Application of deep learning technique in flood management the university of south alabama

--- Page 1 --- Application of Deep Learning Technique in Flood Management Cuong T. Pham Research Interest Paper The University of South Alabama Doctoral Advisor - Dr. Trung Do July 23, 2024 INTRODUCTION Floods are one of the most common natural disasters, occurring on average 163 times each year , causing extensive damage to infrastructure, displacement of populations, and loss of life . During the 20th century, floods caused a significant loss of approximately 7 million lives globally . From 2000 to 2019, floods constituted 44% of all disaster events, impacting 1.6 billion individuals globally, the largest among all types of disasters. Annually, they cause the loss of around 20,000 lives and hurt at least 20 million individuals globally . Many nations have considered and applied various strategies for mitigating and adapting to the damage caused by floods. These strategies include implementing flood control infrastructure, early warning systems, insurance coverage, land use planning, and enforcing construction codes . As climate change increases flood frequency and severity, accurate flood management strategies help communities build resilient infrastructure ​​. Additionally, flood risk data provides information for policy-making and urban planning, guiding the development of flood defenses and land-use practices . --- Page 2 --- In recent years, with the quick development of science and technology, the artificial intelligence (AI)-based approach has emerged as a potential and effective approach for researchers in flood assessment and management. AI is the technology that allows computers and machines to imitate human intelligence and problem-solving abilities . It enables computers to engage in cognitive processes such as acquiring knowledge, logical thinking, finding solutions for problems, and understanding language . Machine Learning (ML) is an AI discipline that uses data and algorithms to enable AI to mimic how people learn with gradually improved accuracy (Jordan & Mitchell, 2015). ML techniques, such as supervised learning, unsupervised learning, and reinforcement learning, are widely used in flood prediction and risk assessment. ML provides high accuracy in predicting flood events due to learning from historical and real-time data . The versatility of ML enables it to use various data sources and integrate with other techniques to enhance the capabilities of flood risk assessment . Deep Learning (DL) is an in-depth branch of ML that employs deep neural networks, which are neural networks with multiple layers, to model the complex cognitive abilities of the human brain . DL models are good at analyzing spatial and temporal data, like satellite images and rainfall or land-use patterns. This leads to high accuracy in flood management. DL eliminates the need for manual involvement in feature extraction and data processing, resulting in more efficient and effective analysis. (a) (b) Fig. 1. Number of publications in the researched field in 10 recent years, from 2014 to 2023: (a) in each year and (b) from different countries. The number of publications over the years and countries in applications of DL in flood mapping is shown in Fig. 1. Geographically, the research contributions are most prevalent in China and the United States, countries with highly developed AI technologies that suffer significant flooding damage yearly. This concentration of studies underscores the critical need and capacity for advanced flood management solutions in these regions. Also, the increase in publications from developing countries shows a rise in AI applications in flood-prone areas, highlighting the need for global collaboration and knowledge sharing. PURPOSE STATEMENT AND RESEARCH QUESTIONS The purpose of this study is to understand and develop DL techniques for flood management. The aim is to enhance the accuracy and reliability of flood risk assessments by integrating available data sources with an advanced DL model. A case study will be conducted for Dauphin Island, Alabama to demonstrate this technique''''s practical application and effectiveness. Some major questions of the research are: 1. How can DL techniques improve the accuracy of management? 2. What are the benefits of DL in collecting and processing data from sources for flood susceptibility mapping? 3. What are the challenges and opportunities in implementing DL techniques in flood management strategies? 4. How can DL techniques be applied specifically to manage flood risks on Dauphin Island? The first question will be answered by the hypothesis that advanced DL techniques provide more accurate predictions and outcomes than traditional methods do. LITERATURE REVIEW Flood Management

