What can GlobalSoilMap expect from Vis-NIR HyperSpectral Imagery in the near future? . :.
Résumé
This paper reports on the use of legacy soil data for estimating three of the soil properties that are in the GlobaSoilMap specifications-Clay (CL), Organic Carbon (OC) and pH-in Languedoc Roussillon, a 27,236 km(2) region of southern France that stretches from the Mediterranean Sea to the Pyrenees and Massif Central mountains. Three Digital Soil Mapping (DSM) methods using two types of soil legacy data produced by the French National Survey program were compared: i) area-weighted means of properties for the soil mapping units of a 1:250,000 scale soil map, ii) Random Forest and iii) Random Forest + kriging predictions from a set of legacy soil measured profiles. There were significant differences in prediction performance across soil properties, DSM methods and depths. For the soil properties, pH was better predicted (best R-2 between 0.69 and 0.75) than OC (best R-2 between 0.11 and 0.42), which was better predicted than CL (best R-2 between 0.12 and 0.23). DSM methods using legacy measured profiles generally outperformed the one using the soil map although the differences were not systematically significant. All differences in performance could be largely explained by the contrasting short-scale spatial variability of the studied soil properties in Languedoc Roussillon. Examples of digital soil maps are presented and these prefigure the production of the Languedoc-Roussillon excerpt of GlobalSoilMap.