Journal Pre-proof

Artificial Intelligence for Flood Risk Management: A Comprehensive State-of-the-Art

Review and Future Directions

Zhewei Liu, Natalie Coleman, Flavia Ioana Patrascu, Kai Yin, Xiangpeng Li, Ali
Mostafavi

PII: S2212-4209(24)00872-0
DOI: https://doi.org/10.1016/j.ijdrr.2024.105110
Reference: IJDRR 105110

To appear in: International Journal of Disaster Risk Reduction

Received Date: 22 September 2024

Revised Date: 12 December 2024
Accepted Date: 13 December 2024

Please cite this article as: Z. Liu, N. Coleman, F.I. Patrascu, K. Yin, X. Li, A. Mostafavi, Artificial
Intelligence for Flood Risk Management: A Comprehensive State-of-the-Art Review and Future
Directions, International Journal of Disaster Risk Reduction, https://doi.org/10.1016/j.ijdrr.2024.105110.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2024 Published by Elsevier Ltd.

 1 Artificial Intelligence for Flood Risk Management: A
2 Comprehensive State-of-the-Art Review and Future
3 Directions
4 Zhewei Liu, Natalie Coleman, Flavia Ioana Patrascu, Kai Yin*, Xiangpeng Li, Ali
5 Mostafavi
6

8 Urban Resilience.AI Lab, Zachry Department of Civil and Environmental

9 Engineering, Texas A&M University
10

f
oo
11 *kai_yin@tamu.edu
12

r
13 Abstract: -p
re
14 Climate hazards are escalating in frequency and severity, with flooding escalating as a
15 major threat. The limitations of the existing analytical necessitate and computational
lP

16 tools for flood risk management necessitates a shift towards more data-driven flood
17 risk management strategies informed by AI-driven tools and methods. This paper
18 explores the forefront of flood risk management focusing on integrating artificial
na

19 intelligence (AI), specifically machine learning (ML) and deep learning (DL)
20 technologies. By reviewing hundreds of relevant studies, we present a comprehensive
ur

21 analysis of AI applications in flood risk management by examining flood types, AI

22 models, spatial scales, input data, and practical applications, to provide a holistic view
Jo

23 of the current landscape and future potential of AI-enhanced flood risk management.
24 We highlight the extent to which AI-driven solutions can complement the existing
25 tools to enhance the reliability of flood predictions and inform mitigation and
26 response strategies. The paper also address prevailing challenges, including data bias
27 and the need for explainable AI models, and proposes pathways for future research to
28 fully harness AI's potential in mitigating flood risks. The analysis underscores AI's
29 promising potential in improving adaptive flood risk management, which is crucial for
30 safeguarding communities and infrastructures against the escalating challenges posed
31 by floods.
32 Keywords: Flood Risk Management, Artificial Intelligence (AI), Machine Learning,
33 Resilience, Equity.
34

1
35 1. Introduction
36 As our planet grapples with the escalating threats of climate hazards, the urgency for
37 innovative approaches to disaster risk management has never been more pronounced.
38 Among the myriad of natural calamities, flooding stands out as a particularly
39 insidious foe, wreaking havoc on communities, economies, and ecosystems with
40 increasing frequency and ferocity (Bush 2018; Rajput, et al. 2024). The confluence of
41 rising sea levels, intensified rainfall patterns, and the relentless expansion of urban
42 sprawl have amplified the vulnerability of our infrastructures and communities to
43 these aqueous onslaughts (Tahvildari and Castrucci 2021; Cao, et al. 2024a). This
44 paper seeks to explore the vanguard of flood risk management: the integration of
45 artificial intelligence (AI), with a particular focus on machine learning (ML) and deep
46 learning (DL) technologies.

47 Legacy flood risk management tools, such as physics-based models, are primarily

f
oo
48 used in predicting a diverse range of flooding scenarios. However, they suffer from
49 some limitations that hinder their effectiveness in the face of evolving threats. One of
50 the challenges lies in the need for extensive hydro-geomorphological monitoring

r
51
52
-p
datasets and intensive computation. For example, physically based models require
detailed and precise input data, such as topography, soil properties, and land use,
re
53 which can be difficult to obtain or update frequently (Fatichi, et al. 2016). The
54 computational complexity of these models also means that real-time or near-real-time
lP

55 flood prediction is often infeasible, particularly in regions where computational

56 resources are limited (Al-Sabhan, et al. 2003). This is especially problematic in
57 rapidly evolving flood situations where near-real-time predictions are crucial for
na

58 effective disaster management. Moreover, the development and application of

59 physically based models require a deep understanding of hydrological parameters,
ur

60 which demands specialized expertise that is not always readily available (Mulligan,
61 2013). This expertise is necessary to calibrate and validate the models accurately,
Jo

62 ensuring that they reflect the complex interactions between various environmental
63 factors. However, this complexity also introduces challenges; as noted by Costabile
64 and Macchione, and there is a significant gap in the short-term prediction capabilities
65 of these models (Costabile, et al. 2015). For instance, in some cases, these models
66 have failed to accurately predict flood events, leading to inadequate preparedness and
67 response efforts (Van, 2011).

68 Given these limitations, there is an urgent need for innovative solutions that can
69 address the challenges posed by traditional flood risk management methods. Artificial
70 intelligence (AI), particularly machine learning (ML) and deep learning (DL), offers
71 transformative potential to overcome these constraints and better anticipate the
72 evolving nature of flood risk. One of the significant advantages of AI, ML, and DL is
73 their ability to handle vast and complex datasets (Ahmed, et al. 2023). AI-based
74 models can process a wide array of data sources, including real-time hydrological
75 measurements, satellite imagery, weather forecasts, and historical flood records,
76 without the intensive computational burden typical of conventional physical models
77 (Chang, et al., 2023; Ye, et al., 2024). This ability to analyze large datasets rapidly

2
 78 enables more timely and accurate flood predictions, making it possible to generate
79 short-term forecasts that are crucial for effective flood response.

80 Moreover, AI and ML techniques can learn from the data itself (Roh , Heo, and
81 Whang, 2019, Shu, et al., 2024, Liu, et al., 2024, Chen et al., 2022, Chen et al., 2023),
82 reducing the need for extensive expertise in hydrological parameters that traditional
83 physically based models require. By identifying patterns and correlations within the
84 data, AI models can develop sophisticated predictive algorithms that do not rely
85 solely on predefined physical laws, allowing them to adapt to new and changing
86 conditions more effectively. Furthermore, DL models excel in extracting high-level
87 features from complex datasets (Chen, et al., 2016), enabling the creation of more
88 nuanced and precise flood risk assessments. These models can integrate various data
89 types and sources to generate comprehensive risk maps and forecasts that account for
90 multiple factors simultaneously, such as terrain, land use, and weather conditions

f
91 (Puttinaovarat, and Horkaew, 2020). As a result, AI-driven approaches can offer a

oo
92 more holistic view of flood risks, leading to more informed decision-making.
93 In the past decade, a growing number of studies have proposed AI-driven tools for

r
94
95
-p
various aspects of flood risk management and the development of AI-driven tools in
the field is growing. Due to the growing recognition of the potential of AI-driven
re
96 methods for flood risk management, this paper undertakes a comprehensive review of
97 the burgeoning field of AI in flood risk management. Through a meticulous
lP

98 examination of over three hundred studies, we endeavor to map the current landscape
99 of AI applications in this domain, identify the current progress and gaps, and spotlight
100 the areas needing further development in future research. Our inquiry is structured
na

101 around the following research questions:

102 ● To what extent AI, ML, and DL being employed to enhance the
ur

103 effectiveness of flood events and the efficacy of response strategies? This
104 includes exploring the deployment of sophisticated flood prediction models
Jo

105 that leverage satellite imagery, hydrological data, and other diverse data
106 sources to provide timely and accurate forecasts.
107 ● What are the prevailing challenges and limitations in integrating these
108 technologies into existing flood risk management frameworks? This
109 encompasses issues related to data quality, model reliability, computational
110 requirements, and the need for explainability in AI-driven decisions to ensure
111 stakeholders can trust and act upon AI insights with confidence.
112 ● In what ways do AI-based solutions enhance flood risk management
113 processes and what future directions should be pursued to ensure their
114 reliability, fairness, inclusivity, and effectiveness? This explores future
115 possibilities for developing open-source tools, ensuring fairness and
116 accountability, enhancing model reliability and explainability, and exploring
117 novel hybrid models and participatory approaches to fully harness AI's
118 potential in creating adaptive and equitable flood management strategies.
119 2. Framework of Literature Review

3
 f
oo
120
121 Figure 1. Five-dimensional framework for evaluating the studies related to AI for

r
122

123
-p
flood risk management
To effectively answer the multifaceted questions of flood risk management through
re
124 artificial intelligence (AI), this paper introduces a novel Five-dimensional framework,
125 as shown in Figure 1. This framework serves as a comprehensive analytical tool for
lP

