Federated Learning-Based Flood Forecasting Model Enhancing Predictive Accuracy and Privacy in Flood Prediction

Flooding is a common natural disaster that significantly impacts property, crops, the economy, and human lives. Researchers have long struggled with flood prediction due to the challenges of data privacy, availability, and security. This paper proposes a flood forecasting model using federated learning, a modern AI innovation that ensures data privacy, availability, and security. The model combines locally trained models of 18 customers, determines the flood station, and provides a five-day flood alarm for a specific client. A nearby feed forward brain organization (FFNN) model is prepared at the client station, using regional factors like snow liquefying, precipitation spillover, stream directing, and hydrodynamics to predict the predicted water level. The model has 84
^%
accuracy in predicting past floods in the selected zone between 2010 and 2015. The authors also use cutting-edge algorithms like Convolution2D Neural Network, which have a successful prediction accuracy of over 90%. The authors have extended CNN2D for flood forecasting to increase accuracy.