Refining a reconnaissance soil map by calibrating regression models with data from the same map (Normandy, France)

Reconnaissance soil maps at 1:250,000 scale are the most detailed source of soil information for large parts of France. For many environmental applications, however, the level of detail and accuracy of these maps is insufficient. Funds are lacking to refine and update these maps by traditional soil survey. In this study we investigated the merit of digital soil mapping to refine and improve the 1:250,000 reconnaissance soil map of a 1580 km2 area in Haute-Normandie, France. The soil map was produced in 1988 and distinguishes nine soil class units. The approach taken was to predict soil class from a large number of environmental covariates using regression techniques. The covariates used include DEM derivatives, geology and land cover maps. Because very few soil point observations were available within the area, we calibrated the regression model by sampling the soil map on a grid. We calibrated three models: classification tree (CT), multinomial logistic regression (MLR) and random forests (RF), and used these models to predict the nine soil classes across the study area. The new and original maps were validated with field data from 123 locations selected with a stratified simple random sampling design. For MLR, the estimate of the overall purity was 65.9%, while that of the reconnaissance map was 55.5%. The difference between the purity estimates of these maps was statistically significant (p = 0.014). The significant improvement over the existing soil map is remarkable because the regression model was calibrated with the existing soil map and uses no additional soil observations