Flood Risk Assessment using Visual Features by Transfer Learning of Pretrained Algorithm

Abstract

Flood risk assessment is a critical component in mitigating the impacts of natural disasters, particularly in vulnerable regions. With the growing availability of remote sensing data and advances in computer vision, deep learning techniques offer significant potential for accurate flood detection and risk analysis. This study presents a novel approach for flood risk assessment using visual features extracted from satellite imagery through transfer learning with pre-trained convolutional neural networks (CNNs). Five popular CNN architectures. VGG16, DenseNet, ResNet50, RegNetY320 and SqueezeNet1_0 were evaluated for their performance in classifying flood-affected areas. The models were fine-tuned on annotated flood datasets and were assessed based on accuracy, precision, latency and their architectural strengths and limitations. The results highlight DenseNet and ResNet50 as the top-performing models, both achieving approximately 96% accuracy with high precision and low latency. VGG16 also performed well (92–94% accuracy), though at the cost of higher computational requirements. RegNetY320 demonstrated moderate accuracy (80–82%) but offered flexibility and low latency, making it suitable for scalable applications. SqueezeNet1_0, while extremely efficient in terms of computation and suitable for deployment on low-power devices, exhibited the lowest accuracy (57–58%) and is less ideal for high-stakes flood risk assessment. The study underscores the importance of selecting an appropriate architecture based on the operational environment, data availability and resource constraints. Overall, transfer learning using visual features from pre-trained models provides a scalable, accurate and efficient solution for flood risk mapping and early warning systems.