Two-Stage Deep Belief Network and Dynamic Adjustable Momentum Algorithm for Early Flood Prediction in Metropolitan Cities

Abstract

Urban flooding poses significant risks to human lives, property and infrastructure, disrupting essential services in densely populated metropolitan areas. Accurate and timely flood prediction models are crucial to mitigate these risks and enhance city resilience. This research highlights the potential of AI-based techniques in improving urban flood prediction accuracy and the importance of incorporating advanced learning algorithms in flood management systems. This study proposes a two-stage Deep Belief Network (DBN) and Dynamic Adjustable Momentum Algorithm (DAM) for early flood prediction in metropolitan cities. The first DBN model predicts rainfall intensity, while the second DBN model is activated when heavy rainfall is forecasted, estimating flood runoff. The proposed strategy incorporates the newly developed DAM algorithm, which combines adaptive momentum and backpropagation techniques to reduce training loss and enhance DBN accuracy. To demonstrate the efficacy of the DAM algorithm, the data is partitioned into various folds for training, testing, and validation. Experimental results show that the proposed DAM algorithm outperforms the traditional backpropagation method, yielding lower training errors in terms of Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). The MAE, MSE and RMSE for DAM were 0.69, 2.29 and 0.11, respectively, while for the traditional BP algorithm, they were 2.67, 10.75 and 0.38, respectively. The two-stage DBN and DAM approach offers a robust and efficient early flood prediction model for metropolitan cities, contributing to the development of reliable flood warning systems and enhancing urban resilience against flood-related disasters.