Statistical sampling design impact on predictive quality of harmonization functions between soil monitoring networks

Abstract Regulations about soil quality are normally imposed at international level while many countries have set up monitoring networks at national scale. Since these networks use different sampling strategies, there is a strong need to harmonize a posteriori the collected data from the national networks in order to answer questions raised by the global regulations. For that purpose, calibration sites where different sampling strategies are carried out are necessary in order to construct harmonization functions between measurements from different sampling protocols. A case study is available for French forest soils that have been sampled twice simultaneously on the same sampling grid but with different sampling and analytical strategies: a first sampling for the French soil quality monitoring network (RMQS) and a second one for the European forest monitoring network (ICP Forests level I second survey i.e. Biosoil). However, the way to define the number and the position of these calibration sites remains a key issue. In this work, we compare both RMQS and Biosoil strategies for a set of measured variables of interest (carbon, potassium and lead contents and pH) and aim to define the minimum number of sites and their best location to establish reliable harmonization functions. Three statistical methods for construction of sampling designs are tested: random sampling, conditioned Latin Hypercube Sampling (cLHS, Minasny and McBratney, 2006 ) and D-Latin Hypercube Sampling (DLHS, Minasny and McBratney, 2010 ). With each method, we investigate the effects of the number of calibration data on the predictive quality of the harmonization functions. First, we show that both cLHS and DLHS are better than simple random sampling. Then, the difference between cLHS and DLHS performance depends mainly on the size of the samples, the nature of the soil property and the form of the pedotransfer functions.