Premium
Erosion Index Derived from Terrain Attributes using Logistic Regression and Neural Networks
Author(s) -
Pike A. C.,
Mueller T. G.,
Schörgendorfer A.,
Shearer S. A.,
Karathanasis A. D.
Publication year - 2009
Publication title -
agronomy journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.752
H-Index - 131
eISSN - 1435-0645
pISSN - 0002-1962
DOI - 10.2134/agronj2008.0207x
Subject(s) - logistic regression , terrain , digital elevation model , elevation (ballistics) , erosion , artificial neural network , regression analysis , field (mathematics) , surface runoff , regression , logistic model tree , hydrology (agriculture) , environmental science , statistics , remote sensing , computer science , geography , cartography , artificial intelligence , geology , mathematics , geomorphology , geotechnical engineering , ecology , geometry , pure mathematics , biology
Maps identifying areas prone to channel erosion within agricultural fields could be useful for conservation planners. The objective of this study was to test an approach for creating such maps with logistic regression and neural networks. Survey grade elevation measurements were obtained from on a Central Kentucky farm. The elevation measurements were used to create 4 by 4‐m digital elevation models (DEMs) from which terrain attributes were derived. Areas exhibiting evidence of erosion caused by overland water flow sufficient to justify the placement of grassed waterways were identified. The terrain attributes were used as predictor variables and models were fit using the field assessments of soil erosion. Leave‐one‐field‐out validation analysis was conducted to assess the quality of predictions maps. For the models created with logistic regression, an average of 14% of the 4 by 4‐m grid cells in noneroded areas were incorrectly classified as being eroded and 16% of cells in eroded areas were incorrectly classified as noneroded. For neural network analysis, these error rates were 15 and 19%, respectively. Most of these errors occurred because the analyses did not exactly define the shapes of the eroded features; however, both logistic regression and neural networks identified most waterway features in all fields. The proposed three‐variable logistic regression model for erosion prediction should only be tested with datasets constructed using identical procedures for DEM creation and terrain analysis. This approach could improve the efficiency and accuracy of field site assessments for conservation planning.