An approach to map soil texture class on the Iowan Erosion Surface

Thumbnail Image
Date
2021-08
Authors
Ehret, Dustin Lane
Major Professor
Advisor
Miller, Bradley A
Gelder, Brian K
Burras, C Lee
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Altmetrics
Authors
Research Projects
Organizational Units
Journal Issue
Is Version Of
Versions
Series
Department
Agronomy
Abstract
The objective of this digital soil mapping (DSM) research was to compare approaches of new and legacy dataset combination to produce surface soil texture class maps on the Iowan Erosion Surface physiographic region (MLRA 104) using standard and novel predictor variables (covariates) for modeling. Hand-textured and laboratory-analyzed legacy particle size fraction (PSF) data were obtained from the NRCS National Soil Information System (NASIS). Four models examined various training dataset combinations of new and legacy data. Three PSF sub-models and maps were created for clay, silt, and sand for each parent model. The PSF maps were converted to a soil texture classification map using SAGA 7.4.0 and compared against the NRCS Soil Survey Geographic Database (SSURGO). A set of 320 covariates were provided to Cubist, a machine learning algorithm, to make the data combination models. Twenty-seven rasters were developed to capture eolian sediment transport distribution across the study area. A detailed methodology and case study was used to determine the utility of these rasters as predictors for surface silt content across the Iowan Erosion Surface. The case study showed that the eolian sediment transport distribution rasters were frequently selected by Cubist, but resulted in a higher RMSE and a lower Concordance Correlation Coefficient (CCC) during independent map validation than the model that excluded those covariates. Streaking artifacts present in the directional distance rasters suggest that methodological refinement is needed. At the PSF level, maps that included legacy data as a supplement to the new dataset returned lower RMSE and higher CCC statistics during independent validation. Similarly, the confusion matrices of the soil texture class maps showed that the combined new and legacy data maps outperformed maps trained exclusively with new data. The new data map correctly predicted soil texture class with 27% accuracy, while the maps with new and legacy data had correct classification between 39% and 42%. The results suggest that incorporating new and legacy data is an effective approach for predicting soil texture class. However, the SSURGO soil texture class map outperformed all new and legacy data combination maps with a classification accuracy of 58%.
Comments
Description
Keywords
Citation
Source
Copyright