• Depth, texture and stoniness classes were mapped over the Karnataka state (India). • DSM models were calibrated using 91 detailed soil maps rather than soil profiles. • Extrapolation areas were identified around these soil maps using soilscape distances. • Texture and stoniness classes were well predicted over 7% of Karnataka. • A hybrid soil mapping approach combining soil survey and DSM was found to be possible. Using detailed soil maps to calibrate DSM models could be an alternative to point observations, as they would account for local soil patterns more accurately than the sparse sets of soil profiles classically used in broad-scale DSM applications. However, the detailed soil surveys are most often scarce on large territories, which generates clustered calibration sets that may not represent the whole unmapped area. It is therefore important to delineate extrapolation areas that have soil-landscape relationships sufficiently similar to those of the soil map perimeter. We developed a DSM approach for mapping soil property classes (depth, texture and stoniness) over a large part (156,499 km 2 ) of Karnataka state, South India. We used a sparse set of 91 soil maps of micro-watersheds (464 km 2 of soil mapped areas) collected from recent land inventory programmes. Soilscape distances between soil maps were first defined by measuring the differences between soil property class distributions for each couple of soil-mapped micro-watersheds. A predictive model (random forest) that can estimate these ’ground-truth’ soilscape distances was then calibrated by using as covariates the differences of distributions and variograms of soil covariates (e.g. relief, climate, remote sensing data and small-scale soil maps), as well as the geographical distance between micro-watersheds. Soilscape distances were then used to select the appropriate DSM model for predicting soil property classes at each location (i.e. the model calibrated with the map of the closest micro-watershed). Soilscape distances served also to delineate extrapolation areas around existing soil maps in which soil property classes can be predicted with the highest accuracy and lowest predicted uncertainty. Using a leave-one-micro-watershed-out evaluation approach, We found that a single model calibrated onto the entire set of soil maps successfully predicted the texture and stoniness classes of soils over an extrapolation area covering 7% of the entire study area. Accuracies of 94% and 90% were obtained for texture, and stoniness, with respective predicted uncertainties of 6% and 7%. However, lower accuracy (57%) and higher uncertainty (31%) were obtained for predicted soil depth classes. Using multiple DSM models, each selected from soilscape distances, did not improve upon these results. This exploratory study paves the way for a possible hybrid approach to mapping soils across large territories. This approach would combine conventional soil surveys for detailed mapping of soil properties with digital soil mapping to extrapolate detailed soil maps. Digital soil mapping sampling techniques should also be employed in the future to select the locations of further detailed soil maps for mapping the target territory in an optimal way, thereby extending the extrapolation area while reducing survey costs.
Lagacherie et al. (Thu,) studied this question.