126 evaluating the integration and effectiveness of AI technologies in flood risk

127 management. The dimensions include Flood Types, AI Models, Spatial Scales, Input
na

128 Data, and Applications, which were made to comprehensively capture the elements
129 that shape the development and use of AI-based tools in flood risk management.
130 Flood Types specify the diverse hydrological characteristics and scenarios that AI
ur

131 models must address, providing the foundation for understanding the unique
132 behaviors of different flood events. AI Models, in turn, determine the computational
Jo

133 approaches and techniques required to effectively analyze and predict these flood
134 types, ensuring that the methodologies align with the complexity and nature of the
135 flooding. Spatial Scales ensure that the analyses and solutions are appropriate for the
136 geographic context, whether local, regional, or global, tailoring the models to the
137 specific area of application. Input Data integrates the various attributes—such as
138 hydrological, topographical, and socio-demographic factors—needed to provide
139 robust, accurate predictions and characterizations of flood risks. Finally, Applications
140 bring these elements together by focusing on how AI models are deployed in real-
141 world flood risk management scenarios, from early warning systems to vulnerability
142 assessments, ensuring practical utility and societal relevance.
143 Flood Types: This dimension assesses the types of flooding events (e.g., coastal, flash
144 floods) studied within the scope of AI applications, reflecting the diverse nature and
145 sources of flood risks. Fluvial floods occur when water levels in rivers, lakes, or
146 streams surpass their banks due to excessive rain or snowmelt, leading to overflow
147 into surrounding areas (Merz, Hall et al. 2010). This type of flooding can affect
148 downstream regions, breaching levees and flooding communities. Their impact varies
149 with the landscape—floods rise slowly but linger in flat areas, whereas in hilly

4
150 regions, they can appear rapidly and carry destructive debris. Pluvial floods result
151 from heavy rainfall that overwhelms drainage systems, causing flooding independent
152 of water body overflows (Falconer, Cobby et al. 2009; Cao et al., 2024b). They can
153 occur anywhere, affecting both urban and rural settings without the need for nearby
154 water bodies. This category includes surface water floods, which gradually flood
155 streets and buildings, and flash floods, which emerge suddenly with intense water
156 flow and debris, posing significant dangers. Coastal floods are caused by seawater
157 inundation along coastlines, mainly from storm surges during high tide or tsunamis
158 (Woodruff, Irish et al. 2013). The severity of coastal flooding depends on the storm's
159 characteristics and the coastal topography (Xu and Liu, 2023). Storm surges, the
160 primary cause of coastal flooding, can lead to devastating losses in low-lying areas
161 (Xu and Liu, 2022). Predictive models for coastal floods take into account the storm's
162 features and historical storm data to estimate risks and potential impacts. As such,
163 differentiating between flood types can tailor AI models to predict and respond to

f
oo
164 specific hydrological behaviors and impacts. Knowing these minute characteristics
165 can enhance the accuracy of risk assessment and management strategies.

r
166 Spatial Scales: This aspect evaluates the spatial resolution at which flood phenomena
167
168
-p
are represented and studied, from localized site-specific analyses to regional or global
assessments. The scale for which AI is used for flood risk management depends on
re
169 the purpose of the application. For this review, the spatial scale is characterized as
170 local, regional, national-wide, and multiple national-wide. The national scale focuses
lP

171 on assessments of entire countries, typically using consistent national data and
172 encompassing numerous watersheds. The regional scale is generally sub-national,
173 referring to a specific province, watershed, or large area. The local scale is the
na

174 smallest, pertaining to a city, a specific stretch of a river, or a specific element. The
175 local scale is often selected for studies concentrating on specific cases, targeting
ur

176 particular small areas or cities (Ke, Tian et al. 2020, Iuppa, Cavallaro et al. 2021), or
177 focusing on individual features like structures (Pyayt, Mokhov et al. 2011, Melnikova,
Jo

178 Jordan et al. 2015, Lin, Lee et al. 2021). Therefore, AI applications need to be
179 scalable and adaptable to different spatial requirements and expectations in order to
180 ensure these interventions are optimized and relevant to the scale of impact.
181 Input Data: The variety and complexity of data inputs for AI models used in flood
182 risk analysis reflect the multifaceted nature of flood phenomena, which include not
183 just the occurrence of floods but also their causes, impacts, and the vulnerability of
184 the affected areas, as shown in Table 1. Types of input data typically for flood risk
185 analysis can be summarized as hydrological data, topographical/ geological data,
186 geological data, built environment data, remote sensing and satellite data, and
187 sociodemographic data.
188 Table 1. Summary of input data for AI models in flood analysis

Type of Input Data Description

River and stream flow

rates, water levels, soil
Hydrological Data
moisture content, and
groundwater levels.

5
 Elevation data, land use,
Topographical and
land cover, and watershed
geological Data
boundaries.

Locations and
Infrastructure and Built characteristics of
Environment Data buildings, roads, bridges,
and other infrastructure.

Records of past flood

events, including dates,
Historical Flood Events
locations, flood extents,
and damages.

High-resolution images

f
and data from satellites

oo
Remote Sensing and
and drones, including
Satellite Data
vegetation and water
bodies.

r
-p Population density,
demographics, economic
re
Socioeconomic Data
activities, and community
resources.
lP

189
190 AI Models: This dimension scrutinizes the specific AI methodologies employed,
na

191 ranging from conventional machine learning techniques to advanced deep learning
192 networks. The choice of model directly influences the accuracy and reliability of
193 flood event predictions and response strategies (Shi, et al., 2021). Classical models,
ur

194 such as linear regression, logistic regression, tree-based, support vector machines, and
195 clustering, form the foundation of machine learning. These models have the
Jo

196 advantage of being simpler to interpret and less computationally expensive, and have
197 been proven effective for modeling the complexity of flood cases (Liu, Felton, and
198 Mostafavi, 2024). Deep learning models, including long short-term memory
199 (LSTM), artificial neural networks (ANN), and convolution neural networks (CNN),
200 are more advanced neural network architectures to handle vast amounts of data (Liu,
201 et al., 2022) . The network consists of input, hidden, and output layers, where each
202 layer is made up of neurons that apply weights to inputs and pass them through a
203 nonlinear activation function (Yu, et al., 2023). This structure allows them to model
204 complex relationships, making them more effective for complex phenomena in which
205 multiple factors are interacting in nonlinear ways. Hybrid models combine the
206 strengths of classical and machine learning approaches. A hybrid model might use
207 classical machine learning techniques to extract and reduce features and then feed the
208 selected features into a deep learning model to predict. This allows for more efficient
209 handling of large-scale data while maintaining interpretability and efficiency in
210 specific parts of the model. Hence, selecting the right AI approach can leverage the
211 strengths of each model, potentially leading to improvements in the predictive
212 accuracy and operational efficiency of flood risk management.

6
213 Applications: Finally, we explore the practical applications of these AI technologies
214 in flood risk management across different phases. This dimension assesses how AI is
215 being utilized in real-world scenarios to enhance resilience and reduce the
216 vulnerability of populations to flooding events (Yin, et al., 2023; Farahmand, et al.,
217 2024; Liu, et al., 2024). Vulnerability assessments help in identifying populations
218 and areas that are at high risk of flooding impacts, which may consider the physical
219 and social attributes of the community. The purpose is to facilitate targeted resilience
220 and mitigation measures. Early warning systems are crucial for immediate alters to
221 inform about impending flood threats. It can assist in evacuation decisions and
222 preparation measures to minimize human and economic losses. Early forecasting
223 and nowcasting provide predictions about flood events right before they occur or in
224 real time. Nowcasting also allows for rapid decision-making during the events to help
225 with rescue and immediate recovery. These applications not only enhance emergency
226 preparedness and response but also support more long-term planning and adaptation

f
oo
227 strategies for flood-prone areas.
228 By structuring our review around these five dimensions, we aim to provide an in-

r
229 depth understanding of the current state and future potential of AI in enhancing the
230
231
-p
efficacy and responsiveness of flood risk management strategies. Understanding these
key concepts is pivotal for grasping the scope and impact of AI technologies in the
re
232 realm of flood risk management, making them central themes throughout our
233 discussion.
lP

234 3. Initial Assessment of Flood Literature

235 The research presents a review of machine learning and artificial intelligence-based
na

236 methods used in flood risk assessment applications. To capture the broad applications,
237 we searched on Google Scholar, which offers broader coverage of relevant literature
ur

238 than other sources like Scopus and Web of Science (Harzing and Alakangas, 2016),
239 for the following words “flood risk, flood vulnerability, flood nowcasting, flood
Jo

240 situational awareness, and flood early warning system” along with “artificial
241 intelligence, machine learning.” Only articles written in English and published from
242 2010 to the middle of 2023 were reviewed. In the initial review, 120 articles were
243 identified based on the relevance of the title and abstract. After the full-text review,
244 articles were excluded for the following reasons: (1) the paper had a format outside
245 the study scope (e.g. literature review, perspective paper, policy report) (2) the paper
246 included limited discussion on artificial intelligence and machine intelligence
247 applications, and (3) the paper focused on precipitation cast with limited to no
248 mention of flooding applications. Through this full-text screening process, 68 articles
249 remained.
250 Figure 2 shows a growing increase in the number of articles published on flood risk
251 management and artificial intelligence, particularly after 2020. This shows the
252 growing interest in applying AI techniques in addressing the limitations of legacy
253 tools in flood risk management.

