An LSTM approach to predict emergency events using spatial features

Abstract

With the global population on the rise, the frequency and severity of emergency events like fires and traffic accidents are becoming more frequent and severe. Attending to these emergencies demands valuable and limited resources, such as professionals and vehicles, so it is important to efficiently allocate them to regions that are more likely to require their services. However, the fact that emergencies can be related to spatial and temporal contexts makes resource allocation a highly complex task requiring specialized tools and techniques to exploit these relationships efficiently. This paper proposes an emergency event prediction solution using spatial segmentation and Long Short-Term Memory (LSTM) neural networks to model associations in space and time domains. We used data from real emergency occurrences in Florianópolis, Brazil, collected over five and a half years. Clustering algorithms combined with the silhouette metric were used to segment the time series in four different city regions. A comparison with traditional forecasting techniques and machine learning models showed that the LSTM network is consistent in its predictions and outperforms other approaches. Compared with a state-of-the-art reference employing LSTM, our solution leads to a 17.8% reduction in mean absolute error. Two methodologies for multi-step lookahead prediction are also presented and compared, showing that reusing the output of LSTM to predict future time steps is better than a full model retraining. To assess the generalizability of the model and proposed methodology, we applied the entire pipeline to new data from a different city. Our results demonstrate that models tailored to specific cities significantly outperform those trained on generalized datasets, highlighting the importance of localized training data.