Application of Deep Learning Technique in Flood Management

Cuong T Pham Research Interest Paper The University of South Alabama

Doctoral Advisor - Dr Trung Do

July 23, 2024

1 INTRODUCTION

Floods are one of the most common natural disasters, occurring on average 163 times each year (UNDRR, 2020), causing extensive damage to infrastructure, displacement of populations, and loss of life (Doocy et al., 2013) During the 20th century, floods caused a significant loss of approximately 7 million lives globally (Doocy et al., 2013) From 2000 to

2019, floods constituted 44% of all disaster events, impacting 1.6 billion individuals globally, the largest among all types of disasters Annually, they cause the loss of around 20,000 lives and hurt at least 20 million individuals globally (Kellens et al., 2013)

Many nations have considered and applied various strategies for mitigating and adapting

to the damage caused by floods These strategies include implementing flood control infrastructure, early warning systems, insurance coverage, land use planning, and enforcing construction codes (Hegger et al., 2016) As climate change increases flood frequency and severity, accurate flood management strategies help communities build resilient infrastructure (Takin et al., 2023) Additionally, flood risk data provides information for policy-making and urban planning, guiding the development of flood defenses and land-use practices (Rezende et al., 2019)

In recent years, with the quick development of science and technology, the artificial intelligence (AI)-based approach has emerged as a potential and effective approach for researchers in flood assessment and management AI is the technology that allows computers and machines to imitate human intelligence and problem-solving abilities (Lucci et al., n.d.) It enables computers to engage in cognitive processes such as acquiring knowledge, logical thinking, finding solutions for problems, and understanding language (Norvig & Russel, 2020) Machine Learning (ML) is an AI discipline that uses data and algorithms to enable AI to mimic how people learn with gradually improved accuracy (Jordan & Mitchell, 2015) ML techniques, such as supervised learning, unsupervised learning, and reinforcement learning, are widely used

in flood prediction and risk assessment ML provides high accuracy in predicting flood events due to learning from historical and real-time data (Mosavi et al., 2018) The versatility of ML enables it to use various data sources and integrate with other techniques to enhance the capabilities of flood risk assessment (Al-Areeq et al., 2022) Deep Learning (DL) is an in-depth branch of ML that employs deep neural networks, which are neural networks with multiple

layers, to model the complex cognitive abilities of the human brain (Lecun et al., 2015) DL models are good at analyzing spatial and temporal data, like satellite images and rainfall or land-use patterns This leads to high accuracy in flood management DL eliminates the need for manual involvement in feature extraction and data processing, resulting in more efficient and effective analysis

Fig 1 Number of publications in the researched field in 10 recent years, from 2014 to 2023: (a)

in each year and (b) from different countries

The number of publications over the years and countries in applications of DL in flood

mapping is shown in Fig 1 Geographically, the research contributions are most prevalent in

China and the United States, countries with highly developed AI technologies that suffer significant flooding damage yearly This concentration of studies underscores the critical need and capacity for advanced flood management solutions in these regions Also, the increase in publications from developing countries shows a rise in AI applications in flood-prone areas, highlighting the need for global collaboration and knowledge sharing

2 PURPOSE STATEMENT AND RESEARCH QUESTIONS

The purpose of this study is to understand and develop DL techniques for flood management The aim is to enhance the accuracy and reliability of flood risk assessments by integrating available data sources with an advanced DL model A case study will be conducted

for Dauphin Island, Alabama to demonstrate this technique's practical application and effectiveness Some major questions of the research are:

1 How can DL techniques improve the accuracy of management?

2 What are the benefits of DL in collecting and processing data from sources for flood susceptibility mapping?

3 What are the challenges and opportunities in implementing DL techniques in flood management strategies?

4 How can DL techniques be applied specifically to manage flood risks on Dauphin Island?

The first question will be answered by the hypothesis that advanced DL techniques provide more accurate predictions and outcomes than traditional methods do

3 LITERATURE REVIEW

3.1 Flood Management

Flood management includes a series of phases designed to effectively address the different stages of flood events These phases are mitigation, preparation, response, and recovery, as shown in Fig 2 We implemented flood mitigation and preparation before the flood event and applied response and recovery stages during and after the disaster Each phase plays a crucial role in minimizing the impact of floods and ensuring a coordinated response

Fig 2 Phases of flood management.

