[Paper Review] Improving Urban Flood Prediction using LSTM-DeepLabv3+ and Bayesian Optimization with Spatiotemporal feature fusion

Zuxiang Situ, Qi Wang+5 more|arXiv (Cornell University)|Apr 19, 2023

Flood Risk Assessment and Management13 citations

TL;DR

This paper proposes a CNN-RNN hybrid feature fusion model for urban flood prediction, identifying influential factors via Bayesian optimization and selecting an optimal LSTM-DeepLabv3+ architecture for accurate and fast predictions.

ABSTRACT

Deep learning models have become increasingly popular for flood prediction due to their superior accuracy and efficiency compared to traditional methods. However, current machine learning methods often rely on separate spatial or temporal feature analysis and have limitations on the types, number, and dimensions of input data. This study presented a CNN-RNN hybrid feature fusion modelling approach for urban flood prediction, which integrated the strengths of CNNs in processing spatial features and RNNs in analyzing different dimensions of time sequences. This approach allowed for both static and dynamic flood predictions. Bayesian optimization was applied to identify the seven most influential flood-driven factors and determine the best combination strategy. By combining four CNNs (FCN, UNet, SegNet, DeepLabv3+) and three RNNs (LSTM, BiLSTM, GRU), the optimal hybrid model was identified as LSTM-DeepLabv3+. This model achieved the highest prediction accuracy (MAE, RMSE, NSE, and KGE were 0.007, 0.025, 0.973 and 0.755, respectively) under various rainfall input conditions. Additionally, the processing speed was significantly improved, with an inference time of 1.158s (approximately 1/125 of the traditional computation time) compared to the physically-based models.

Motivation & Objective

• Motivate improved urban flood prediction with integrated spatial and temporal features.
• Develop a hybrid CNN-RNN framework to fuse static and dynamic flood-related features.
• Use Bayesian optimization to identify the most influential flood drivers and the best feature/model combination.
• Assess multiple CNN backbones and RNN architectures to find an optimal LSTM-DeepLabv3+ configuration.
• Demonstrate improved prediction accuracy and faster inference compared to traditional models.

Proposed method

• Combine four CNNs (FCN, UNet, SegNet, DeepLabv3+) with three RNNs (LSTM, BiLSTM, GRU) to form a hybrid feature fusion model.
• Apply Bayesian optimization to select the seven most influential flood-driven factors and the best combination strategy.
• Identify the optimal model as LSTM-DeepLabv3+.
• Evaluate performance under various rainfall input conditions using MAE, RMSE, NSE, and KGE as metrics.
• Measure inference speed and compare against physically-based models.

Experimental results

Research questions

• RQ1 Which spatiotemporal feature fusion strategy yields the best flood prediction accuracy and speed?
• RQ2 What combination of CNN backbones and RNN types best captures urban flood dynamics?
• RQ3 Which seven flood-driven factors are most influential for prediction under Bayesian optimization?
• RQ4 How does the LSTM-DeepLabv3+ model perform relative to other hybrid CNN-RNN configurations under varying rainfall inputs.

Key findings

• LSTM-DeepLabv3+ achieved the highest prediction accuracy among tested hybrids.
• Under various rainfall inputs, the model attained MAE 0.007, RMSE 0.025, NSE 0.973, and KGE 0.755.
• Inference time was 1.158 s, roughly 1/125 of traditional computation time.
• Bayesian optimization identified seven influential flood-driven factors and the optimal combination strategy.
• The approach provides both static and dynamic flood predictions with improved efficiency compared to physically-based models.