Predicting the susceptibility to gully initiation in data-poor regions

- A two-step method to predict susceptibility to gully initiation in data-poor regions
- Published slope-area thresholds (S-A) are used as field measurement proxies
- Initiation areas of the current gullies are reconstructed from the S-A thresholds
- Logistic regression analysis to predict the spatial occurrence of initiation areas
- Susceptibility models are geomorphologically sound

Permanent gullies are common features in many landscapes and quite often they represent the dominant soil erosion process. Once a gully has initiated, field evidence shows that gully channel formation and headcut migration rapidly occur. In order to prevent the undesired effects of gullying, there is a need to predict the places where new gullies might initiate. From detailed field measurements, studies have demonstrated strong inverse relationships between slope gradient of the soil surface (S) and drainage area (A) at the point of channel initiation across catchments in different climatic and morphological environments. Such slope-area thresholds (S-A) can be used to predict locations in the landscape where gullies might initiate. However, acquiring S-A requires detailed field investigations and accurate high resolution digital elevation data, which are usually difficult to acquire. To circumvent this issue, we propose a two-step method that uses published S-A thresholds and a logistic regression analysis (LR). S-A thresholds from the literature are used as proxies of field measurement. The method is calibrated and validated on a watershed, close to the town of Algiers, northern Algeria, where gully erosion affects most of the slopes. The gullies extend up to several kilometres in length and cover 16% of the study area. First we reconstruct the initiation areas of the existing gullies by applying S-A thresholds for similar environments. Then, using the initiation area map as the dependent variable with combinations of topographic and lithological predictor variables, we calibrate several LR models. It provides relevant results in terms of statistical reliability, prediction performance, and geomorphological significance. This method using S-A thresholds with data-driven assessment methods like LR proves to be efficient when applied to common spatial data and establishes a methodology that will allow similar studies to be undertaken elsewhere.