UNLEASHING THE POWER OF MULTI-SOURCE DATA FOR BUILDING ATTRIBUTE PREDICTION BASED ON DEEP LEARNING IN FLOOD RISK ASSESSMENT

Abstract

Knowledge of a structure’s physical attributes is critical for analyzing the risks associated with natural disasters such as floods and hurricanes. However, collecting these data over a wide geographical area is a difficult task because it often requires manual labor. Even in Louisiana, which is at considerable risk for flood-related natural disasters, the state government has sparse structure level data about the composition of building type, number of stories, foundation type, foundation height, presence of a basement, and presence of a garage, all of which are significant predictors of flood risk in many flood risk assessment methods. In this study, an automated model is proposed to predict these six attributes by utilizing a data fusion-based methodology. The model extracts feature from a building’s Google Street View (GSV) image through a Convolutional Neural Network (CNN) architecture. Likewise, an Artificial Neural Network (ANN) architecture is employed to extract features from community data taken at the census-block-group level. Another ANN is used to extract features from structure-level data obtained through a real estate database. These three feature streams are then fused and processed with a fusion module to predict a building’s attributes needed for flood risk assessment. Through this technique, accuracies close to or surpassing 90% are achieved on all five classification tasks. Likewise, the Mean Absolute Error (MAE) for foundation height estimation is small enough to make usable improvements to flood risk estimates over existing data sources and modeling assumptions. Although CNN and ANN models are used in this study, the basic framework can be applied with other machine learning models as well, if features can be extracted from them. Likewise, all the experiments are performed in the context of flood risk assessment in Louisiana. However, the framework can be easily extended to other natural disasters and geographical regions.