A Multi-Modal Data Fusion System for Flood Detection: A Privacy-Preserving Federated Learning Approach Leveraging Transfer Learning and Real-World Field Data
Year: 2026
Type: conference-paper
Abstract: One of the most devastating natural disasters is that of floods and this has necessitated early detection and rapid response systems which are usually constrained by the privacy threats, delays of communications and a lack of scalability of centralized data processing. The present paper suggests TriModalFloodNet, a federated multi mode network of data fusion based on privacy preservation in order to increase the accuracy and security of flood monitoring solution through local processing of data by a number of nodes. The architecture gets three different real-world field data streams, which are spatial visuals of the Sentinel-2 satellite image, temporal meteorological patterns of rainfall history, and dynamic hydrological reacts on the river water-level sensors. To learn these heterogeneous inputs, the framework makes use of specialized model branches, i.e., a CNN-based branch that visual processing, an all-MLP TSMixer that can analyse rainfall, and a Nonlinear Auto-Regressive with exogenous inputs (NARX) to predict water-level using the model. The system can be used to accomplish collaborative training in a heterogeneous environment with a variety of different clients, such as high-compute satellite centers and low-power IoT sensors, and at the same time keep raw data in a local and safe location by means of Heterogeneous Federated Learning (HeteroFL). The experimental findings reveal that TriModalFloodNet attains the highest accuracy of detection rate of 94 percent and that is much better than the traditional one modality contributions. The global model showed consistent convergence, and training loss decreased to around 0.10 at the 50 th epoch, which gave a very strong, scalable, and community-oriented solution to disaster management and its early warning features.
Institution: VIT Bhopal University
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7221136864