Flood Hazard and Vulnerability Mapping using Deep Learning and Earth Observation Data
Abstract
Urban flood risk assessment is critical for safeguarding lives and infrastructure amid frequent floods. Recent advances in Earth Observation (EO) data enable the creation of flood risk maps with enhanced spatial and temporal resolutions. While traditional Machine Learning (ML) algorithms lack optimal feature selection, Deep Learning (DL) algorithms excel in extracting complex patterns from EO data. This dissertation delves into the estimation of flood hazard and vulnerability within urban environments through DL algorithms and EO data. Several innovative methodologies were proposed: 1) A Convolutional Siamese Network (CSN) was devised for urban flood mapping using SAR satellite imagery. This method employed two parallel Convolutional Neural Networks (CNNs), processing pre-event and co-event images, respectively. By measuring feature space similarity, pixels were classified as flood or background using Contrastive Loss, Weighted Double Margin Contrastive Loss (WDMCL), and Triplet Loss functions. Testing with VGG16 and ResNet50 architectures yielded Precision, Recall, and F1 Score values of 0.75, 0.6, and 0.67, respectively, for the SEN12-FLOOD dataset, which notably improved upon integrating DEM into input features. 2) First Floor Height (FFH) estimation based on vertical measurements from Google Street View (GSV) images facilitated the creation of a flood vulnerability map for Toronto's Lower Don River region. Utilizing the proposed vulnerability index, derived from water depth minus FFH, buildings were categorized by vulnerability levels. Additionally, First Floor Elevation was estimated using FFH and Lowest Adjacent Grade (LAG) heights, achieving RMSE and Bias values of 81 cm and −50 cm for the Greater Toronto Area (GTA) and 95 cm and −20 cm for Virginia. 3) A Dense Attention Network (DAN) CNN architecture was proposed for building footprint extraction from LiDAR and RapidEye images, demonstrating an improved F1 Score (0.71) over DL models like U-net (0.42) and ResUnet (0.49). 4) A fusion method combining CNN classifications from GSV, LiDAR, and Orthophoto data for building land-use type classification achieved superior accuracy indices compared to a previous CNN-based study, with an Overall Accuracy of 75%. These methodologies represent significant advancements in utilizing DL algorithms and EO data for urban flood risk assessment, promising enhanced accuracy and efficiency in mapping and mitigation efforts.