7
 f
oo
254

r
255
256
-p
Figure 2. Number of journal publications related to AI in flood risk management
re
257
258
lP
na
ur
Jo

8
 f
r oo
-p
re
lP
na
ur
Jo

259
260 Figure 3. Distribution of Studies on AI applications in Flood Risk Management,
261 Categorized by Country and Spatial Scale. Countries like Iran (9 studies) and India (8
262 studies) predominantly focus on regional-scale studies. There is also a strong
263 emphasis on regional-level analyses across various countries, while fewer studies
264 address flood risks at a national-wide scale, suggesting that much of the AI-driven
265 flood risk management research is tailored to specific regional contexts, reflecting
266 localized approaches to flood mitigation.
267
268 4. Evaluation of Dimensions in Framework
269 This section presents concepts covered in studies with applications of AI-based
270 methods for flood risk management. These concepts are divided based on the
271 dimensions of our analysis framework shown in Figure 1.
272 The frequency of studies in each of the dimensions are summarized in the figure in
273 Supplemental Information. In terms of flood type, it includes fluvial (55.1%), pluvial
274 (31.5%), and coastal floods (13.5%). Coastal floods may be less represented in AI
275 applications due to the complexity of the data and models required, and less frequent

9
276 occurrence compared to fluvial and pluvial floods. Regarding spatial scales, the rank
277 order is regional (53.0%), local (45.5%), and national-wide (1.5%). The notable
278 disparity in the number of studies focusing on local and regional scales compared to
279 those on the nationwide scale can be attributed to the fact that flood assessments are
280 usually done at watershed level which involves a regional scale. While regional
281 models are necessary and useful, the extent to which models trained on data in one
282 region would be applicable to other regions is not examined and very limited efforts
283 have been made to evaluate the cross-regional transferability of models.
284 In addition, the types of AI models span across classical (51.2%), deep learning
285 (40.4), and hybrid (8.4%). Classical models may be favored in flood management
286 application due to their simplicity, lower data requirements, and greater transparency,
287 and lower computational costs. Deep learning models, though potentially more
288 accurate and powerful, require extensive data and significant computational resources.

f
289 As shown in Figure 4, the top methods for classical models were random forest

oo
290 (12.0%), support vector machine (10.9%), and logistic regression (10.9%) while the
291 top methods for deep learning were variations of neural networks such as artificial

r
292 neural network (56.2%), multilayer perceptron (16.7%), convolution neural network
293
294
-p
(10.4%), and long short term memory (8.3%). Hybrid models combine both
approaches. Lastly, the distribution of flood risk management applications is
re
295 vulnerability assessment (42.9%), early forecasting/ nowcasting (30.0%), and early
296 warning systems (27.1%). The prioritization of vulnerability assessment indicates a
lP

297 strong focus on long-term planning and risk reduction. The relatively limited research
298 on nowcasting and early warning systems indicates untapped opportunities to further
299 develop AI-based models for these applications. In fact, AI-based models would be
na

300 highly suitable for nowcasting and flood warning application since they can be pre-
301 trained and then use even-specific data to generate near-real-time predictions and
ur

302 characterization and address the limitation of legacy methods whose use in
303 nowcasting and flood warning are rather limited.
Jo

304

10
 f
r oo
-p
re
lP
na
ur
Jo

305
306 Figure 4. Top used methods in machine learning models
307

308

11
 f
r oo
-p
re
lP

309
na

310 Figure 5. Sankey diagram between flood scale, AI model, and data input
311
ur

312 Consideration of Spatial Scales: Figure 5 illustrates comparable variability between

Jo

313 the Classical and Deep Learning models with respect to spatial scales, with
314 approximately half of the studies conducted at the regional scale and the other half at
315 the local scale. Classical models, are robust and interpretable, making them suitable
316 for both detailed local studies and broader regional analyses. Similarly, Deep
317 Learning models, known for their ability to handle large and complex datasets, are
318 also well-suited for both scales, providing high accuracy and the ability to uncover
319 intricate patterns in data. A significant distinction appears with the Hybrid Learning
320 Model, where the majority of studies focus on regional studies.
321 For local-scale studies, the focus of Hybrid Learning models is often on detailed,
322 high-resolution data specific to a smaller area. In Zhou, Wu et al. (2023), the local
323 scale was used to demonstrate the method's effectiveness in a controlled environment
324 using physical information such as geometry, river systems, and topographic
325 attributes. This suggests that Hybrid Learning Models can effectively leverage
326 detailed local data for specific applications where precision and specificity are crucial.
327 Local-scale applications often benefit from the ability to model fine-grained variations
328 in terrain and infrastructure, which Hybrid Learning Models can handle by
329 incorporating detailed physical and spatial data. For regional-scale studies, a broader

12
330 range of data types has been integrated. For example, in Hadian, Afzalimehr et al.
331 (2022), the flood influential features included data such as altitude, slope, aspect, plan
332 curvature, distance from rivers, rainfall, soil type, lithology, and land use. This
333 implies that Hybrid Learning Models are versatile and can synthesize extensive and
334 diverse datasets to address complex, large-scale environmental challenges. The ability
335 to incorporate and process varied data sources makes Hybrid Learning Models
336 particularly effective at the regional scale, where the interplay of multiple factors
337 must be considered to accurately predict and characterize flood risks.
338 Importance of data inputs: Hydrological data are fundamental in flood risk analysis,
339 as they provide insights into the weather conditions that can lead to flooding. Wang
340 and Abdelrahman (2023) focus on a flood prediction system implemented in
341 Kristianstad, Sweden, utilizing a series of local sensors to monitor environmental
342 conditions such as rainfall, along with third-party weather forecasts. Hydrological

f
343 data form a critical component of the datasets used in AI models for flood risk

oo
344 analysis, providing essential insights into the dynamics of water movements and
345 storage within the earth's hydrological cycles (Cloke, Di Baldassarre et al. 2017).

r
346 These data include river flow rates (Kim, Park et al. 2023), water levels in bodies of
347
348
-p
water (Lin, Sreng et al. 2023), and soil moisture content (ElSaadani, Habib et al.
2021). Antzoulatos, Kouloglou et al. (2022) present a methodological framework for
re
349 dynamically assessing flood hazard and risk levels by combining data such as water
350 velocity and water depth. Xu, Wang et al. (2022) focus on deep learning models like
lP

351 CNNs, LSTMs, and GRUs for long-term hydrological simulation, highlighting their
352 application in capturing temporal and spatial features of hydrological data.
na

353 Topographical and geological data, especially Digital Elevation Models (DEMs) and
354 Land Use and Land Cover (LULC) data, offer a foundation for analyzing how water
355 interacts with the earth's surface (Rahman, Ningsheng et al. 2021; Avand, Kuriqi et al.
ur

356 2022; Li et al., 2022), are used to understand flood plains and the potential spread of
357 floodwater. LULC data (Costache, Bao Pham et al. 2020) analyzes how different
Jo

358 terrains and vegetation impact flood dynamics. This information is usually derived
359 from satellite imagery, aerial photography, and ground surveys. Data related to
360 rainfall, elevation, land use, and soil type were used in Rahayu, Mathias et al. (2023)
361 to examine their impact on flood risk using classical machine learning methods
362 including classification, and linear regression. It is indispensable for understanding
363 how geological features influence water flow, flood propagation (Yin and Mostafavi
364 2023), and the impact on inhabited areas (Yin and Mostafavi 2023). For satellite and
365 aerial imagery, high-resolution images from satellites or drones could provide detailed
366 information on land surface conditions before, during, and after flood events, aiding
367 in the detection of flooded areas (Stateczny, Praveena et al. 2023) and assessment of
368 damages (Kaur, Lee et al. 2023). Remote sensing data includes data on precipitation,
369 soil moisture, and water bodies obtained from remote sensing technologies (Shi et al.,
370 2024). These factors could be used for predicting flood risks and understanding flood
371 dynamics (Bentivoglio, Isufi et al. 2022). These images are pivotal in observing and
372 analyzing the physical characteristics of the land surface, including areas covered by
373 water, vegetation, urban developments, and other land cover types (Cheng and Li,
374 2023). These are frequently used in land cover and vegetation mapping (Wang and

