Enhanced Distributed Flood Detection and Alert System Using Deep Learning

Abstract: Among natural disasters, flood has greater impact on farming, property, human lives and their money situation, which further affects the overall economy. We are seeing that AI alert systems for floods are only necessary to reduce the damage. Basically, this paper proposes the Federated Learning approach for flood prediction model. Federated Learning itself is a distributed ML technique that further reduces data transfer from flood sites to the central server. As per the system design, FL can give security and privacy to data and keep it available always. Regarding data protection, this method ensures information stays safe and accessible at all times. Moreover, in FL based flood prediction model, each location trains its individual model using local flood data, and this trained model is further shared with the server itself. The server combines all local models to prepare a global model further. This process itself creates a unified model from individual contributions. We are seeing this method can avoid network delays only, so the global model can take quick decisions on floods. As per this paper, the server combines local models to make a global model regarding five- day flood warning plans for each place. We are seeing that a local model is trained using regional conditions to give predictions on possible flood areas and their maximum water levels only. The data set actually has flood history information from five rivers between 2015 to 2021, which is used for training the model. Basically, it includes four features - rainfall overflow, snow melting rate, water movement dynamics, and current river flow, which are the important factors for water analysis. Basically, the results show flood prediction data collected from 2010 to 2015 for the selected area using the proposed method with 84% accuracy. The model is further improved by adding a Convolutional 2D Neural Network (CNN2D) itself. The improved version of our proposed method can enhance flood predictions with 90% accuracy. Moreover, this approach provides reliable results for better flood forecasting.

Keywords: Feedforward Neural Network (FFNN), Random Forest, Decision Tree, Gradient Boost, Convolutional Neural Network (CNN).