Generation of soil drainage equations from an artificial neural network-analysis approach

Abstract

Zhao, Z., MacLean, D. A., Bourque, C. P.-A., Swift, D. E. and Meng, F.-R. 2013. Generation of soil drainage equations from an artificial neural network-analysis approach. Can. J. Soil Sci. 93: 329–342. Soil properties, especially soil drainage, are known to be related to topo-hydrologic variables derived from digital elevation models (DEM), such as vertical slope position, slope steepness, sediment delivery ratio, and topographic wetness index. Such relationships typically are strongly non-linear and thus difficult to define with conventional statistical methods. In this study, we used artificial neural network (ANN) models to establish relationships between soil drainage classes and DEM-generated topo-hydrologic variables and subsequently formulated the relationships to generate soil drainage equations for soil mapping. A high-resolution field soil map of the Black Brook Watershed in northwest New Brunswick, Canada, was used to calibrate/validate the ANN models, and the obtained equations. Independent data from an experimental farm, about 180 km away, were also used for validation. Results indicated that vertical slope position was the best predictor of soil drainage classes (r=0.55), followed by slope steepness (r=0.44), sediment delivery ratio (r=0.39), and topographic wetness index (r=0.38). The obtained soil drainage equations fitted well to the ANN model predictions (r2=0.78–0.99; root mean squared error=0.39–4.55). Analyses indicated that soil drainage equations clearly reflected the actual relationships between soil drainage classes and DEM-generated topo-hydrologic variables, and have the potential to minimize bias originated from over-training the ANN models when applied outside the area of calibration, especially when the ranges of input variables were outside of the range of calibration data.