Soil mapping identifies different soil types by compressing a unique suite of spatial patterns and processes across multiple spatial scales. It can be quite difficult to quantify spatial patterns of soil properties with remotely sensed predictor variables. More specifically, matching the right scale of a given predictor variable with a specific soil type remains a challenge in digital soil mapping. An example of how scaling predictor variables derived from remote sensing and topographic indices can affect classification accuracy of soil types in southeastern Arizona is presented in the March–April 2017 issue of the Soil Science Society of America Journal. Support vector machine and random forest approaches produced the best results across a range of circular buffer windows for averaging predictors (0–180 m) compared with other commonly used machine learning models. Classification accuracy increased with the size of averaging window up to 150 m; however, larger window resolution produced general soil maps that poorly identified soils with narrow, linear features. Incorporating multiscale predictors produced soil–landscape features with a mix of general and detailed patterns.
These findings suggest that traditional soil survey concepts can inform new soil mapping efforts and optimize the extraction of landscape information to predict soil properties.