Some Challenges on Quantifying Soil Property Predictions Uncertainty for the GlobalSoilMap Using Legacy Data

Abstract

The GlobalSoilMap project aims to create digital soil property maps in a raster format for six standard depths (0–5; 5–15; 15–30; 30–60; 60–100; 100–200 cm) and, for the first time, with estimates of uncertainty for predicted soil property maps. Data-driven methods and expert knowledge methods have been proposed, both of which present unique challenges. Initially, the majority of the predicted soil property maps will be derived from legacy soil data. The quantification of uncertainty, in particular, presents challenges due to the inherent nature of legacy data coming from different vintages (varying scales, formats, degree of completeness, differences in methods of observations, measurements, and classifications). We discuss the merits of each approach and potential practical and temporary solutions using two case studies from the USA, North America, and Llanos Orientales, Columbia, South America. Both case studies have limited data with insufficient point observations for a meaningful statistical approach for the estimation of prediction interval (PI) uncertainty. For the US case study, the available point measurements are not adequate for PI uncertainty quantification at soil map unit level and furthermore have been purposively collected to support the assignment of estimated mean, upper and lower property values to soil map units. We compared the estimated soil map unit upper and lower limits and 90 % CI from measured pedon for soil pH and found no significant differences between the two. The results suggest that the estimated upper and lower values from soil map units can be used for estimating the 90 % PI uncertainty at least initially until other independent measured point data become available. The available points in Llanos Orientales were collected for soil fertility evaluations and were independent of soil map unit polygons. However, they were surficial samples, clustered, and biased toward cultivated fields. As a result, only the 90 % CI was calculated and was found to be as wide as the range of the mean predicted soil property. These examples highlight few challenges in quantifying the 90 % PI and the need for more measured point data and flexible approaches when dealing with uncertainty quantification.