Assessment of gully erosion susceptibility in Northeast China's black soil region using new stacking model with multiple machine learning algorithms

Gully erosion is a severe, globally widespread manifestation of soil erosion, threatening land, food, ecological, and human settlement security. Recently, machine learning methods have been increasingly applied to Gully erosion susceptibility (GES) assessment; however, most studies mainly focus on algorithm comparison, with limited attention to result rationality, classification suitability, and factor interpretability. Moreover, ensemble models remain underexplored. This study uses multisource data from a typical watershed, including UAV and field-surveyed gullies and environmental variables, to assess gully erosion susceptibility. Through factor selecting methods, 12 influencing factors were identified. Seven machine learning models (Random Forest, Extremely Randomized Trees, Light Gradient Boosting Machine, Extreme Gradient Boosting, K-Nearest Neighbors, Support Vector Machine and stacking), combined with 3 classification methods, namely, Equal Interval Method (EIM), Geometric Interval Method (GIM), and Natural Breaks Method (NBM), were used to classify the GES results through different grading approaches. Model and classification effectiveness were analyzed using standard metrics and rationality validation, followed by interpretability analysis of the best-performing model. The following results were obtained: (1) Among the 3 classification methods, the GIM yields a larger range of high susceptibility, and NBM provided more balanced result; both are more suitable than the EIM. (2) The stacking model outperformed all 6 single algorithms, with improvements ranging from 0.2 % to 27.8 %. (3) Rationality validation revealed that the classification performance of 7 models was mainly affected by gully target omission. The stacking model exhibited the highest congruence with field observations. (4) NDVI, elevation, distance to residence, land use, topographic wetness index, ridge direction, and distance to cropland were identified as the key factors. The interpretability analysis results consistent with gully development patterns in the study area, supporting the model's rationality. (5) The GES performance depended more on selecting key independent factors than on factor numbers. These findings support for improving GES and disaster early warning and prevention. • A stacking model combining six machine learning algorithms was developed to assess gully erosion susceptibility(GES). • Mutual information and multicollinearity tests were jointly used to select key factors for GES assessment. • GES outputs from 7 machine learning models were evaluated under 3 grading methods for comparison. • A rationality validation index was proposed and applied to verify GES results using independent gully samples.