Balancing method for landslide monitoring samples and construction of an early warning system

Given that machine learning is adept at uncovering implicit patterns from heterogeneous data sources, it is well suited for predicting landslide deformation with multi-factor monitoring. The sample dataset forms the foundation for training the models, and the quality and quantity of the dataset directly affect its accuracy and generalization ability. However, significant deformation in landslide bodies is relatively rare, leading to an imbalance in the collected sample dataset. To address this issue, this study proposed the genetic algorithm improved multi-classification-genetic-synthetic minority oversampling technique (SMOTE)-algorithm (GAMCGSA). Building on the multi-classification-genetic-SMOTE-algorithm (MCGSA), it integrated genetic algorithms to determine the optimal sampling rate. Based on this rate, new samples were generated, avoiding the creation of a large number of synthetic samples and effectively addressing the issue of sample imbalance. Subsequently, a convolutional neural network (CNN) was employed to process non-image data from multiple sources, resulting in the development of an intelligent landslide warning model. According to the test results, the F1 score of this model reached 84.2% with an accuracy of 90.8%, it possesses strong classification capabilities for both majority and minority classes, especially outperforming many current models (such as TabNet and RF) in classifying minority classes. This indicates that the CNN model has a superior ability to identify large-scale landslides. Based on the developed warning model and utilizing popular development frameworks, geographic information systems, and database technologies, an intelligent landslide monitoring warning system was constructed. This system integrates intelligent landslide monitoring and warning services, and provides scientific and reliable technical support for landslide disaster prevention and reduction.