Regional soil thickness mapping based on stratified sampling of optimally selected covariates

Abstract A regional soil thickness prediction strategy based on stratified sampling was implemented. It can collect samples reflecting major topographic characteristics in a mesoscale spatial area to achieve the fine extension prediction of soil thickness from points to the entire area. The strategy firstly selected significant covariates based on historical soil thickness data by using random forest (RF) algorithm, and then these covariates could be used for clustering analysis by fast mean shift (FMS) method. In this way, a 73 km2 research area (27°54′7″–27°59′16″ N, 120°19′30″–120°26′46″ E) was divided into several subregions according to the similarity of variable characteristics between raster data generated in ArcGIS. Subsequently, we collected the new samples from each subregion through field investigations, so that the samples can reflect the characteristics of each subregion. Finally, four geographically weighed regression (GWR) models trained separately with the new samples from each of four large subareas were applied for the extension prediction of unobserved points. The results are as follows: (1) nine significant environmental covariates (i.e. slope position, slope length, slope of slope, land use, parent material, elevation, slope, topographic wetness index, normalized difference vegetation index) were obtained; (2) the whole research area were divided into 12 subregions, and 66, 49 and 44 samples were collected from the first three large subareas respectively; the fourth large subarea consists of the remaining nine small subregions, and a total of 93 samples were collected from it. (3) For the test set, the GWR prediction results of four large subareas as follows: for the first large subarea, the coefficients of determination (Rp2) is 0.789, the root mean square error of the prediction (RMSEP) is 1.518 and the relative prediction deviation (RPD) is 1.708; for the second large subarea, Rp2 = 0.704, RMSEP = 0.069, and RPD = 4.549; for the third large subarea, Rp2 = 0.667, RMSEP = 0.693, and RPD = 1.726; for the fourth large subarea, Rp2 = 0.623, RMSEP = 0.948, and RPD = 1.684. The above results basically meet the requirements of a qualified prediction model. Using the proposed method, the unobserved points in the whole research area were predicted and the soil thickness map was produced.