13
375 Abdelrahman 2023), water bodies monitoring, and flood extent and damage
376 assessment (Kaur, Lee et al. 2023).
377 Sociodemographic data plays a crucial role in understanding and assessing the
378 potential impacts of flooding events on human populations and their built
379 environment. Previous works have found a strong relationship between
380 socioeconomic status and the risk of flood. Population (including population number,
381 race, income, etc.) related maps are detailed representations of the distribution of
382 individuals across a geographic area, they are essential for flood risk analysis as they
383 help in identifying high-risk areas, assessing potential human impact, and
384 vulnerability analysis (Zhou et al., 2024b; Sun et al., 2024). Infrastructure and asset
385 maps provide detailed information about the built environment, including the location,
386 type, and value of buildings, roads, bridges, utilities, and other critical infrastructure
387 (Zhou et al., 2021; Zhou et al., 2024a). Their significance in flood risk analysis

f
388 includes: economic loss estimation (Chen, You et al. 2021), critical infrastructure

oo
389 protection (Mester, Frieler et al. 2023), and reconstruction planning (Khanam, Sofia et
390 al. 2023). Road networks disruptions (Yuan, Xu et al. 2022), drainage systems

r
391 (Kazadi, Doss-Gollin et al. 2022), and other infrastructural elements (Bentivoglio,
392
393
-p
Isufi et al. 2023) are frequently modeled by GNNs to model how floods spread
through urban areas and impact communities.These models help in identifying
re
394 vulnerable elements in the infrastructure and improving the design of resilient urban
395 systems.
lP

396 Moreover, the selection of input data for flood management studies varies
397 significantly across spatial scales, reflecting differences in study objectives and data
na

398 availability. Local-scale studies often utilize highly detailed and site-specific data,
399 such as hydrological features and physical sensor data from rain gauges and vibration-
400 based sensors (Farahmand, Xu et al., 2023; Suwarno et al. 2021). These datasets,
ur

401 characterized by their fine resolution, are well-suited to capturing localized conditions
402 and dynamics in small catchment areas. At the regional scale, the focus shifts toward
Jo

403 broader topographical and land-use/land-cover data, including slope, elevation,

404 curvature, and generalized land-use classifications, often supplemented with
405 meteorological data like rainfall at a coarser resolution (Nakhaei et al., 2023;
406 Tabbussum and Dar, 2021). National-scale studies predominantly rely on generalized
407 datasets, such as large-scale meteorological trends, land cover, and remote sensing
408 data, emphasizing broader patterns and trends that inform policy-making (Müller et
409 al., 2022). Across scales, granularity decreases as the focus transitions from highly
410 specific physical and hydrological features at the local level to aggregated
411 meteorological and land-use data at the national level. This trend highlights the
412 importance of tailoring data selection to the spatial scale to align with study goals and
413 ensure analytical precision.
414 Application of AI models: A variety of model types have been used in the literature.
415 For example, Ha, Bui et al. (2023) used Adaboost and logistic regression to evaluate
416 flooded and non-flooded areas in Quang Binh province Vietnam. The ensemble of
417 adaboost-logistic models (91.6%) scored higher accuracy than the individual models
418 of adaboost (71.0%) and logistic (82.2%). Tree-based models of decision-tree and
419 random forest can model paths to outcomes which would be useful in outlining the

14
420 specific flood conditions. Support vector machines can classify and predict nonlinear
421 boundaries by transforming inputs into higher dimensions. One study by Rafiei-
422 Sardooi, Azareh et al. (2021) used support vector machines, boosted regression tree,
423 and random forest to create flood hazard maps of Jiroft city, Iran. Random forest had
424 the best area under the curve performance for testing data at 95.8%, followed by
425 support vector machine at 89.9%, and boosted-regression tree at 86.5%. However,
426 there is variability in the performance of classical models given specific features,
427 location, and flood application meaning that a specific model may highly perform in
428 one study and have a lower level of performance in another. A study by Ke, Tian et
429 al. (2020) used various machine learning models to classify flooded and non-flooded
430 areas in Shenzhen, China. Among the classical models of decision trees, discriminant
431 analysis, support vector machine, k-nearest, and subspace models, the discriminant
432 subspace models had the highest accuracy of 99%- although these classical models
433 are not commonly found in the literature. Additionally, classical models can be

f
oo
434 enhanced with optimization algorithms. In a study of Bui, Pradhan et al. (2016), the
435 metaheuristic optimization algorithms and neural fuzzy Inference model with genetic
436 and particle swam optimization (91.1% validation area under the curve) outperformed

r
437 other classical models of decision tree, random forest, support vector machine, fuzzy,
438
439
-p
and even a neural network model. Our analysis also show that classical models are
particularly favored for fluvial floods due to their simplicity, interpretability, and
re
440 effectiveness in modeling well-understood hydrological processes. These floods,
441 caused by river overflow from excessive rainfall or snowmelt, are typically analyzed
lP

442 using structured datasets like river flow rates and historical flood records. Classical
443 models, such as Decision Trees and Logistic Regression, excel in providing clear,
na

444 interpretable results that support decision-making in flood risk management. Their
445 low computational requirements make them practical for regions with limited
446 resources or real-time prediction needs. Additionally, enhancements like
ur

447 metaheuristic optimization further boost their accuracy, making them a reliable choice
448 for fluvial flood modeling.
Jo

449 Deep learning models, including long short-term memory (LSTM), artificial neural
450 networks (ANN), and convolution neural networks (CNN), are more advanced neural
451 network architectures to handle vast amounts of data. The network consists of input,
452 hidden, and output layers, where each layer is made up of neurons that apply weights
453 to inputs and pass them through a nonlinear activation function. This structure allows
454 them to model complex relationships, making them more effective for complex
455 phenomena in which multiple factors are interacting in nonlinear ways. Farahmand,
456 Xu et al. (2023) uses physics-based and human-sensed data to predict urban flood
457 events in near real-time. It uses an attention-based spatial-temporal graph convolution
458 network, specifically designed to account for spatial and temporal dependencies in
459 flood propagation in 2017 Hurricane Harvey for Harris County, TX, and the study
460 compares other neural network models. Yin and Mostafavi (2023b) proposed a graph
461 deep learning framework to estimate the community flood risk level in an
462 unsupervised manner. Therefore, these models excel in understanding dependencies
463 in data which is beneficial for unraveling the intercorrelations between features. In
464 another study of the convolutional LSTM model (Moishin, Deo et al. 2021), it
465 outperformed a classical machine learning model of support vector regression as well

15
466 as benchmark models of LSTM and CNN-LSTM for multiple-day periods. For
467 instance, Muñoz, Orellana-Alvear et al. (2021) compare five machine learning models
468 for the tombembda catchment in the southern Andes in Ecuador. The neural-network-
469 based model achieved the highest performance for “pre-alert” and “alert” states which
470 are crucial for flood early warning systems. These models demonstrate the powerful
471 predictive power in flooding events; however, this increased performance may be a
472 trade-off with computational power and speed processing. Deep Learning models are
473 more widely adopted for pluvial floods compared to fluvial floods, which may be due
474 to their ability to handle the unpredictable and highly dynamic nature of these events.
475 Pluvial floods, caused by sudden and intense rainfall leading to flash flooding or
476 surface water accumulation, require advanced computational approaches capable of
477 capturing real-time changes and spatial-temporal patterns. Models such as LSTMs,
478 CNNs, and Graph Neural Networks excel in processing large-scale, high-dimensional
479 datasets like weather radar imagery, sensor networks, and topographical data. Unlike

f
oo
480 fluvial floods, which often involve structured datasets, pluvial floods demand
481 sophisticated models to account for the complex and non-linear interactions between
482 precipitation, terrain, and drainage systems. This greater adoption is driven by the

r
483 ability of Deep Learning models to provide accurate short-term predictions and
484
485
-p
identify critical hotspots, which are essential for real-time flood warnings and
mitigation strategies. However, the reliance on such models may require substantial
re
486 computational resources, posing challenges for their deployment in regions with
487 limited infrastructure.
lP

488 Hybrid models combine the strengths of classical and machine learning approaches. A
489 hybrid model might use classical machine learning techniques to extract and reduce
na

490 features and then feed the selected features into a deep learning model to predict. This
491 allows for more efficient handling of large-scale data while maintaining
ur

492 interpretability and efficiency in specific parts of the model. For example, Zhou, Wu
493 et al. (2023) integrated support vector regression, back propagation neural network,
Jo

