Modelling the occurrence of gullies at two spatial scales in the Olteţ Drainage Basin (Romania)

Gully erosion is both a significant natural hazard and an important sediment source. The design of proper prevention measures relies firstly on the prediction of the future locations of gullies. But, as recent progress has shown, methods and results in any environmental modelling greatly depend upon scale, the selection of which should be based on management needs. This research deals with predicting the spatial potential for gullying at two different scales grouped in a top-down framework, i.e. starting with a preliminary, regional scale analysis (1:100,000–1:200,000) to develop a simplified model, and performing a more detailed, intermediate level analysis at a medium scale (1:25,000–1:50,000) for the basin sector revealed as the most threatened by the process. At the same time, the study searches for relationships among: scale of analysis, area of investigation, precision and accuracy of input data, and the quality of expected results and their applicability. The study area is the Olteţ Drainage Basin (2439 km 2 ) in southern Romania, which extends over four landform types: mountains, hills, a plateau (piedmont) and a plain. Aiming to investigate the scale effect, the same statistical method is selected for both analyses, namely Classification and Regression Trees (CART). Scale-adapted procedures and resolutions are applied for defining the dependent variable, deriving the environmental attributes and deciding the sampling strategy needed to provide information for the statistical analyses. In order to detect the degree of gullying at the regional scale, the statistical method is used over the entire basin, and gully density is selected as the target variable. For the analysis of gully susceptibility at the medium scale, within the most affected area of the Olteţ Basin identified as the piedmont sector, the single gully and the intensely gullied spot are defined as the target variables. The best validated maps obtained at the two scales are compared. The results reveal that both individual maps are characterized by statistical accuracy (a NRMSE value of 0.05–0.08 and an AUC of 0.86 for the regional scale and the medium scale models, respectively). Yet, the regional scale map is affected by high uncertainties when compared to the medium scale one. The scale dependency of results and hence the relative nature of their accuracy and reliability are highlighted in the context of both fundamental and applied research. Copyright Springer Science+Business Media Dordrecht 2015