Assessing the predictive ability of multispectral and geomorphometric data for soil rock fraction and bulk density

Abstract

In recent decades, significant progress has been made in digital soil mapping methods. However, gaps remain, particularly with respect to the quality of maps for sloped and stony land. Attempts to increase agricultural production on such land, for instance through robotic planters, require the development of higher resolution soil maps emphasizing properties like bulk density and stoniness that determine planter energy use. I test our current ability to predict these properties by relating them to widely available multispectral and geomorphometric data. Three study areas in Colorado, Minnesota and Missouri were selected to cover a range of climate and landscape settings. About 50 sample locations for each site were selected using a Latin hypercube of representative curvature and topographic wetness index to set sampling locations. Unusually large sample volumes were used in triplicate at each sample location to reduce negative bias in stoniness measurements. A series of commonly used geomorphometric variables, along with multispectral data, were then employed in regression models to assess predictive ability for bulk density and rock fraction. Results indicate the models for stoniness were incapable of establishing a relationship between it and the model variables in both Minnesota and Missouri. In Colorado, however, a moderate r-square of 0.52 was observed, suggesting some relationship may be identified. With respect to the bulk density, all three models reported weak r-squares, though the Colorado model performed the best. Possible explanations for this performance include the low variation in stoniness at the Colorado study area, the limited relief therein, or a lack of obstructive vegetation. However, additional research, such as the inclusion of multispectral indices, is likely necessary.