| Summary: | Flood is the most destructive natural disaster in Malaysia, causing significant economic and human life losses. Recent flood management paid more attention to flood risk assessment to provide information on risk of flooding. It consists of three components i.e., flood hazard, flood exposure, and flood vulnerability, which can be supported directly and indirectly by remotely sensed data. The research aims to utilise a geospatial approach for flood vulnerability assessment in the Kelantan River Basin (KRB), Malaysia. For flood hazard modelling, the flood event in December 2014 over Kelantan has been simulated using the Rainfall-Runoff-Inundation (RRI) model. The model is supported by four different Satellite Rainfall Products (SRPs) (i.e., Integrated Multi-satellitE Retrievals-Late, -Early (IMERG-L, -E), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Cloud Classification System (PERSIANN-CCS) and Global Satellite Mapping of Precipitation (GSMaP)) and rain gauge data. The simulation results were compared to the observed discharge flow. For the flood exposure analysis, a framework to extract detailed element-at-risk information was developed by classifying detailed land use and land cover in heterogeneous urban areas of Kota Bharu using Very-High-Resolution (VHR) satellite imagery and Light Detection and Ranging (LiDAR) data. The Feature Selection (FS) algorithm was also adopted to achieve an efficient classification framework. References points extracted from existing land use was used to validate the classification results. Finally, a geospatial approach for flood vulnerability assessment of buildings has been developed, which combines outputs from flood hazard and flood exposure. The framework incorporates relevant building parameters derived from the VHR satellite image and LiDAR data, and flood hazard for a vulnerability analysis of building using Machine Learning (ML) approach. The geospatial approach-based vulnerability results were validated using in-situ flood damage data derived from questionnaire method. The results of flood inundation showed that the GSMaP has the best performance in simulating hourly runoff with the lowest relative bias (RB) and the highest Nash-Sutcliffe efficiency (NSE) of 4.9% and 0.79, respectively. The results of image classification showed that there was a significant difference between classification accuracies using two datasets (50 features and 107 features) in object-based approach. The overall accuracy increased to 93.7% from 79.2% using the Random Forest (RF) classifier. Nevertheless, the RF classifier have achieved significantly better classification results compared to other classifiers (k-NN and Support Vector Machine (SVM)). The extreme gradient boosting (xgbTree) FS method improved the classification accuracy from 93.7% to 94.1% using 26 features However, there is a statistically significant difference between the results produced by the Recursive Eliminate Feature (RFE) and Simulated Annealing (SA). The result of vulnerability showed that the geospatial approach achieved a good prediction result with an overall accuracy of 81.8%. In conclusion, the SRPs can be used to support spatially distributed flood hazard assessment in a scarcity rain gauge station area. High resolution remotely sensed data can be used to extract information of element-at-risk in highly heterogenous location in Malaysia, which allows ML method for flood vulnerability assessment.
|
|---|