A hybrid machine learning model for disaster prediction using historical geological disaster data

Abstract

As disasters are occurring more frequently and weather extremes get harsher, communities and governments around the world are facing more and more difficulties. In order to address these problems, disaster preparedness must be done precisely. This research examines in detail the transformative impact that machine learning algorithms have on bolstering disaster preparedness and response systems. Beyond a simple synopsis, our study’s Enhanced Pelican Optimization based on Disaster Prediction long short-term Memory (EPO-DPLSTM) is remarkable and shows off the advanced capabilities of Machine Learning (ML) in predicting a wide range of patterns of the weather and natural disasters, such as waves in heat, hurricanes, floods, droughts, and more. In order to assist the improved efficacy of prediction models in disaster preparedness, we made useful observations into the intricacies of application using ML. In addition to outlining the theoretical underpinnings, the study offers empirical evidence of the substantial advantages that machine learning algorithms offer. By using these precise forecasts of past geological disasters and new weather trends, preventative measures might be put in place, ultimately saving lives and lessening the extent of the damage. Regional landslide catastrophe early-warning is a crucial tool for disaster prevention and mitigation in China, where disasters are severe. A proposed approach to regional disaster warning was presented in this research. The model creation process includes warning output, model parameter optimization, sample learning and training, sample-set construction, and so forth. Eighty percent of the training sample set was used as the trained set, and twenty percent was utilized as the testing set for cross-validation in the sample learning and training process. The model parameters were optimized using the Enhanced Pelican Optimization based on the Disaster Prediction Long Short-Term Memory (LSTM) algorithm, and the accuracy, Receiver Operating Characteristic (ROC curve), and Area Under the Curve (AUC) value were utilized to confirm the model’s generalization capacity and accuracy. To improve model training, five machine learning methods were used; the results indicated that the suggested algorithm was the model with the best generalization capacity (AUC was 0.989) and performed the best, with an accuracy of 99.5%. The findings of this study provide critical scientific support for policymakers, emergency planners, and local stakeholders, enabling the development of more targeted, data-driven disaster mitigation strategies and strengthening regional resilience against future geological hazards.