Zoning gully erosion susceptibility using ANN, CART and RF models

Gully erosion is a water erosion that has a great contribution to land degradation and is known as one of the most important environmental hazards in the world and especially in Iran. In recent years, machine learning techniques and geographic information systems have been highly effective in determining areas sensitive to gully erosion and have increased accuracy and speed in the evaluation and potential of gully erosion and in determining effective factors on gully erosion has also been effective. The loess lands of Golestan Province are more susceptible to water erosion due to sufficient depth and almost uniform silty graining, excessive use, cultivation on sloping lands, and wrong land management so that all types of erosion can be observed in these areas. The most common type of erosion in these sediments is gully erosion. The studied watershed is faced with the increase of dry and abandoned land, land use change, the presence of surplus livestock in the forests, and also the population increase. Therefore, this area is facing an increase in sensitivity to gully erosion, and areas with the potential for gully erosion should be identified and managed. 

The studied watershed with an area of 222,000 ha and an elevation range of 58 to 2168 m is located in the northeast of Golestan Province. The average rainfall of the area is between 224 and 736 mm. In this research, first, the location of the gullies was obtained from the General Directorate of Natural Resources and Watershed Management of Golestan Province. Then, from the total of 1127 gullies position, 70% were randomly classified as training data and 30% as validation data. To determine the effective variables in gully erosion sensitivity, 14 factors were identified and in the next step, the collinearity test between the variables was performed using SPSS software. By using the indices of tolerance coefficient and variance inflation factor, if there is collinearity between the variables, they were removed from the modeling process. Considering the importance of the DEM map and its application in the preparation of various factors of the current research, a DEM was prepared using ALOS satellite images. The layers of slope and aspect are prepared by using a digital elevation model and slope and aspect functions respectively. Slope length index in SAGA GIS software, layers of distance from stream based on the map of stream, and distance from roads based on existing roads, and using the Euclidean distance function in the ArcGIS software was prepared. Stream density and road density layers were obtained based on the map of existing streams and roads in the region and using the line density function in ArcGIS. The lithology layer was extracted from the geological map of the region and the land use layer was obtained from the General Directorate of Natural Resources and Watershed Management of Golestan province. The rainfall map has been prepared using the information from 35 rain gauge stations. First, the average rainfall of 26 years was calculated for each station, and then rainfall zoning was done using the global Kriging Method (due to the lowest RMSE) in ArcGIS. The TPI layer was prepared using the DEM and using the SAGA GIS software. The HAND index is a topographic-hydrological index of the DEM of the nearest drain, representing the hydrological behavior of the watershed. To evaluate the models, the relative performance detection curve (ROC) was used for the predictive power of the models.  

The results showed that there is no co-linearity between the variables and therefore all the variables were used in the modeling process. The relationship between gully erosion and elevation showed that lower elevations are more sensitive than higher elevations and more susceptible to gully erosion near waterways. The results showed that with the increase in drainage density, the sensitivity of gully erosion increases, and the possibility of gully erosion increases. The results showed that the old barracks, shale, and loess have the greatest impact on the sensitivity of gully erosion. The results show a decrease in the sensitivity of gully erosion with a decrease in the HAND index. This result indicates that in the areas where the level of saturation in the watershed level increases, the possibility and sensitivity of gully erosion increases. The results showed that among the types of land use, canals, poor pastures, and agricultural land use have the highest sensitivity to gully erosion. This is even though the forest areas have the lowest sensitivity to this erosion. The results showed that in the rainfall range of 220 to 420 mm, the possibility of gully erosion has increased, and the range of 420 to 500 mm has shown the highest level of sensitivity, and with the increase of rainfall from 500 mm to above, a reduction in the sensitivity of gully erosion has been encountered. One of the reasons for reducing the sensitivity of gully erosion in higher rainfalls is the increase in vegetation and the creation of suitable conditions for landslides. The results showed that the depth of the valley up to 235 meters have increased the probability of gully erosion, and from 235 meters above, it has decreased the probability of erosion. The results showed that the sensitivity of gully erosion increases near roads, and this case shows the effects of road construction and the aggravation of conditions for gully erosion. 

This research was conducted to determine the effective factors on gully erosion and zone its spatial distribution in the northeast of Golestan Province. In this study, by considering 14 important factors and using RF, ANN, and CART models, a sensitivity map of gully erosion was prepared. Because the identification of gully erosion-sensitive areas based on traditional methods and expert opinions do not have acceptable accuracy, it is necessary to use modern machine learning methods. The results showed that the factors of distance from the road and land use are the most important factors affecting the sensitivity of gully erosion, which requires land use management as human activities. The ROC curve showed that the accuracy of the models in estimating areas with gully erosion sensitivity was excellent in the test stage (ANN) and very good in the test and validation stage (RF and CART), which means the excellent performance of the models.