Differences in model application effectiveness, insufficient numbers of disaster samples, and unreasonable selection of non-hazard samples are common problems in landslide susceptibility studies. Therefore, in this paper, we propose a semi-integrated supervised approach to improve the prediction performance of machine learning (ML) models in landslide susceptibility studies. First, taking the lower reaches of the Jinsha River as the study area, a geospatial dataset consisting of 349 landslides, an equal number of randomly selected non-landslide points, and 12 environmental factors were randomly divided into training (70%) and testing (30%) datasets. Then, K-nearest neighbors (KNN), random forest (RF), and Bayesian-regularized neural network (BRNN) models were built. Second, the three models were combined to form an integrated weighted model. Very high- and low-prone areas were selected and, combined with the prediction results and remote sensing images, landslide and non-landslide samples were identified. The identified samples were then combined with the original samples to form new samples, which were used to sequentially construct the ensemble-supervised K-nearest neighbors (ESKNN), ensemble-supervised random forest (ESRF), and ensemble-supervised Bayesian-regularized neural network (ESBRNN) models. Finally, the area under the curve (AUC), true skill statistic (TSS), and frequency ratio (FR) values were used to test the accuracy of each model. The traditional ML model results and accuracy were improved by the semi-integrated supervised method. The ESRF model had the best prediction effect (AUC = 0.939, TSS = 0.440, and FR = 95.8%). The proposed semi-integrated supervised ML model solved the problems observed in traditional landslide susceptibility studies and provided insights for reducing variations in model applications, expanding landslide data sources, and improving non-landslide sample selection.