494 and gradient-boosted decision tree to predict urban flooding in Zhengzhou City. The
495 combined model was 11.4-50.4% higher in accuracy compared to individual models.
496 In another study, Mahato, Pal et al. (2021) used an ensemble-based approach of
497 random forest, artificial neural networks, and radial basis function neural networks to
498 measure river flooding of the Atreyee River Basin in India and Bangladesh. The
499 ensemble method scored the highest at 93.1% AUC score. The review of the AI types
500 shows that the use case and data availability are two important aspects in the selection
501 of the AI approach to be used. All three groups of methods (i.e., classical, deep
502 learning, and hybrid learning) have been shown to be effective and reliable across
503 different use cases and applications.
504 5. Flood Applications
505 In terms of use case applications, Figure 6 shows the connections between types of
506 machine learning models and applications. Regarding flood vulnerability
507 assessment, there are more classical models than deep learning while in flood risk
508 prediction, there are more deep learning models. This may be because classical
509 models are sufficient in categorizing flood risk at low, medium, and high level which
510 is adequate for vulnerability assessments. For example, a study using a support vector

16
511 machine, k-nearest neighbors, and linear discriminant analysis to measure flood
512 susceptibility (low, medium, and high flood risk categories) (Khan, Shahid et al.
513 2019; Liu, and Mostafavi 2024). Based on the dataset of precipitation, temperature,
514 and windspeed, the models produced accurate results. Costache and Bui (2019)
515 showed that classification and regression trees (CART) outperformed multi-layer
516 precipitation (MLP) in identifying areas of flood vulnerability. In other cases, deep
517 learning models do outperform classical machine learning. As shown in Pham, Ali et
518 al. (2022), MLP-ANN and DLNN-DL4j had higher AUC scores (0.975 and 0.943,
519 respectively) than Bayesian logistic regression models (0.697). These models were
520 not just based on weather-related data, but also more complex granular scale data in
521 topography, land use, and building data, where deep learning could be the better
522 predictor. In addition, deep learning models, given enough quality data and
523 computational power, can capture complex and non-linear interactions among various
524 features that shape flood risk to provide accurate predictions and characterizations.

f
oo
525 For early warning forecasting, Tabussum and Dar (2021) used artificial neural
526 networks, fuzzy logic, and adaptive neuro-fuzzy inference systems to model and

r
527 predict flood events of the Jhelum River. Several of the early warning forecasting
528
529
-p
used wireless sensors (Sahagun et al 2017). For instance, Wahyono et al (2019)
utilized distributed wireless sensor networks data in training a model for flood early
re
530 warning. These sensors collect real-time data which is processed t to provide rapid
531 and accurate flood predictions. The TV-BMA model, in particular, showed an
lP

532 improved accuracy of 11.4-50.4% to benchmark models. Several of the papers that
533 used early warning forecasting also proposed, used, and tested early warning
534 systems. Hayder et al (2023) focused on flood early forecasting using a combination
na

535 of machine learning and deep learning algorithms to integrate in advanced Internet-of-
536 Things sensors. It is also important to account for temporal dynamics and event-
ur

537 specific characteristics for near-real-time predictions. As Zhou, Wu et al. (2023)

538 showed, models decreased in performance as the time of forecast increased from 10
Jo

539 minutes to 120 minutes. This suggests that models for near-real-time forecasting that
540 use sensor data perform better for near-term forecasting and their performance
541 diminishes with increased prediction time horizon.
542 Few research studies highlighted early warning coastal flooding. Among them,
543 Chondros et al (2021) trained an artificial neural network model to predict coastal
544 flood risk using various sea-state scenarios. Although most of the articles used a
545 binary alert system, Munoz et al (2021) used multiple machine learning models to
546 alert different flood stages (no-alert, pre-alert, alert). The study found that logistic
547 regression and multi-layer perception had the overall best performance. In addition to
548 sensor data, early warning systems can combine historical weather data, sensors, and
549 crowdsourced data for improved results, as shown by Illukkumbure et al (2021).

17
 f
r oo
-p
re
550
lP

551 Figure 6. Application of flood techniques on types of models and data

na

552 5. Enhancements and Future Directions in AI-Enhanced Flood Risk

553 Management
ur

554 Looking ahead, the application of Artificial Intelligence (AI) in flood risk
555 management is an emerging field that promises to revolutionize the ways in which we
Jo

556 predict, prepare for, and respond to flood events. This exploration aims to highlight
557 the significant enhancements AI brings to flood risk management processes, from
558 improving the accuracy of flood predictions and the efficiency of emergency
559 responses to advancing long-term mitigation strategies. By integrating state-of-the-art
560 machine learning algorithms, deep learning models, and predictive analytics, AI is
561 poised to enhance situational awareness, responsiveness, and adaptability in managing
562 flood risks. To unleash the full potential of AI for enhancing flood risk management
563 capabilities, there are multiple areas that future research needs to address.

18
564

f
oo
565 Figure 7. Vision for the future perspective in AI for flood risk management
566 Future Directions for AI in Flood Risk Management

r
567

568
Open Source and Capacity Building -p
The analysis of previous sections underscored the effectiveness of AI tools in
re
569 addressing diverse flood scenarios across different spatial scales and data contexts,
570 highlighting the need for solutions that are both scalable and adaptable to various
lP

571 geographic and socio-economic conditions (Ye, et al. 2021). Recognizing the
572 importance of inclusivity in deploying AI-driven solutions, ensuring broader
na

573 accessibility to these technologies is essential for democratizing their benefits and
574 enhancing their utility in diverse contexts. Despite the significant advancements in AI
575 technology, the accessibility and customization of these tools are still constrained,
ur

576 particularly in vulnerable areas that are often most impacted by flooding (Bhuyan, et
577 al. 2023; Chang, et al., 2023). This is crucial in the context of flood risk management,
Jo

578 where the ability to rapidly disseminate and replicate models across different regions
579 can significantly enhance cross-regional transferability of models and their use. Open-
580 source models can help develop more robust and adaptable AI tools that are better
581 suited to address the unique and dynamic challenges posed by flood risks in different
582 geographical and socio-economic contexts (Horita, et al., 2015 ).

583 Focusing on capacity building ensures that AI technologies are not only more
584 accessible but also more relevant and effective for diverse user groups, particularly
585 those in underserved or high-risk areas where flood risks are most acute.
586 Democratizing AI in this way is crucial for ensuring that innovative flood
587 management solutions reach and benefit the most vulnerable populations, who often
588 bear the brunt of flooding disasters. Furthermore, by empowering local communities
589 with the skills to use and adapt AI tools, we can foster more resilient and sustainable
590 flood management practices. This local ownership of tools and strategies enhances the
591 effectiveness of flood risk management efforts and ensures that the solutions
592 developed are directly responsive to the specific needs and challenges faced by these
593 communities.

19
594 Fairness-Based and Responsibility-Based Approaches

595 As found in previous sections, AI models are applied in diverse contexts, spanning
596 different geographic, socio-economic, and demographic conditions. This diversity
597 underscores the importance of ensuring that AI-driven solutions are equitable and
598 responsive to the unique vulnerabilities of affected populations. The integration of AI
599 into flood risk management must be grounded in the principles of fairness and
600 responsibility to fully realize its potential to improve outcomes (Gevaert, et al., 2021;
601 Yang, et al., 2020). There is a recognized risk that AI systems, if not carefully
602 designed and implemented, could perpetuate or even exacerbate existing inequalities,
603 particularly when deployed in areas with varying socio-economic conditions
604 (Kozcuer, Mollen, and Bießmann 2024). For instance, AI models trained on biased
605 data or those that fail to account for the diverse needs of different populations might
606 unintentionally reinforce disparities, leading to unequal levels of protection and

f
607 potentially widening the gap between different communities during flood events.

oo
608 As AI becomes more prevalent in flood risk management, it is imperative to develop

r
609 and implement robust responsible AI guidelines to govern its use. This involves
610
611
-p
designing algorithms that are acutely aware of the socio-economic and demographic
contexts in which they are deployed. For example, algorithms must be able to adjust
re
612 their predictions based on the varying needs and vulnerabilities of different
613 communities, ensuring that flood warnings and resources are distributed equitably. By
lP

614 embedding these accountability mechanisms into the deployment of AI technologies,

615 we can build trust among stakeholders and ensure that AI’s role in flood risk
616 management is both fair and equitable.
na

617 Explainability of ML/DL Models

ur

618 Another gap identified in the preceding analysis is the challenge of model
619 interpretability in the deployment of advanced AI systems for flood risk management.
Jo

620 While machine learning (ML) and deep learning (DL) models demonstrate significant
621 potential to enhance the accuracy and efficiency of flood predictions, their complexity
622 often limits the ability of stakeholders to understand and trust the outputs. This
623 highlights the critical need for explainable systems that can bridge the gap between
624 technical performance and practical decision-making (Pradhan, et al., 2023; Albahri,
625 et al., 2024; Başağaoğlu, et al., 2022). The literature identifies existing models' lack of
626 interpretability, emphasizing the need for more explainable systems to better inform
627 decisions and actions.
628 The explainability of AI models is vital for gaining acceptance among the various
629 stakeholders involved in flood risk management. The literature underscores the
630 importance of making AI-driven decisions transparent and understandable, especially
631 for non-technical users such as emergency responders, community leaders, and
632 policymakers (Bird, Dudík et al. 2020). This transparency not only enhances trust in
633 the technology but also facilitates more informed and confident decision-making
634 during flood events. For instance, if an AI model specifies an area as a high-risk area,
635 it should provide explanation of features importance that led to the specific
636 classification (Liu, and Mostafavi 2024).

