Prediction of landslide susceptibility in Wenchuan County based on pixel-level samples

The essence of landslide susceptibility assessment is to conduct a probability assessment of landslide occurrences in a specific area based on historical landslide data. The majority of the results of landslide susceptibility evaluation depend on the fineness of the samples. Traditional sample production methods utilize statistical methods for quantification of dependent variables, and statistical formulas lead to a loss of information on the precise locations of landslides. This leads to uncertainty in the final prediction results. In this work, a new form of pixel-level sample production is proposed to preserve the landslide's boundary location information as much as possible. Three machine learning models, namely, a logistic regression model, a deep neural network model, and a transformer model, are combined with the sample production method proposed in this paper. The accuracy was verified using receiver operating characteristic curves. The three models' areas under the curves were 0.935, 0.963, and 0.980, respectively. The results of the susceptibility zoning showed that the TR model achieves a much finer classification of very high-landslide susceptibility areas, which makes it convenient to conserve human and material resources and focus on high-landslide-occurrence areas.