An Adaptive Uncertainty-guided Sampling Strategy For Digital Soil Mapping

Spatial distribution of soil attributes information is an important part of land resource utilization,hydrological process simulation and environmental resource management and protection.In recent years,digital soil mapping(DSM)based on GIS has gradually become the mainstream method for obtaining soil information.The sampling stage is one of the key issues in digital soil mapping.Since there are parts of samples previously collected in many regions,the research on how to design additional samples under conditions with a certain number of samples is an urgent issue in the field of current sampling design methods.At present,one of the latest research on the supplemental sampling method for digital soil mapping are theoretically based on knowledge of soil-environment relations.By calculating the similarity between the sample points and the unvisited points,the prediction uncertainty of each unvisited point can be obtained.With the use of this prediction uncertainty and user-defined uncertainty thresholds,the predictable and unpredictable areas can be defined.Based on this,we can select the points that maximize the predictable area and reduce the prediction uncertainty as the new supplementary samples.However,the existing method has the disadvantages of complex algorithm design and artificially subjectively setting parameters,which results in not only the technical complexity but also the problem of low sampling efficiency.To solve this problem,this paper focuses on how to improve the algorithm design and proposes an adaptive uncertainty-guided sampling strategy for digital soil mapping.The key research ideas of this paper are as follows:(1)Constructing the unified objective function based on prediction uncertainty;(2)designing the adaptive approaches to adjust parameters,including the weights of sub-objectives and uncertainty threshold.In order to validate this method,we selected two research areas.One is Raffelson area which is in Wisconsin,USA.Another is Xuancheng area,Anhui Province in China.The experimental results show that the proposed method can adaptively adjust the weight parameters of sub-objectives adaptively,and ensure that the total objective function can be reduced continuously during the iterative process so as to help find the optimal complementary samples.This method can achieve rapid reduction of the total prediction uncertainty and increase the predictable area,and improves the sampling efficiency.We also compares the proposed adaptive sampling method with the existing non-adaptive sampling method and stratified random sampling method,and selects three commonly used prediction models,which are classification and regression tree(CART),random forest and individual predictive soil mapping(iPSM).The results show that the proposed method can obtain highest accuracy in most cases,and can get more stable inference results than other methods.We also analyzes the key adaptive adjustment parameters contained in the method.The results show that either parameter setting too large or too small will lead to higher prediction uncertainty.And found the optimal combination of parameters.This result also provides a reasonable default reference of parameters for the users to use the method in this paper.In summary,this paper presents an adaptive uncertainty-guided sampling strategy for digital soil mapping.We improve the original non-adaptive complement sample to an adaptive sampling optimization method.Under the ensuring of a slight improvement in accuracy,the disadvantage of subjective manual adjustment in the original parameter is eliminated,which reduces the difficulty for the designer and improves the usability and sampling efficiency.