Flood mitigation involves management actions to reduce the long-term risk, impact, and loss caused by floods (Gougelet, 2016) Examples of mitigation activities include mapping the flood-prone areas, constructing and analyzing infrastructures such as levees, dams, floodways, and other barriers to control flooding, analyzing and applying strategies for mitigation, and educating the general public on the risk of floods

Flood preparation, or flood preparedness, includes those activities and measures that would ensure adequate response mechanisms before a flood event is likely to occur The phase encompasses preparing and maintaining early warning systems, displaying and disseminating notices and information to alert communities to the dangers of flooding and emergency response, stocking necessary supplies, and planning evacuation for a flood event

Flood response comprises the short-term actions taken during and after a flood to protect life and property The process includes implementing disaster response plans and taking search and rescue action; providing evacuation, emergency housing, and first aid to the victims; and coordinating the activities of local and national agencies to ensure a comprehensive response

The recovery phase focuses on making life return to normal when the floodwaters recede The recovery phase involves damage and impact assessment, physical restoration, the reconstruction of infrastructure, and financial and community resilience programs Depending on the resources available for recovery, it can be short-term or long-term

3.2 Deep Learning (DL)

Fig 3 Relationship between Artificial Intelligence, Machine Learning, and Deep Learning.

DL is a powerful branch of ML that employs neural networks with multiple layers to model complex patterns and relationships within data (Lecun et al., 2015) The relationship

between DL, ML and AI is shown in Fig 3 The DL techniques can be categorized into several groups based on structure and application Deep neural networks (DNNs) are ML algorithms similar to the artificial neural network but have multiple layers between the input and output layers (Samek et al., 2021) Convolutional neural networks (CNNs) are one of the most widespread models of DNNs, which capture local dependencies and spatial hierarchies in data (Li et al., 2022) These networks are highly effective for tasks involving spatial data, such as image recognition, classification, and semantic segmentation Some popular variants of CNNs are fully convolutional networks (FCNs), U-Net, residual neural networks (ResNets), and VGGNets (He et al., 2015; Long et al., 2015; Ronneberger et al., 2015; Simonyan & Zisserman, 2014) Recurrent neural networks (RNNs) are used to handle sequential data, making them suitable for time-series forecasting, natural language processing, image captioning, and speech recognition (Salehinejad et al., 2017) They can capture temporal dependencies and patterns in sequences of data Long short-term memory networks (LSTMs) are a special type of RNNs, which can learn long-term dependencies (Cheng et al., 2016) They are particularly effective for tasks where it is essential to remember information over long sequences Multi-layer perceptrons (MLPs) are networks that consist of multiple layers of neurons using nonlinear activation functions to learn complex patterns in data like nonlinear relationships (Kruse et al., 2022) Group method of data handling (GMDH) is a set of algorithms for different problem solutions, including parametric, clusterization, analogues complexing, rebinarization and probability algorithms (Madala & Ivakhnenko, 2019) Autoencoder is a type of ANN used to learn efficient codings of unlabeled data, typically for dimensionality reduction (Li et al., 2023) Overall, these

DL models excel at addressing complex patterns in data for many applications, with potential impact across scientific and industrial domains

3.3 Deep Learning Techniques in Flood Management:

3.3.1 Deep Learning in Flood Mitigation

Key applications of AI in flood mitigation include flood susceptibility mapping, infrastructure analysis, and mitigation strategy analysis

