Historical water fraction data and flood susceptibility mapping using deep learning model

다국어 초록 (Multilingual Abstract)

Flooding caused by continuous high intensity rainfall has caused severe damage to residents and property in parts of North Chungcheong Province, South Korea in July 2023 and Central Java Province, Indonesia in March 2024. Given the highly dynamic nature of flood disasters, rapid and timely spatial monitoring is considered essential for early disaster prevention, mid-term relief, and post-disaster reconstruction. In this study, inundation extent was detected using Global Navigation Satellite System-Reflectometry (GNSS-R) at L-band, with the Cyclone Global Navigation Satellite System (CYGNSS) being utilized for its high spatial (3 km) and temporal (3 days) resolution. The results demonstrate a strong agreement with independent remote sensing data, confirming the method's proficiency in rapidly updating the spatial extent of submerged areas. Despite its promise, this approach is subject to constraint by the inherent trade-off between spatial and temporal resolution due to pseudo-random sampling of CYGNSS observations, as well as uncertainties arising from simple threshold-based water classification, topography-related errors, and the use of previously observed behavior interpolation (POBI) surface reflectivity (SR) values rather than direct measurements. A long short-term memory neural network (LSTM) is proposed for flood susceptibility map prediction. Eight flood conditioning factors were considered in the analysis for developing the flood susceptibility map (FSM). The final FSM was categorized into five zones, ranging from very low, low, moderate, high and very high susceptibility. The experimental results demonstrate that satisfactory prediction performance was achieved by the LSTM model, with an area under the receiver operating characteristic curve (AUC) of 0.930 (Cheongju City, North Chungcheong Province, South Korea) and 0.914 (Kudus City, Central Java Province, Indonesia). The frequency ratio (FR) model was employed to analyze the relationship between historical flooded sites and environmental variables. The highest FR results for each factor's sub-classes (FR ≥ 1.0) in the two cities showed similarities that 0 - 5 degrees of slope, flat curvature, aspect of the field facing north, flat - lower slope of topographic positioning index, low power stream water of stream power index, less vegetation from normalized difference vegetation index, and crops area for land use – land cover type were crucial for flood exposure assessment. These findings are expected to provide critical information for environmental planners, policymakers, and stakeholders engaged in flood risk management to mitigate the adverse consequences of flash floods.

Flooding caused by continuous high intensity rainfall has caused severe damage to residents and property in parts of North Chungcheong Province, South Korea in July 2023 and Central Java Province, Indonesia in March 2024. Given the highly dynamic nature of flood disasters, rapid and timely spatial monitoring is considered essential for early disaster prevention, mid-term relief, and post-disaster reconstruction. In this study, inundation extent was detected using Global Navigation Satellite System-Reflectometry (GNSS-R) at L-band, with the Cyclone Global Navigation Satellite System (CYGNSS) being utilized for its high spatial (3 km) and temporal (3 days) resolution. The results demonstrate a strong agreement with independent remote sensing data, confirming the method's proficiency in rapidly updating the spatial extent of submerged areas. Despite its promise, this approach is subject to constraint by the inherent trade-off between spatial and temporal resolution due to pseudo-random sampling of CYGNSS observations, as well as uncertainties arising from simple threshold-based water classification, topography-related errors, and the use of previously observed behavior interpolation (POBI) surface reflectivity (SR) values rather than direct measurements. A long short-term memory neural network (LSTM) is proposed for flood susceptibility map prediction. Eight flood conditioning factors were considered in the analysis for developing the flood susceptibility map (FSM). The final FSM was categorized into five zones, ranging from very low, low, moderate, high and very high susceptibility. The experimental results demonstrate that satisfactory prediction performance was achieved by the LSTM model, with an area under the receiver operating characteristic curve (AUC) of 0.930 (Cheongju City, North Chungcheong Province, South Korea) and 0.914 (Kudus City, Central Java Province, Indonesia). The frequency ratio (FR) model was employed to analyze the relationship between historical flooded sites and environmental variables. The highest FR results for each factor's sub-classes (FR ≥ 1.0) in the two cities showed similarities that 0 - 5 degrees of slope, flat curvature, aspect of the field facing north, flat - lower slope of topographic positioning index, low power stream water of stream power index, less vegetation from normalized difference vegetation index, and crops area for land use – land cover type were crucial for flood exposure assessment. These findings are expected to provide critical information for environmental planners, policymakers, and stakeholders engaged in flood risk management to mitigate the adverse consequences of flash floods.

목차 (Table of Contents)

• Chapter 1. INTRODUCTION 1
• Chapter 2. MATERIALS AND METHODS 7
• 2.1 CYGNSS and flood label data 7
• 2.2 Flood conditioning factor data 8
• 2.3 Water Fraction Retrieval Method Based on CYGNSS Surface Reflectivity 13
• 2.4 Flood Susceptibility Map - LSTM model 19
• Chapter 3. RESULT AND DISCUSSION 25
• 3.1 CYGNSS flood 25
• 3.1.1 July 2023 Korean flood 25
• 3.1.2 March 2024 Indonesian flood 28
• 3.2 Flood susceptibility map (FSM) 33
• 3.2.1 FSM of Cheongju City 33
• 3.2.2 FSM of Kudus City 35
• 3.3 Model and environmental analysis 37
• Chapter 4. CONCLUSION 47
• Reference 50
• 논문요약 59