Evaluation and Analysis of Transfer Learning Models Towards the Prediction of Flood

Abstract

This study presents a comparative analysis of three transfer learning-based models such as EfficientNet, Vision Transformer (ViT), and ResNetfor predicting pluvial flood. A flood dataset comprising 144,401 records with eight key conditioning variables was collected from Kaggle repository organized by the United States Geological Survey (USGS) and Copernicus Climate Data Store and was further used for the implementation of this study.Additionally, historical rainfall and meteorological data were obtained from the Nigerian Meteorological Agency (NiMet) through their official data request portal.Subsequently, the dataset was pre-processed by cleaning and normalizing, transforming features and augmenting them, and partitioned into training, validation and testing sets. All the models were pretrained on ImageNet weights and trained to learn flood-specific spatial patterns. As the experimental findings indicate, ViT has the best accuracy (93.1%), F1-score (0.925), and AUC-ROC (0.95) that are used to capture long-range spatial dependencies. EfficientNet was more accurate (92.3) and had the highest F1-score of 0.915; however, it took the least amount of time to be trained, which is acceptable in terms of real-time use. ResNet obtained 91.5% accuracy and 0.905 F1-score, showing stable feature acquisition at a modest computational price. The paper shows the success of transfer learning in improving the flood prediction in low-data areas. Generally, ViT should be used in the context of high-accuracy, EfficientNet in the context of computational efficiency, and ResNet in the context of robust and reliable modeling. These results help to justify the creation of AI-based flood early warning systems to enhance urban flood risk management.