Forecasting vertical ground surface movement from shrinking/swelling soils with artificial neural networks

Artificial neural networks (ANNs) are used to estimate vertical ground surface movement when soils expand and contract due to changes in soil moisture content caused by changing climate conditions. Several counterpropagation ANN test cases were investigated to map climate data (i.e. temperature and rainfall) to vertical ground surface movement at field sites in Texas and Australia. Three of the four ANN test cases use a historical time series of climate data to forecast ground surface elevation relative to a specified datum. The fourth ANN test case predicts the rate of ground surface movement, and requires post-processing of the predicted rates to calculate ground surface elevation relative to a specified datum. The counterpropagation network has demonstrated a successful mapping of temperature and rainfall data to vertical ground surface movement at a field site when it is trained with a subset of data from the same field site (test cases 1 and 2). The results of training an ANN on one field site and testing it on another field site (test cases 3 and 4) demonstrate the ability of the ANN to capture trends in vertical ground surface movement. When compared with the predictions from a physics-based method (shrink test-water content method) that requires measurements/estimates of changes in soil water content, the ANN-based predictions (based on climatic changes) captured the trends in the field measurements of shrinking–swelling soil surface movements equally well. These findings are promising and merit further investigation with data from additional field sites. Copyright q 2007 John Wiley & Sons, Ltd.