International Soil and Water Conservation Research (Dec 2024)
Modeling gully initiation by two codeless nonlinear methods: A case study in a small watershed on the Tibetan Plateau
Abstract
Land and soil resources are scarce in the Tibetan Plateau, and the region is facing ecological pressure from climate warming and increasing human activities. As a major ecological problem, gully erosion is destroying land and soil resources on the Tibetan Plateau, but related research is limited, and susceptibility areas and influencing factors are unclear. Machine learning methods are often applied to study gully initiation susceptibility, but they require a programming foundation. Therefore, the Redui watershed on the southern Tibetan Plateau with severe gully erosion was selected to evaluate the susceptibility and influencing factors of gully initiation through 12 influencing factors including topography, human activity, and underlying surface conditions, and all 2310 gully headcut sites. Two non-code nonlinear modeling methods, the categorical Regression (CATREG) and geographical detector (Geodetector) methods, were first used in the spatial modeling of gully initiation susceptibility. The results showed that the gully initiation susceptibility of the hillslope around the alluvial fan was highest. The very high susceptibility areas of the CATREG model and Geodetector model account for 18.2% and 16% of the total, respectively. The main influencing factors of gully initiation were elevation, relief, and soil type recognized by CATREG, and elevation, human footprint, and soil type recognized by Geodetector. Elevation is the primary factor controlling downstream susceptibility in both models. The primary factors in the upper and middle reaches are soil type and relief identified by CATREG. Human footprint, soil type, and distance to road are primary factors in the upper and middle reaches identified by Geodetector. The explanatory power of elevation, elevation-relief interaction, Geodetector model and CATREG model were 39%, 54%, 46.4% and 73.8%, respectively, at extremely significant levels (P < 0.001), which means that the influencing factors were well considered and that the methods have great application potential in the future.