Spatial and Temporal Patterns for Disaster Prediction Using the Global Disaster Dataset

Natural disasters, such as floods, earthquakes, and storms, cause significant human and economic losses worldwide, necessitating accurate early prediction to enhance preparedness. This study leverages spatial (location-based) and temporal (time-based) patterns in the EM-DAT global disaster dataset, containing over 22,071 records with variables like disaster type, location, magnitude, and impacts, to predict disaster occurrence and severity. Three deep learning models Convolutional Long Short-Term Memory (ConvLSTM), Graph Neural Network (GNN), and Long Short-Term Memory (LSTM) were developed to capture spatial-temporal dynamics. The dataset was pre-processed to handle missing values, normalize features, and construct spatial graphs and temporal sequences. Exploratory data analysis (EDA) revealed patterns in disaster frequency, geographic hotspots, and temporal trends. On a held-out test set, the ConvLSTM model achieved the highest performance, with an AUROC of 98.78%, accuracy of 98.45%, Recall of 96.62%, Precision of 97.59% log loss of 0.0912 and F1-score of 97.10%, followed by LSTM and GNN. Visualizations, including spatial heatmaps, temporal prediction curves, confusion matrix and geospatial plots, enhance interpretability. These findings underscore the potential of specialized spatial and temporal models for disaster forecasting, supporting proactive mitigation strategies.