20
637 When stakeholders understand how a model works and why it generates specific
638 predictions for flood events, they are better equipped to interpret the results accurately
639 and take appropriate, timely action. This is particularly important in flood risk
640 management, where decisions often need to be made quickly to mitigate the impacts
641 of imminent flooding. Future research should prioritize the development of ML/DL
642 models that not only deliver reliable performance but also offer a high degree of
643 interpretability. This will enable stakeholders to confidently rely on AI-driven
644 insights, ensuring that flood risk management practices are both effective and
645 equitable. By focusing on both reliability and explainability, we can advance the use
646 of AI in flood risk management, making it a more trusted and integral tool in
647 safeguarding communities against the devastating impacts of floods.
648 Addressing Data Bias
649 Finally, the analysis of AI applications across diverse flood types, spatial scales, and

f
oo
650 data sources also leads to the critical need to address data bias in AI-driven flood risk
651 management models. While these models effectively utilize large datasets for accurate
652 predictions, unbalanced or unrepresentative data can result in inequitable outcomes.

r
653 Ensuring fairness and inclusivity in data-driven predictions is essential to safeguard
654
-p
vulnerable communities and enhance the overall reliability of flood risk strategies. In
655 the context of flood risk management, this can result in models that systematically
re
656 underpredict risks for certain communities, leaving them more exposed to the dangers
657 of flooding (Schubert, et al. 2024).
lP

658 The implications of data bias in flood risk management are significant. For instance, if
659 an AI model is trained primarily on data from urban areas with well-developed
na

660 infrastructure, it may fail to accurately predict flood risks in rural or low-income areas
661 that lack similar infrastructure. This could lead to a lack of timely warnings and
ur

662 inadequate preparedness measures in these underrepresented regions, resulting in

663 disproportionate damage and loss during flood events. Such biases not only fail to
Jo

664 protect all communities equally but can also reinforce and deepen existing socio-
665 economic disparities.
666 To effectively address data bias in flood risk management, it is essential to adopt
667 rigorous data validation processes and to incorporate a diverse range of data sources
668 that reflect the full spectrum of conditions and scenarios present across different
669 regions. The literature suggests that integrating data from various sources—such as
670 satellite imagery, historical flood records, topographical data, and real-time sensor
671 inputs—can help create more balanced and representative models (Wan, Li et al.
672 2020). By ensuring that these models are trained on a comprehensive dataset that
673 includes diverse geographic and socio-economic contexts, we can reduce the risk of
674 bias and improve the accuracy of flood predictions for all communitie
675 6. Concluding Remarks
676 This manuscript provides a comprehensive review of the current landscape and future
677 potential of AI applications in flood risk management. By examining over three
678 hundred studies, we categorize and analyze AI-based solutions across five key
679 dimensions: flood types, AI models, spatial scales, input data, and practical

21
680 applications. The review highlights significant advancements in flood risk
681 management, such as AI-driven flood prediction, susceptibility mapping, and damage
682 assessment, emphasizing the transformative role of AI in enhancing the resilience,
683 responsiveness, and adaptability of flood risk management strategies.
684 The value of our work lies in its holistic approach to evaluating and proposing AI's
685 integration into flood risk management. Our five-dimensional framework not only
686 facilitates a deeper understanding of the technological landscape but also identifies
687 critical areas for future research. By addressing key challenges such as data bias,
688 model reliability, and explainability, future works can explore the way for more
689 effective, equitable, and inclusive AI-based flood management solutions.
690 Furthermore, the emphasis on open-source tools, capacity building, and ethical
691 guidelines will also ensure that AI technologies are accessible and beneficial to all
692 communities, particularly those historically marginalized and disproportionately

f
693 affected by flooding.

oo
694 By leveraging state-of-the-art machine learning (ML) and deep learning (DL) models,
695 AI can significantly enhance the accuracy of flood predictions, optimize emergency

r
696 response strategies, and advance long-term mitigation efforts. The successful
697
-p
deployment of AI in this domain necessitates addressing several key challenges,
698 including ensuring data quality, model reliability, and explainability. Future research
re
699 can develop open-source AI tools and platforms to encourage widespread adoption
700 and customization. Responsible AI guidelines and standards can also be established to
lP

701 ensure fairness and accountability, preventing the perpetuation of existing

702 inequalities. In addition, exploring novel AI models and methodologies, such as
na

703 hybrid models and participatory approaches, will further unlock AI's potential in
704 enhancing flood resilience. By pursuing these innovative research directions, AI
705 technologies can create more adaptive, equitable, and effective flood risk management
ur

706 strategies, ultimately safeguarding communities against the escalating threats posed
707 by climate change-induced flooding.
Jo

708

709

22
710 References:
711 Ahmed, S. F., Alam, M. S. B., Hassan, M., Rozbu, M. R., Ishtiak, T., Rafa, N., ... &
712 Gandomi, A. H. (2023). Deep learning modelling techniques: current progress,
713 applications, advantages, and challenges. Artificial Intelligence Review, 56(11),
714 13521-13617.
715
716 Albano, R., Mancusi, L., Sole, A., & Adamowski, J. (2017). FloodRisk: a
717 collaborative, free and open-source software for flood risk analysis. Geomatics,
718 Natural Hazards and Risk, 8(2), 1812-1832.
719
720 Albahri, A. S., Khaleel, Y. L., Habeeb, M. A., Ismael, R. D., Hameed, Q. A., Deveci,
721 M., ... & Alzubaidi, L. (2024). A systematic review of trustworthy artificial
722 intelligence applications in natural disasters. Computers and Electrical Engineering,
723 118, 109409.

f
724

oo
725 Al-Sabhan, W., Mulligan, M., & Blackburn, G. A. (2003). A real-time hydrological
726 model for flood prediction using GIS and the WWW. Computers, Environment and
727 Urban Systems, 27(1), 9-32.

r
728
729
-p
Antzoulatos, G., et al. (2022). "Flood hazard and risk mapping by applying an
730 explainable machine learning framework using satellite imagery and GIS data."
re
731 Sustainability 14(6): 3251.
lP

732
733 Avand, M., et al. (2022). "DEM resolution effects on machine learning performance
734 for flood probability mapping." Journal of Hydro-Environment Research 40: 1-16.
na

735
736 Başağaoğlu, H., Chakraborty, D., Lago, C. D., Gutierrez, L., Şahinli, M. A.,
ur

737 Giacomoni, M., Şengör, S. S. (2022). A review on interpretable and explainable

738 artificial intelligence in hydroclimatic applications. Water, 14(8), 1230.
Jo

739
740 Bentivoglio, R., et al. (2022). "Deep learning methods for flood mapping: a review of
741 existing applications and future research directions." Hydrology and Earth System
742 Sciences 26(16): 4345-4378.
743
744 Bentivoglio, R., et al. (2023). "Rapid spatio-temporal flood modelling via hydraulics-
745 based graph neural networks." Hydrology and Earth System Sciences 27(23): 4227-
746 4246.
747
748 Bhuyan, K., Van Westen, C., Wang, J., & Meena, S. R. (2023). Mapping and
749 characterising buildings for flood exposure analysis using open-source data and
750 artificial intelligence. Natural Hazards, 119(2), 805-835.
751
752 Bird, S., et al. (2020). "Fairlearn: A toolkit for assessing and improving fairness in
753 AI." Microsoft, Tech. Rep. MSR-TR-2020-32.
754
755 Bush, M. J. (2018). Climate change adaptation in small island developing states, John
756 Wiley & Sons.
23
757
758 Cao, D., Tam, C. Y., & Xu, K. (2024a). Impacts of El Niño diversity on East Asian
759 summertime precipitation extremes. Climate Dynamics, 1-17.
760
761 Cao, D., Tam, C. Y. F., & Xu, K. (2024b). Simulating springtime extreme rainfall
762 over Southern East Asia: unveiling the importance of synoptic-scale
763 activities. Climate Dynamics, 62(9), 9073-9096.
764
765 Chang, F. J., Chang, L. C., & Chen, J. F. (2023). Artificial intelligence techniques in
766 hydrology and water resources management. Water, 15(10), 1846.
767
768 Chen, A., et al. (2021). "The economic loss prediction of flooding based on machine
769 learning and the input-output model." Atmosphere 12(11): 1448.
770 Chen, X., & Di, X. (2022, October). How the covid-19 pandemic influences human

f
771 mobility? similarity analysis leveraging social media data. In 2022 IEEE 25th

oo
772 International Conference on Intelligent Transportation Systems (ITSC) (pp. 2955-
773 2960). IEEE.

r
774 Chen, X., Wang, Y., & Di, X. (2023, September). Whose Attitudes Toward Transit
775
776
-p
Are Most Affected by Rising Subway Crimes in New York City?. In 2023 IEEE 26th
International Conference on Intelligent Transportation Systems (ITSC) (pp. 4153-
re
777 4158). IEEE.
778
lP

779 Chen, Y., Jiang, H., Li, C., Jia, X., & Ghamisi, P. (2016). Deep feature extraction and
780 classification of hyperspectral images based on convolutional neural networks. IEEE
na

781 transactions on geoscience and remote sensing, 54(10), 6232-6251.

