Quasi-Real time prediction of storm surge inundation for the coastal region of Bangladesh

Abstract

Bangladesh coast is infamous for the negative impacts caused by storm surge inundation that initially leads to damage of life and property but longer term devastation to coastal ecosystems affects the livelihoods years after the event. Considering the devastating impacts of cyclones several modeling approaches were developed and implemented in this region to understand the patterns of genesis and movement of TCs, forecasting peak surge height and associated coastal inundation. Alongside Bangladesh Meteorological Department (BMD) is capable of predicting cyclone track, intensity and peak surge in time for necessary actions to be taken before a storm strikes. But prediction of storm surge inundation on a real time basis is still under construction. Therefore with a great urgency to explore the options for inundation prediction over the coastal regions of Bangladesh this study aims to propose a feasible real-time storm surge inundation prediction system to that would add value to emergency planning and to frame a feasible warning system compatible to the BMD forecast. First a Delft 3D hydrodynamic model coupled with cyclone generating Delft Dashboard model has been applied over the coastal regions of Bangladesh to simulate the inundation due to 31 recent cyclones. The hydrodynamic model is calibrated with varying roughness conditions for land, river and ocean while the coupled cyclone model is validated through calibrating the tidal data for two major cyclones Sidr (2007) and Aila (2009) for multiple wind drag coefficient conditions. After verifying the model parameters and completing the simulations an inundation database is prepared using the simulated inundation maps and cyclone track and intensity data collected from global sources. The database is then applied to prepare a feedforward backpropagation Artificial Neural Network (ANN).To fit the inundation map in the neural network each map is provided a unique number based on their cyclonic characteristics and maximum inundation height. Subsequently the ANN model is trained with four input layers containing landfall latitude, longitude, maximum wind speed and maximum pressure drop information for all cyclonic event while the output layer consists of the respective inundation map number. The network parameters are chosen through trial and error approach so it can classify and predict the future inundation from the existing inundation maps for any cyclone with known track and intensity data. The results showed that the ANN model trained with one hidden layer and Levenberg-Marquardt training algorithm performs the best and provides almost accurate inundation predictions for four randomly selected cyclones from the database. Whilst there is ample scope for advancement and limitations are present in the prediction system, findings of the study demonstrated that with small database and proper combination of neural network parameters it is possible to predict inundation map for future cyclone with a given set of data which BMD provides for cyclones in the Bay of Bengal. Since real time inundation prediction with the help of numerical model is very challenging and still under research stage the study leaves a broad avenue for the advancement of inundation prediction system and generate more accurate results. It is also suggested that the ANN model can be dynamic through upgrading the database with more parameters and cyclonic events that in future this approach might be able to produce better prediction for future cyclones. To this end, beginning of real time storm surge inundation prediction system will not only increase the lead time but also fit to an integral part of a holistic disaster management framework that allow the policy makers to devise long term plans for a better coastal zone management.