Classification of erodible soils using artificial neural networks and biogeochemical indicators

Past research has highlighted the need to classify non-point-pollution sources in a watershed in order to work toward environmental remediation of the streams and rivers. The present work couples field measurements of carbon and nitrogen isotopes and the carbon to nitrogen atomic ratio with artificial neural network, multivariate logistic regression, and multivariate discriminant analysis modeling to perform the classification of erodible soils from forest and agricultural land uses. Results show the importance of all three biogeochemical indicators in aiding classification and it is proposed that all three indicators be measured and evaluated in future studies. Multivariate logistic regression performed adequately in classification and at the same time represented the most parsimonious nonnormal model, which highlights the need to avoid creating nonparsimonious models by application of advanced artificial neural network modeling techniques without underlying rationale. Model misclassification occurs due to a shift in biogeochemical cycles in a subset of the agricultural erodible soil data that is reflective of conversion of soil from tilled winter-wheat/barley rotation to conservation, and misclassification of forest soils is attributed to the high variability of soil organic matter cycling. Quantification of samples that can be tagged as potentially misclassified is performed in order that these samples can be further analyzed prior to watershed remediation.