Low-Latency Urban Flood Forecasting and Early Warning System Based on Edge-Cloud Collaborative AI Architecture

Abstract

Infrastructure and overall community safety and economic stability of fast urbanizing areas are in jeopardy due to urban flooding. A new Edge-Cloud Collaborative AI Architecture is proposed in this paper, which integrates Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNN) and attention mechanisms to predict urban floods and provide early warnings. The suggested Edge-Cloud LSTM-Attention Network (ECLAN) takes advantage of distributed computing, by executing lightweight models on the edges (IoT sensors, ESP32 microcontrollers) to perform real-time local inferences and offload heavyweight deep learning models to the cloud resource to do a thorough analysis. Data preprocessing includes data of multi-source hydrological parameters, including rainfall intensity, water levels, soil moisture, and drainage capacity, measured by the LoRa and MQTT protocols. Emphasis of the time features is dynamically applied to the attention system to enhance the prediction quality. Real world experimental validation using urban data reveals improved performance: 96.8 accuracy, 95.3 precision, 94.7 recall, 95.0 F1-score, and 0.97 AUC-ROC compared to the standalone CNN, LSTM, Random Forest and SVM models (89.2, 91.5, 84.3 and 81.7 accuracy, respectively). The end-to-end latency of the edge-cloud design is 2.8 seconds, while cloud-only and edge-only designs are 8.5 seconds and 5.2 seconds, respectively. Results suggest that the accuracy of prediction and response time of collaborative edge-cloud architectures surpass the individual architectures, enabling an active mitigation of floods in accordance with SDG 11 and 13.