Regional variability of terrain index and machine learning model applications for prediction of ephemeral gullies

Abstract

Terrain, or topographic, index models aggregate morphometric features of a landscape and use them to predict trajectories and initiation points of classical or ephemeral gullies. As an alternative to the index-based models, statistical machine learning approaches have been recently gaining an increased attention for identification of areas of gully susceptibility. Application of index-based or statistical models is normally restricted to the area of study, and regional model transferability is not well understood. In this study, seven terrain index models and two machine learning algorithms were applied to eight watersheds in Kansas and two watersheds in Sicily. The predictive ability of the nine models was measured by using both cut-off independent (area under the receiver operating characteristic curve) and dependent (Cohen's kappa index, sensitivity, specificity) statistics. The performance statistics of both terrain index and statistical models were cross-examined by finding an optimal threshold in one watershed and applying it to other watersheds. In Kansas, the model based on stream order (GORD) was found to perform similar to machine learning approaches, whereas the modified stream power index (MSPI) model overperformed other terrain index models in Sicily. Different landform characteristics in cultivated areas of the High Plains region in Kansas and in the steep hillslopes of Central Sicily caused models to have variable success rate when the indexes from one region were transferred to another region. The results also showed that a well-calibrated terrain index model, especially based on stream grid order, can be viewed as a valid alternative to a data driven approach for ephemeral gully mapping, but transferability of the optimal thresholds can be applied with caution.