Anticipating a Risky Future: Deep Neural Networks for Predictive Exposure Modeling

Kurzfassung

Anticipating future geospatial population distribution is essential for numerous applications such as hazard risk assessment, accessibility detection, and other health, economic, and environment-related fields. Existing population forecasts are provided country- or district wise and are thus not suitable for an exposure analysis on a city level. In this study, freely available population grids are utilized together with machine learning models, that were tailored for time series analysis. In detail, Long Short-Term Memory (LSTM) and Convolutional LSTM (ConvLSTM) networks are trained with WorldPop and ancillary geospatial data to forecast population on a three-year interval for a hazard exposure analysis. The experimental setup is conducted for Peru‘s capital Lima, which features a high population dynamic and a strong seismic activity. To assess the competitive performance of LSTM models for this application, also a Multivariate Linear Regression and Random Forest Regression are implemented for comparison. To gain insight into the influence of different driving factors for population dynamics and the influence of the number of input years on the prediction, the models are trained with different sets of input features and time steps. The results suggest that the LSTM network outperforms the ConvLSTM and the benchmark methods, specifically in terms of lower Root Mean Squared Error (RMSE) and better learning of spatial characteristics. Regions of strong population change are captured much better by the LSTM than by the base models. The included driving factors don’t improve the results. On the contrary, models with the population as the only input result in the lowest median absolute error. With the forecasted population grids and a modeled hazard scenario, it was found that the population in the tsunami inundation area is expected to increase by 61% until 2035 and that 70% of the additional population of Lima will accumulate in areas of high earthquake intensities. This comparative analysis underlines the usefulness of the LSTM networks for forecasting gridded population data for exposure assessment.