Optimal scaling of predictors for digital mapping of soil properties

Abstract

Scientists have long been developing better and more efficient methods to improve the prediction of the spatial distribution of soils and their presence in the landscape, but research in this field is still needed. This study introduces an algorithm to derive terrain attributes at multiple scales and automatically calibrate the optimal scale for each predictor based on the robust and powerful Random Forests (RF) method, to improve the accuracy of soil property mapping. Experiments are conducted to evaluate to what extent optimally scaled predictors lead to improved accuracy of digital mapping of nine soil properties. The procedure starts with the resampling of the original 12.5 m digital elevation model (DEM) to 25 m, then in 25 m increments to 1000 m, thus resulting in 40 broader versions of the DEM. Ten terrain attributes were derived from each downscaled version of the DEM, resulting in 40 downscaled versions of each terrain attribute. Soil property values are then used to create both a RF model and a linear correlation with every scaled terrain attribute. The script exports two sets of optimally scaled terrain attributes, as defined by the maximum value of the R-squared value and the correlation coefficient, respectively. For each soil property, the predictors were prepared into four pools: optimally scaled predictors based on the RF model; optimally scaled predictors based on the correlation coefficient; all multiscale predictors, and original not scaled predictors. The results proved that more accurate and less uncertain soil property maps could be obtained when predictors are optimally scaled, as compared to maps created with original not scaled or all multiscale predictors. The results further confirm earlier findings that a subset of carefully selected predictors works better for mapping a given soil property: a subset of only 27–53% of predictors led to better maps, as compared to the models based on all the available predictors.