Improving urban flood prediction using LSTM-DeepLabv3+ and Bayesian optimization with spatiotemporal feature fusion

Abstract

Deep learning models have become increasingly popular for flood prediction due to their superior accuracy and efficiency compared to traditional methods. However, current models often rely on separate spatial or temporal feature analysis and have limitations on the types, numbers, and dimensions of input data. This study proposes a novel framework to combine the strengths of Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) by connecting the output of RNN to the deepest part of CNN (i.e., the layer with the richest features). The innovative spatiotemporal feature fusion method is developed to strategically integrate the temporal (e.g., rainfall and flood series) and spatial driving factors (e.g., DEM, imperviousness, drainage network, and their related features). The framework focuses on three critical problems: the identification of key driving factors, the design of hybrid deep learning models, and problem formulation and associated optimization algorithms. We verified the framework through a case study in North China. Bayesian optimization was first applied to identify the seven most influential factors and determine their best combination strategy as the model inputs. Then, the optimal hybrid model LSTM-DeepLabv3+ was identified from 12 model combinations and achieved high prediction accuracies in terms of Mean Absolute Error, Root Mean Square Error, Nash-Sutcliffe Efficiency, and Kling-Gupta Efficiency of 0.0071, 0.0253, 0.9730, and 0.7549 under various rainfall conditions. This study demonstrates that the new framework provides effective hybrid models with significantly improved computational efficiency (about 1/125 of the traditional process-based computation time) and offers a promising solution for real-time urban flood prediction.