782
783 Cheng, X., & Li, Z. (2023). Modeling information flow from multispectral remote
ur

784 sensing images to land use and land cover maps for understanding classification
785 mechanism. Geo-spatial Information Science, 1-17.
Jo

786
787 Cloke, H., et al. (2017). "Hydrological risk: floods." Science for disaster risk
788 management 2017: knowing better and losing less. Poljanšek, K., Marín Ferrer, M.,
789 De Groeve, T., Clark, I.,(Eds.): 198-238.
790
791 Costabile, P., Macchione, F., Natale, L., & Petaccia, G. (2015). Flood mapping using
792 LIDAR DEM. Limitations of the 1-D modeling highlighted by the 2-D approach.
793 Natural Hazards, 77, 181-204.
794
795 Costache, R., et al. (2020). "Using GIS, remote sensing, and machine learning to
796 highlight the correlation between the land-use/land-cover changes and flash-flood
797 potential." Remote sensing 12(9): 1422.
798
799 Costache, R. and D. T. Bui (2019). "Spatial prediction of flood potential using new
800 ensembles of bivariate statistics and artificial intelligence: A case study at the Putna
801 river catchment of Romania." Science of the Total Environment 691: 1098-1118.
802

24
803 ElSaadani, M., et al. (2021). "Assessment of a spatiotemporal deep learning approach
804 for soil moisture prediction and filling the gaps in between soil moisture
805 observations." Frontiers in artificial intelligence 4: 636234.
806
807 Falconer, R. H., et al. (2009). "Pluvial flooding: new approaches in flood warning,
808 mapping and risk management." Journal of Flood Risk Management 2(3): 198-208.
809
810 Farahmand, H., et al. (2023). "A spatial–temporal graph deep learning model for
811 urban flood nowcasting leveraging heterogeneous community features." Scientific
812 Reports 13(1): 6768.
813
814 Farahmand, H., Yin, K., Hsu, C. W., Savadogo, I., Alegre, X. E., & Mostafavi, A.
815 (2024). Integrating climate projections and probabilistic network analysis into

f
oo
816 regional transport resilience planning. Transportation Research Part D: Transport and
817 Environment, 104229.

r
818
819
820
-p
Fatichi, S., Vivoni, E. R., Ogden, F. L., Ivanov, V. Y., Mirus, B., Gochis, D., ... &
Tarboton, D. (2016). An overview of current applications, challenges, and future
re
821 trends in distributed process-based models in hydrology. Journal of Hydrology, 537,
822 45-60.
lP

823
824 Gevaert, C. M., Carman, M., Rosman, B., Georgiadou, Y., & Soden, R. (2021).
825 Fairness and accountability of AI in disaster risk management: Opportunities and
na

826 challenges. Patterns, 2(11).

827
828 Ha, H., et al. (2023). "A practical approach to flood hazard, vulnerability, and risk
ur

829 assessing and mapping for Quang Binh province, Vietnam." Environment,
830 Development and Sustainability 25(2): 1101-1130.
Jo

831
832 Hadian, S., et al. (2022). "Determining flood zonation maps, using new ensembles of
833 multi-criteria decision-making, bivariate statistics, and artificial neural network."
834 Water 14(11): 1721.
835
836 Harzing, A. W., & Alakangas, S. (2016). Google Scholar, Scopus and the Web of
837 Science: a longitudinal and cross-disciplinary comparison. Scientometrics, 106, 787-
838 804.
839
840 Horita, F. E., de Albuquerque, J. P., Degrossi, L. C., Mendiondo, E. M., & Ueyama, J.
841 (2015). Development of a spatial decision support system for flood risk management
842 in Brazil that combines volunteered geographic information with wireless sensor
843 networks. Computers & Geosciences, 80, 84-94.
844
845 Iuppa, C., et al. (2021). Coastal flooding risk assessment through artificial
846 intelligence. Recent Advances in Environmental Science from the Euro-
847 Mediterranean and Surrounding Regions (2nd Edition) Proceedings of 2nd Euro-

25
848 Mediterranean Conference for Environmental Integration (EMCEI-2), Tunisia 2019,
849 Springer.
850
851 Kaur, N., et al. (2023). "Large‐scale building damage assessment using a novel
852 hierarchical transformer architecture on satellite images." Computer‐Aided Civil and
853 Infrastructure Engineering 38(15): 2072-2091.
854
855 Kazadi, A. N., et al. (2022). "Flood prediction with graph neural networks." Climate
856 Change AI. Climate Change AI.
857
858 Ke, Q., et al. (2020). "Urban pluvial flooding prediction by machine learning
859 approaches–a case study of Shenzhen city, China." Advances in Water Resources
860 145: 103719.

f
oo
861
862 Khan, T. A., et al. (2019). Early flood risk assessment using machine learning: A
863 comparative study of svm, q-svm, k-nn and lda. 2019 13th International Conference

r
864 on Mathematics, Actuarial Science, Computer Science and Statistics (MACS), IEEE.
865
-p
re
866 Khanam, M., et al. (2023). "Predictive Understanding of Socioeconomic Flood Impact
867 in Data-Scarce Regions Based on Channel Properties and Storm Characteristics:
lP

868 Application in High Mountain Asia (HMA)." Natural Hazards and Earth System
869 Sciences Discussions 2023: 1-27.
870
na

871 Kim, D., et al. (2023). "Application of AI-Based Models for Flood Water Level
872 Forecasting and Flood Risk Classification." KSCE Journal of Civil Engineering: 1-12.
ur

873
874 Kozcuer, C., Mollen, A., & Bießmann, F. (2024). Towards Transnational Fairness in
Jo

875 Machine Learning: A Case Study in Disaster Response Systems. Minds and
876 Machines, 34(2), 1-26.
877
878 Leal Filho, W., et al. (2022). "Deploying artificial intelligence for climate change
879 adaptation." Technological Forecasting and Social Change 180: 121662.
880
881 Li, S., Hu, G., Cheng, X., Xiong, L., Tang, G., & Strobl, J. (2022). Integrating
882 topographic knowledge into deep learning for the void-filling of digital elevation
883 models. Remote Sensing of Environment, 269, 112818.
884
885 Lin, J., et al. (2023). "NeuralFlood: an AI-driven flood susceptibility index." Frontiers
886 in Water 5: 1291305.
887
888 Lin, Y.-B., et al. (2021). "The artificial intelligence of things sensing system of real-
889 time bridge scour monitoring for early warning during floods." Sensors 21(14): 4942.
890

26
891 Liu, Z., Shi, W., Yu, Y., Chen, P., & Chen, B. Y. (2022). A LSTM-based approach
892 for modelling the movement uncertainty of indoor trajectories with mobile sensing
893 data. International Journal of Applied Earth Observation and Geoinformation, 108,
894 102758.
895
896 Liu, Z., Felton, T., & Mostafavi, A. (2024). Interpretable machine learning for
897 predicting urban flash flood hotspots using intertwined land and built-environment
898 features. Computers, Environment and Urban Systems, 110, 102096.
899
900 Liu, C., & Mostafavi, A. (2024). Floodgenome: Interpretable machine learning for
901 decoding features shaping property flood risk predisposition in cities. arXiv preprint
902 arXiv:2403.10625.
903
904 Liu, C. F., Huang, L., Yin, K., Brody, S., & Mostafavi, A. (2024). FloodDamageCast:
905 Building Flood Damage Nowcasting with Machine Learning and Data Augmentation.

f
906 arXiv preprint arXiv:2405.14232.

oo
907
908 Liu, S., Chen, X., & Di, X. (2024). Scalable Learning for Spatiotemporal Mean Field

r
909 Games Using Physics-Informed Neural Operator. Mathematics, 12(6), 803.
910
911
-p
Mahato, S., et al. (2021). "Field based index of flood vulnerability (IFV): A new
re
912 validation technique for flood susceptible models." Geoscience Frontiers 12(5):
913 101175.
lP

914
915 Melnikova, N. B., et al. (2015). "Experience of using FEM for real-time flood early
na

916 warning systems: Monitoring and modeling Boston levee instability." Journal of
917 Computational Science 10: 13-25.
918
ur