Flood susceptibility mapping is a critical part of flood mitigation, aimed at identifying areas most at risk of flooding This process involves using various techniques to analyze and

predict flood-prone zones based on environmental and geographical factors DL methods, which leverage complex neural network architectures, have been a potential approach in predicting and mapping flood Muñoz et al (2021) combined CNNs with data fusion for regional flood mapping Bonafilia et al (2020) applied CNNs to the Sen1Floods11 dataset for training and testing flood algorithms on Sentinel-1 data, aiming at real-time disaster response Xing et al (2023) integrated high-resolution remote sensing and street view images with CNNs for assessing urban flood vulnerability Ghobadi & Ahmadipari (2024) created a hybrid model of DNNs and statistical learning to predict flood-prone areas Pham et al (2021) utilized DNNs and multi-criteria decision analysis for broad flood risk assessments Fang et al (2021) used LSTM networks to analyze temporal data for predicting flood susceptibility Mia et al (2022) incorporated LSTMs with blockchain for sustainable flood risk assessment Ahmadlou et al (2021) combined autoencoders to enhance mapping accuracy in a novel DL model Hybrid models that combine DL with other ML techniques are also common Costache et al (2020) developed a novel DNNs and statistical learning ensemble for flash flood susceptibility mapping Ramayanti et al (2022) compared the performance of different DL models for flood susceptibility in Mozambique, demonstrating the benefits of integrating various approaches NER, a technique of NPL, was applied to extract information about historical events to build a flood susceptibility map (Lai et al., 2022; Nasution et al., 2022)

Infrastructure analysis is another critical part of flood mitigation, aimed at assessing the resilience and vulnerability of critical structures such as dams, levees, and urban infrastructure

DL techniques have significantly enhanced the accuracy and reliability of these assessments by analyzing large datasets and identifying risk factors Khandel & Soliman (2021) developed an integrated framework using DNNs to assess the time-variant flood fragility of bridges From the flood susceptibility mapping and infrastructure analysis, mitigation strategies might be developed and evaluated using AI techniques For example, Puttinaovarat & Horkaew (2020) developed an internetworking flood disaster mitigation system that integrates remote sensing, mobile GIS, and DL to enhance the accuracy and timeliness of flood mitigation measures

In conclusion, while much research is dedicated to flood susceptibility mapping using DL technologies, there is a relative scarcity of studies leveraging DL in infrastructure analysis and analyzing mitigation strategies

3.3.2 Deep Learning in Flood Preparation

Flood preparation involves activities designed to predict and respond to potential flood events, aiming to minimize their impact and ensure swift, effective responses It includes a range

of strategies, such as early warning systems, real-time disaster prediction and detection, and evacuation planning However, DL was only utilized in real-time disaster prediction and detection projects In these projects, DL improved prediction speed and accuracy by using advanced algorithms to process real-time data for actionable understanding Bonafilia et al (2020) introduced the Sen1Floods11 dataset, georeferenced to train and test DL algorithms for Sentinel-1 satellite imagery This dataset improves real-time flood detection by supplying high-quality training data for DL models Hernández et al (2022) developed a flood detection system leveraging CNN to analyze images from UAVs on an edge-computing platform for quick detection and response Wang et al (2020) utilized NPL and DL to monitor flooding phase transitions and establish a passive hotline through social media data analysis, providing timely flood condition insights Zhang et al (2021) applied NER and other NLP techniques to identify urban flooding locations and assessed semantic risk using social sensing data Barker & Macleod (2019) created a real-time Twitter data mining pipeline to analyze flooding impacts and aid disaster response efforts Baldazo et al (2019) utilized decentralized multi-agent DL in drones for flood monitoring, implementing deep Q-networks to coordinate multiple drones and enhance real-time flood detection and monitoring

3.3.3 Deep Learning in Flood Response

Flood response involves immediate actions to protect lives and property during and after

a flood event The main applications of AI in flood response include flood extent mapping and disaster rescue and relief

Understanding the distribution of floodwaters, assessing damage, and guiding response efforts require essential flood extent mapping Recent advancements in flood mapping technologies have leveraged CNNs and various data sources For example, Gebrehiwot et al (2019) harnessed deep CNNs combined with UAV data to generate high-resolution flood maps Hashemi-Beni & Gebrehiwot (2021) used the same approach, which combines DL with UAV imagery Sarker et al (2019) developed a flood mapping method utilizing CNNs, focusing on