Search

Advances in the availability of multi-sensor, remote sensing-derived datasets, and machine learning algorithms can now provide an unprecedented possibility to predict flood events and risk. Therefore, this study was undertaken to develop a flood vulnerability map and to assess the exposure of buildings to flood risk in Warsaw, the capital of Poland. This goal was pursued in four research phases. The thirteen flood predictors were evaluated using information gain ratio (IGR), and finally reduced to eight of the most causative ones and used for flood vulnerability mapping with three machine learning algorithms, Artificial Neural Network Multi-Layer Perceptron (ANN/MLP), Deep Learning Neural Network based approach—DL4j (DLNN-DL4j) and Bayesian Logistic Regression (BLR). These algorithms show a good predictive performance with the receiver operating curve (ROC) value of 0.851, 0.877 and 0.697, respectively. The buildings' exposure to flood was assessed in line with criteria established in European and national legal regulations. The introduced new buildings' flood hazard index (BFH) revealed a significant similarity of potential flood risk for both models, highlighting the greatest risk in zones with high vulnerability to flooding. Depending on the method used, the BFH value was 0.54 (ANN), 0.52 (DLNNs) or 0.64 (BLR). The holistic approach proposed in this study could assist local authorities in improving flood management. [ABSTRACT FROM AUTHOR]

Flood is considered as the most devastating natural hazards that cause the death of many lives worldwide. The present study aimed to predict flood vulnerability for Warsaw, Poland, using three machine learning models, such as the Bayesian logistic regression (BLR), the artificial neural networks (ANN), and the deep learning neural networks (DLNNs). The perfomance of these three methods was assessed in order to select the best method for flood vulnerability mapping in densely urbanized city. Thus, initially, thirteen flood predictors were evaluated using the information gain ratio (IGR), and eight most important predictors were considered from model training and testing. The performance of the applied models and accuracy of the result was evaluated through the area under the curve (AUC) and statistical measures. By using the testing dataset, the result reveals that DLNN (AUC = 0.877) is the more performant model in comparison to ANN (AUC = 0.851) and BLR (AUC = 0.697). However, the BLR model has the lowest predictive capability. The results of the present study could be effectively used for the urban flood management strategies. [ABSTRACT FROM AUTHOR]

The provision of detailed information on the impact of potential fluvial floods on urban population health, quantifying the impact magnitude and supplying the location of areas of the highest risk to human health, is an important step towards (a) improvement of sustainable measures to minimise the impact of floods, e.g., by including flood risk as a design parameter for urban planning, and (b) increase public awareness of flood risks. The three new measures of the impact of floods on the urban population have been proposed, considering both deterministic and stochastic aspects. The impact was determined in relation to the building's function, the number of residents, the probability of flood occurrence and the likely floodwater inundation level. The building capacity concept was introduced to model population data at the building level. Its proposed estimation method, an offshoot of the volumetric method, has proved to be successful in the challenging study area, characterised by a high diversity of buildings in terms of their function, size and density. The results show that 2.35% of buildings and over 122,000 people may be affected by 500-year flooding. However, the foreseen magnitude of flood impact on human health is moderate, i.e., on average ten persons per residential building over the 80% of flood risk zones. Such results are attributed to the low inundation depth, i.e., below 1 m. [ABSTRACT FROM AUTHOR]