Predicting gully initiation: Comparing data mining techniques, analytical hierarchy processes and the topographic threshold

Predicting gully initiation at catchment scale was done previously by integrating a geographical information system (GIS) with physically based models, statistical procedures or with knowledge-based expert systems. However, the reliability and validity of applying these procedures are still questionable. In this work, a data mining (DM) procedure based on decision trees was applied to identify areas of gully initiation risk. Performance was compared with the analytic hierarchy process (AHP) expert system and with the commonly used topographic threshold (TT) technique. A spatial database was used to test the models, composed of a target variable (presence or absence of initial points) and ten independent environmental, climatic and human-induced variables. The following findings emerged: using the same input layers, DM provided better predictive ability of gully initiation points than the application of both AHP and TT. The main difference between DM and TT was the very high overestimation inherent in TT. In addition, the minimum slope observed for soil detachment was 2°, whereas in other studies it is 3°. This could be explained by soil resistance, which is substantially lower in agricultural fields, while most studies test unploughed soil. Finally, rainfall intensity events >62.2mmh ^-1 (for a period of 30min) were found to have a significant effect on gully initiation.