Artificial Intelligence and Copula-Probabilistic Models for Early Flood Warning and Community Risk Management: Case Studies in Fiji Islands

Artificial Intelligence and Copula-Probabilistic Models for Early Flood Warning and Community Risk Management: Case Studies in Fiji Islands

Flooding is one of the most prevalent natural hazards, impacting numerous regions worldwide. The repercussions of such disasters are especially severe in developing nations, particularly in small island countries like Fiji. The absence of advanced flood risk monitoring resources and relevant data in these developing nations presents significant challenges to implementing effective early flood warning systems. To address this issue, this research develops innovative flood monitoring, assessment, and forecasting tools using artificial intelligence (AI) and copulastatistical methods to enhance and contribute to developing effective early flood warning systems, thereby assisting in better flood preparation and management strategies to mitigate the severe impacts of flooding. The first objective is to develop a novel hourly flood monitoring index (SWRI24-hr-S) to identify flood events and compute their associated characteristics, including flood volume (V), duration (D) and peak (Q). The feasibility of this index as an hourly flood risk monitoring tool is demonstrated for various flood-prone sites in Fiji. The 3-dimensional (3D) vine copula model is employed to model the joint distribution between D, V, and Q to extract their joint exceedance probability for probabilistic flood risk assessment across these study sites. In the second objective, a hybrid deep learning model (C-GRU) is designed by fusing the Convolutional Neural Network (CNN) with the Gated Recurrent Unit (GRU) model to forecast the proposed SWRI24-hr-S over a short-term (i.e., 1-hourly forecast horizon) to assess the future flood risk for five flood-prone study sites in Fiji. The objective model is trained using the statistically significant lagged SWRI24-hr-S and realtime hourly rainfall data. The state-of-the-art models, i.e., CNN, GRU, Long Short- Term Memory (LSTM), and Random Forest Regression (RFR), were also developed for benchmarking. The hyperparameters of the objective and benchmarking models were optimised using the efficient Bayesian Optimization (BO) technique. Overall, the outcomes of this research are expected to assist Fiji and other flood-prone regions worldwide in enhancing their existing early flood warning systems by integrating the flood monitoring, assessment, and forecasting tools developed in this study. This integration will enhance the decision-support framework for flood preparedness and response efforts, thus mitigating the severe impacts of flooding and improving community risk management.

File reproduced in accordance with the copyright policy of the publisher/author/creator.