Gully erosion zonation mapping using integrated geographically weighted regression with certainty factor and random forest models in GIS

Every year, gully erosion causes substantial damage to agricultural land, residential areas and infrastructure, such as roads. Gully erosion assessment and mapping can facilitate decision making in environmental management and soil conservation. Thus, this research aims to propose a new model by combining the geographically weighted regression (GWR) technique with the certainty factor (CF) and random forest (RF) models to produce gully erosion zonation mapping. The proposed model was implemented in the Mahabia watershed of Iran, which is highly sensitive to gully erosion. Firstly, dependent and independent variables, including a gully erosion inventory map (GEIM) and gully-related causal factors (GRCFs), were prepared using several data sources. Secondly, the GEIM was randomly divided into two groups: training (70%) and validation (30%) datasets. Thirdly, tolerance and variance inflation factor indicators were used for multicollinearity analysis. The results of the analysis corroborated that no collinearity exists amongst GRCFs. A total of 12 topographic, hydrologic, geologic, climatologic, environmental and soil-related GRCFs and 150 gully locations were used for modelling. The watershed was divided into eight homogeneous units because the importance level of the parameters in different parts of the watershed is not the same. For this purpose, coefficients of elevation, distance to stream and distance to road parameters were used. These coefficients were obtained by extracting bi-square kernel and AIC via the GWR method. Subsequently, the RF-CF integrated model was applied in each unit. Finally, with the units combined, the final gully erosion susceptibility map was obtained. On the basis of the RF model, distance to stream, distance to road and land use/land cover exhibited a high influence on gully formation. Validation results using area under curve indicated that new GWRCFRF approach has a higher predictive accuracy 0.967 (96.7%) than the individual models of CF 0.763 (76.3%) and RF 0.776 (77.6%) and the CF-RF integrated model 0.897 (89.7%). Thus, the results of this research can be used by local managers and planners for environmental management.