Flood risk assessment using multi-sensor remote sensing, geographic information system, 2D hydraulic and machine learning based models

University of Technology Sydney. Faculty of Engineering and Information Technology.Flooding events threaten the population, economy and environment worldwide. In recent years, several spatial methods have been developed to map flood susceptibility, hazard and risk for predicting and modelling flooding events. However, this research proposes multiple state-of-the-art approaches to assess, simulate and forecast flooding from recent satellite imagery. Firstly, a model was proposed to monitor changes in surface runoff and forecast future surface runoff on the basis of land use/land cover (LULC) and precipitation factors because the effects of precipitation and LULC dynamics have directly affected surface runoff and flooding events. Land transformation model (LTM) was used to detect the LULC changes. Moreover, an autoregressive integrated moving average (ARIMA) model was applied to analyse and forecast rainfall trends. The parameters of the ARIMA time series model were calibrated and fitted statistically to minimise prediction uncertainty through modern Taguchi method. Then, a GIS -based soil conservation service-curve number (SCS-CN) model was developed to simulate the maximum probable surface runoff. Results showed that deforestation and urbanisation have occurred upon a given time and have been predicted to increase. Furthermore, given negative changes in LULC, surface runoff increased and was forecasted to exceed gradually by 2020. In accordance with the implemented model calibration and accuracy assessment, the GIS-based SCS-CN combined with the LTM and ARIMA model is an efficient and accurate approach to detecting, monitoring and forecasting surface runoff. Secondly, a physical vulnerability assessment of flood was conducted by extracting detailed urban features from Worldview-3. Panchromatic sharpening in conjunction with atmospheric and topographic corrections was initially implemented to increase spatial resolution and reduce atmospheric distortion from satellite images. Dempster–Shafer (DS) fusion classifier was proposed in this part as a feature-based image analysis (FBIA) to extract urban complex objects. The DS-FBIA was investigated among two sites to examine the transferability of the proposed method. In addition, the DS-FBIA was compared with other common image analysis approaches (pixel- and object-based image analyses) to discover its accuracy and computational operating time. k-nearest neighbour, Bayes and support vector machine (SVM) classifiers were tested as pixel-based image analysis approaches, while decision tree classifier was examined as an object-based image analysis approach. The results showed improvements in detailed urban extraction obtained using the proposed FBIA with 92.2% overall accuracy and with high transferability from one site to another. Thirdly, an integrated model was developed for probability analysis of different types of flood using fully distributed GIS-based algorithms. These methods were applicable, particularly where annual monsoon rains trigger fluvial floods (FF) with pluvial flash flood (PFF) events occur simultaneously. A hydraulic 2D high-resolution sub-grid model of hydrologic engineering centre river analysis system was performed to simulate FF probability and hazard. Moreover, machine learning random forest (RF) method was used to model PFF probability and hazard. The RF was optimised by particle swarm optimisation (PSO) algorithm. Both models were verified and calibrated by cross validation and sensitivity analysis to create a coupled PFF– FF probability mapping. The results showed high accuracy in generating a coupled PFF–FF probability model that can discover the impact and contribution of each type to urban flood hazard. Furthermore, the results provided detailed flood information for urban managers to equip infrastructures, such as highways, roads and sewage network, actively. Fourthly, the risk of a flood can be assessed through different stages of flood probability, hazard and vulnerability. A total of 13 flood conditioning parameters were created to construct a geospatial database for flood probability estimation in two study areas. To estimate flood probability, five approaches, namely, logistic regression, frequency ratio (FR), SVM, analytical hierarchy process and combined FR–SVM, were adopted. Then, a flood risk map was generated by integrating flood hazard and vulnerability. The accuracy of flood probability indices indicated that the combined FR–SVM method achieved the highest accuracy among the other approaches. The reliability of the results obtained from this research was also verified in the field. The most effective parameters that would trigger flood occurrence were rainfall and flood inundation depth. In this research, transferable residency from one study area to another was verified through all the implemented methods. Therefore, the proposed approaches would be effectively and easily replicated in other regions with a similar climate condition, that condition that is, having a sufficient amount of flooding inventory events. Moreover, the results of the proposed approaches provided solid-detailed information that would be used for making favourable decisions to reduce and control future flood risks