of precision agriculture. It makes soil surveying and mapping less cost in manpower and material resources. Decision tree analysis is being increasingly used in digital soil mapping. It can reveal the relationship between soil information and landscape variations straightly and clearly.In the paper, several landscape variations were chosen to build the models of soil type-landscape and SOM-landscape by the C4.5 algorithm. The map of soil type and SOM were predicted using insufficient soil samples. The contents and main result of this study included four parts as follows:(1) Soil types and the distribution of SOM are closely related to landscape variations. The C4.5 algorithm was used to extract the relation of soil type-landscape and SOM-landscape. As a result the models of soil type-landscape and SOM-landscape were built.(2) Based on the mechanistic model for soil development, ten variations, such as soil land use, geology, elevation, aspect, slope plan curvature, profile curvature, normalized difference vegetation index(NDVI), normalized difference water index (NDWI), soil color index(SCI) were chosen to predict soil types in study area. The result shows that the overall accuracy is 70.5%. The accuracy of predicting red soil and paddy soil is better than other three types.(3) The ten variations in (2) and another variation, soil type were used to predict SOM. The decision tree model of SOM was built to get the predicted map of SOM in study. The result shows that the overall accuracy of SOM is 67.0%. It comes up to the requirements of general soil type mapping.(4) Two spatial interpolation methods, kriging and IDW, are used in predicting SOM. The accuracies of the two methods are 49.5% and 50.5%. Another geostatistical method co-kriging was used with elevation as an auxiliary variation. Its accuracy is 52.4%, which shows elevation is positive in predicting SOM using co-kriging. Be compared to decision tree model, spatial interpolation models emphasized the overall distributing of SOM while decision tree model reveals more details and tendency of transfers.