Predict soil texture distributions using an artificial neural network model

Abstract

High-resolution soil maps are important for planning agriculture crop production, forest management, hydrological analysis and environmental protection. However, high-resolution soil maps are generally only available for small areas because obtaining these maps through field survey is time consuming and expensive. The objective of this study was to develop an artificial neural network (ANN) model to predict soil texture (sand, clay and silt contents) based on soil attributes obtained from existing coarse resolution soil maps combined with hydrographic parameters derived from a digital elevation model (DEM). The calibrated ANN model then can be used to produce high-resolution soil maps in area with similar conditions without additional field surveys. The hydrographic parameters derived from DEM were soil terrain factor, sediment delivery ratio and vertical slope position. Field measured soil texture in the Black Brook Watershed (BBW) in northwestern New Brunswick, Canada was used to train and test the ANN model. Results indicated that the Levenberg–Marquardt optimization algorithm was better than the commonly used training method based on the resilient back-propagation algorithm. The root mean square errors between model predictions and field determination were 4.0 for clay and 6.6 for sand contents. The relative overall accuracy (within ±5% of field measurement) was 88% for clay content and 81% for sand content. The trained ANN model has been tested in an experimental farm located in southeastern NB about 180 km from the Black Brook Watershed where the model was first calibrated. Results indicated that with proper training, the ANN model can be used in the areas where the model was calibrated (for interpolations), or other areas provided that the relative range of input parameters were similar to the region where the model was calibrated.