Lightweight Deep Learning-Based Multimodal Sensor Fusion for Disaster Detection in Smart Urban Infrastructure

Over the last few years, real-time disaster detection in smart environments has become crucial because of the increasing frequency of extreme weather events, complexity of urban infrastructure, and vulnerabilities. In particular, the existing Internet of Things (IoT) sensor-based method integrates fog computing with ensemble Machine Learning (ML) to detect flood risks in real-time. However, the existing method faced limitations, such as limited ability to handle multimodal sensor inputs, high latency due to cloud dependency, and lack of temporal pattern modeling for evolving disasters. Hence, to address these limitations, a novel multilayer framework, namely, DETECT-FusionNet, is proposed. Initially, the input data is collected from a heterogeneous IoT sensor implemented across an urban infrastructure. Next, multimodal sensor fusion is performed by applying a Kalman filter to integrate the data streams. The fused data is then preprocessed using Interquartile Range (IQR) to remove outliers and min-max normalization. Subsequently, the preprocessed data are forwarded to MobileNetv2, which is optimized for edge devices for device classification. Finally, a Temporal Convolutional Network (TCN) is incorporated to perform sequential modeling and final prediction. Therefore, the proposed DETECT-FusionN et achieved better performance than the existing model based on accuracy (99.7%), precision (99.5%), F1-score (99.6%), and area under curve (99.8%).