ارزيابي حساسيت فرسايش خندقي با استفاده از تكنيك‌هاي يادگيري ماشين (منطقه مورد مطالعه: حوضه آبخيز ميان¬آب- شوشتر)

Soil is one of the most important an‎d valuable natural resources, providing around 95% of human food. Without healthy soil, life on Earth would not be possible. Soil erosion is a physical process an‎d one of the most dangerous mech-anisms of lan‎d degradation that leads to the reduction of fertility an‎d damage to agricultural lan‎ds. One of the major processes of soil degradation is gully ero-sion, which causes significant soil loss an‎d produces large amounts of sediment in various climates. Gully erosion is considered one of the most complex an‎d destructive types of water erosion. Therefore, studying, identifying, an‎d mapping areas prone to gully erosion has become highly necessary. The aim of this study is to identify the factors influencing gully erosion occurrence, predict, an‎d map the likelihood of its occurrence in the Mianab-Shushtar watershed. Initially, us-ing Google Earth images, satellite imagery an‎d field surveys, a map of gully dis-tribution points was prepared. In this research, 4000 points with the presence of gullies an‎d another 4000 points without gullies were used. Next, the most im-portant topographic indices were prepared as environmental parameters influ-encing gully erosion occurrence. The independent variables affecting gully ero-sion include elevation, slope degree, slope direction, slope length index (LS), topographic wetness index (TWI), stream power index (SPI), sediment transport index (STI), lithology, lan‎d use, five vegetation indices including NDVI, EVI, ARVI, SAVI an‎d OSAVI, distance to streams, stream density, distance to faults, fault density, distance to roads, road density, average annual precipitation, rain-fall erosivity index (R), soil texture (san‎d, clay an‎d silt), an‎d soil erodibility in-dex (K).Among the 14 machine learning algorithms implemented in this study, CatBoost, LightGBM, an‎d AdaBoost algorithms showed higher accuracy with an F1-score over 94% compared to other algorithms. Additionally, among the 23 indices influencing gully erosion occurrence in the area, lan‎d use was identified as the most important variable. The next most influential indices were ARVI an‎d stream density. Based on the classification results of the three best algorithms, about 67/984% of the area falls in the very low-risk class, 14/074% in the low-risk class, 7/739% in the moderate-risk class, 5/363% in the high-risk class, an‎d 4/837% in the very high-risk class. The results of this study will play a signifi-cant role in preventing an‎d reducing the damage caused by gully erosion in the studied region.

5

138065