A New Approach to Estimate Prediction Uncertainty using Sparse Ad hoc Samples in Digital Soil Mapping application

Tuesday, 16 December 2014
Jing Liu1, A-xing Zhu1,2, Fei Du1 and Shujie Zhang3, (1)University of Wisconsin Madison, Madison, WI, United States, (2)Nanjing Normal University, Nanjing, China, (3)IGSNRR Institute of Geographic Sciences and Natural Resources Research, CAS, Beijing, China
The importance of the spatial variation of prediction uncertainty in digital soil mapping (DSM) application has gained increasing awareness in the past decades. An appropriate estimation on prediction uncertainty is useful not only in revealing the quality of the predicted soil map, but also in identifying areas where further efforts for reducing prediction error can be carried out. Existing probabilistic approaches for estimating prediction uncertainty can be inapplicable in practice due to the violation of the stationary assumption and the limited number of field samples available for constructing the probability distribution function of the targeted soil property. This paper presents a new approach to estimate the spatial variation of prediction uncertainty in digital soil mapping application only using sparse ad hoc field samples. The approach assumes that the more similar environmental conditions between two locations the more similar their soil properties will be. The prediction uncertainty at every un-visited location is estimated by analyzing a similarity representation, which indicates how reliable each of the existing field samples can be used to predict the value of targeted soil property at that location. Two case studies were conducted to evaluate the methodology and to illustrate how the quantified uncertainty can be used to assess the quality of the soil map derived from digital soil mapping techniques. The case studies showed that the areas where the predicted soil organic matter content (%, mass percentage) in the topsoil layer along with high prediction uncertainty were associated with low prediction accuracy and vice versa. A clear positive relationship between the prediction errors (residuals) and the quantified prediction uncertainty has been found in both case studies. In conclusion the method presented in this paper is effective in estimating prediction uncertainty when only sparse ad hoc field samples are available. The quantified prediction uncertainty can be an effective indicator to the accuracy of the derived soil map and can serve as an effective tool for allocating error reduction effort in a cost-effective way.