A hybrid convolutional neural network model coupled with AdaBoost regressor for flood mapping using geotagged flood photographs

Flood mapping has been crucial to flood hazard identification, mitigation, and preparedness. Development of accurate flood maps in data-scarce regions has always been a challenge. This study presents a multi-stage approach involving topographic feature extraction, CNN-based flood classification, and regression modeling for flood mapping. The study aims to provide accurate estimates of flood depth by leveraging both spatial data and image-based contextual information. The methodology comprises two CNN models followed by a regressor for estimating flood depth using geo-tagged flood photographs. The first CNN model identifies the existence of flood, while the second one identifies the flood severity class (greater than or less than 1 m depth). A modified VGG-16 CNN architecture is employed in the present study for both stages. Finally, the AdaBoost Regressor is employed to estimate precise flood depth using topographical data such as elevation, slope, and topographic position index (TPI) values as the input. The model results showed excellent performance with R2 of 0.93 and RMSE of 25.01% when tested on manually collected flood data for VGP Selva Nagar, a residential area in Chennai, India, for the December 2023 flood. Comparison of the VGG-16 CNN architecture with other standard architectures, such as ResNet50 and InceptionV3, showed the efficacy of the presented model. The presented multi-staged approach, thus, proves to be an effective tool that relies only on geo-tagged flood photographs as the input to develop accurate flood maps. The models developed in this study have significant implications for flood management that can help inform emergency response teams about flood severity and extent, facilitating prompt and effective interventions.