<p>Focusing on the complex challenges faced in the field of earthquake emergency response, this paper innovatively introduces a semi-supervised incremental learning (SSIL) strategy, which skillfully integrates the fast response characteristics of the physics-based analytical method with the data mining capabilities of the data-driven method. The framework relies on the Newmark method to build the semi-supervised learning foundation, and iteratively optimizes the machine learning (ML) model by continuously absorbing new data through the incremental algorithm, which demonstrates excellent information extraction performance and data fusion capability under resource-limited conditions. The study applies Bayesian optimization (BO) algorithms to tune the parameters of various machine learning models, such as Convolutional Neural Network (CNN) and Support Vector Machine (SVM). which significantly enhances the flexibility and prediction accuracy of the models, thus advocating the inclusion of BO in the process of standardizing machine learning models. In addition, this paper innovatively proposes a new evaluation criterion for <i>Landslide Sensitivity Interval Frequency Ratios Index</i> (<i>LSIFRs</i>), which directly maps the regional landslide risk sensitivity and can be used as a scale for landslide sensitivity prediction (LSP) accuracy. The results show that the SSIL strategy proposed in this paper is an ideal solution to meet the needs of post-earthquake emergency response. After a comprehensive assessment of the model performance, it was found that the SSIL-BOSVM model, which underwent BO enhancement, demonstrated significant utility and efficiency in practical applications, with a high area under the ROC curve (<i>AUC</i>) of 0.884 and an <i>LSIFRs</i> value of 0.416. The model can serve in future earthquake emergency management and post-disaster reconstruction work with technical support and data support.</p>