919 Merz, B., et al. (2010). "Fluvial flood risk management in a changing world." Natural
920 Hazards and Earth System Sciences 10(3): 509-527.
Jo

921
922 Mester, B., et al. (2023). "Human displacements, fatalities, and economic damages
923 linked to remotely observed floods." Scientific data 10(1): 482.
924
925 Moishin, M., et al. (2021). "Designing deep-based learning flood forecast model with
926 ConvLSTM hybrid algorithm." Ieee Access 9: 50982-50993.
927
928
929 Mulligan, M. (2013). WaterWorld: a self-parameterising, physically based model for
930 application in data-poor but problem-rich environments globally. Hydrology
931 Research, 44(5), 748-769.
932
933 Muñoz, P., et al. (2021). "Flood early warning systems using machine learning
934 techniques: The case of the Tomebamba Catchment at the southern Andes of
935 Ecuador." Hydrology 8(4): 183.
936

27
937 Müller, R., Barleben, A., Haussler, S., & Jerg, M. (2022). A Novel Approach for the
938 Global Detection and Nowcasting of Deep Convection and Thunderstorms. Remote
939 Sensing, 14(14), 3372.
940
941 Nakhaei, M., Nakhaei, P., Gheibi, M., Chahkandi, B., Wacławek, S., Behzadian,
942 K., ... & Campos, L. C. (2023). Enhancing community resilience in arid regions: A
943 smart framework for flash flood risk assessment. Ecological Indicators, 153, 110457.
944
945 Pham, B. T., et al. (2021). "Flood risk assessment using hybrid artificial intelligence
946 models integrated with multi-criteria decision analysis in Quang Nam Province,
947 Vietnam." Journal of Hydrology 592: 125815.
948
949 Pham, Q. B., et al. (2022). "Flood vulnerability and buildings’ flood exposure
950 assessment in a densely urbanised city: comparative analysis of three scenarios using

f
951 a neural network approach." Natural Hazards 113(2): 1043-1081.

oo
952
953 Pradhan, B., Lee, S., Dikshit, A., & Kim, H. (2023). Spatial flood susceptibility

r
954
955
-p
mapping using an explainable artificial intelligence (XAI) model. Geoscience
Frontiers, 14(6), 101625.
re
956
957 Puttinaovarat, S., & Horkaew, P. (2020). Flood forecasting system based on
lP

958 integrated big and crowdsource data by using machine learning techniques. IEEE
959 Access, 8, 5885-5905.
960
na

961 Pyayt, A., et al. (2011). Artificial intelligence and finite element modelling for
962 monitoring flood defence structures. 2011 IEEE Workshop on Environmental Energy
ur

963 and Structural Monitoring Systems, IEEE.

964
Jo

965 Rafiei-Sardooi, E., et al. (2021). "Evaluating urban flood risk using hybrid method of
966 TOPSIS and machine learning." International Journal of Disaster Risk Reduction 66:
967 102614.
968
969 Rahayu, R., et al. (2023). "Impact of land cover, rainfall and topography on flood risk
970 in West Java." Natural Hazards 116(2): 1735-1758.
971
972 Rahman, M., et al. (2021). "Flooding and its relationship with land cover change,
973 population growth, and road density." Geoscience Frontiers 12(6): 101224.
974
975 Rajput, A. A., Liu, C., Liu, Z., & Mostafavi, A. (2024). Human-centric
976 characterization of life activity flood exposure shifts focus from places to people.
977 Nature Cities, 1(4), 264-274.
978
979 Roh, Y., Heo, G., & Whang, S. E. (2019). A survey on data collection for machine
980 learning: a big data-ai integration perspective. IEEE Transactions on Knowledge and
981 Data Engineering, 33(4), 1328-1347.

28
 982
983 Schubert, J. E., Mach, K. J., & Sanders, B. F. (2024). National‐scale flood hazard
984 data unfit for urban risk management. Earth's future, 12(7), e2024EF004549.
985
986 Shi, W., Liu, Z., An, Z., & Chen, P. (2021). RegNet: a neural network model for
987 predicting regional desirability with VGI data. International Journal of Geographical
988 Information Science, 35(1), 175-192.
989
990 Shi, S., Zhong, Y., Liu, Y., Zhang, L., & Li, D. (2024). Cross-temporal high spatial
991 resolution urban scene classification and change detection based on a class-weighted
992 deep adaptation network. Urban Informatics, 3(1), 3.
993
994 Shu, Y., Chen, X., & Di, X. (2024). Mobility Pattern Analysis during Russia–Ukraine
995 War Using Twitter Location Data. Information, 15(2), 76.
996

f
oo
997 Stateczny, A., et al. (2023). "Optimized Deep Learning Model for Flood Detection
998 Using Satellite Images." Remote sensing 15(20): 5037.

r
999
1000
1001
-p
Sun, K., Zhou, R. Z., Kim, J., & Hu, Y. (2024). PyGRF: An improved Python
Geographical Random Forest model and case studies in public health and natural
re
1002 disasters. Transactions in GIS.
1003
1004 Suwarno, I., Ma’arif, A., Raharja, N. M., Nurjanah, A., Ikhsan, J., & Mutiarin, D.
lP

1005 (2021). IoT-based lava flood early warning system with rainfall intensity monitoring
1006 and disaster communication technology. Emerging Science Journal, 4(0), 154-166.
na

1007
1008 Tabbussum, R., & Dar, A. Q. (2021). Performance evaluation of artificial intelligence
1009 paradigms—artificial neural networks, fuzzy logic, and adaptive neuro-fuzzy
ur

1010 inference system for flood prediction. Environmental Science and Pollution Research,
1011 28(20), 25265-25282.
Jo

1012
1013 Tahvildari, N. and L. Castrucci (2021). "Relative sea level rise impacts on storm surge
1014 flooding of transportation infrastructure." Natural Hazards Review 22(1): 04020045.
1015
1016 Van den Honert, R. C., & McAneney, J. (2011). The 2011 Brisbane floods: causes,
1017 impacts and implications. Water, 3(4), 1149-1173.
1018
1019 Wang, Q. and W. Abdelrahman (2023). "High-Precision AI-Enabled Flood Prediction
1020 Integrating Local Sensor Data and 3rd Party Weather Forecast." Sensors 23(6): 3065.
1021
1022 Woodruff, J. D., et al. (2013). "Coastal flooding by tropical cyclones and sea-level
1023 rise." Nature 504(7478): 44-52.
1024
1025 Xu, C., et al. (2022). "Comprehensive analysis for long-term hydrological simulation
1026 by deep learning techniques and remote sensing." Frontiers in earth science 10:
1027 875145.
1028

29
1029 Xu, C., & Liu, W. (2022). The Spatiotemporal Characteristics and Interactions
1030 between Urban Expansion and Tidal Flat Dynamics: A Case Study of Three Highly
1031 Urbanized Coastal Counties in the Southeastern United States. Earth, 3(2), 557–576.
1032
1033 Xu, C., & Liu, W. (2023). The spatiotemporal assessments for tidal flat erosion
1034 associated with urban expansion in the conterminous coastal United States from 1985
1035 to 2015. Science of the Total Environment, 899, 165660.
1036
1037 Yang, Y., Zhang, C., Fan, C., Mostafavi, A., & Hu, X. (2020). Towards fairness-
1038 aware disaster informatics: an interdisciplinary perspective. IEEE Access, 8, 201040-
1039 201054.
1040
1041 Ye, X., Wang, S., Lu, Z., Song, Y., & Yu, S. (2021). Towards an AI-driven
1042 framework for multi-scale urban flood resilience planning and design. Computational
1043 Urban Science, 1, 1-12.

f
1044

oo
1045 Ye, X., Li, S., Gao, G., Retchless, D., Cai, Z., Newman, G., ... & Duffield, N. (2024).
1046 3D visualization of hurricane storm surge impact on urban infrastructure. Urban

r
1047 Informatics, 3(1), 1-14.
1048
1049
-p
Yuan, F., et al. (2022). "Spatio-temporal graph convolutional networks for road
re
1050 network inundation status prediction during urban flooding." Computers,
1051 Environment and Urban Systems 97: 101870.
lP

1052
1053 Yu, Y., Yao, Y., Liu, Z., An, Z., Chen, B., Chen, L., & Chen, R. (2023). A Bi-LSTM
na

1054 approach for modelling movement uncertainty of crowdsourced human trajectories

1055 under complex urban environments. International Journal of Applied Earth
1056 Observation and Geoinformation, 122, 103412.
ur

1057 Yin, K., Wu, J., Wang, W., Lee, D. H., & Wei, Y. (2023). An integrated resilience
Jo

1058 assessment model of urban transportation network: A case study of 40 cities in China.
1059 Transportation Research Part A: Policy and Practice, 173, 103687.
1060 Yin, K., & Mostafavi, A. (2023). Unsupervised graph deep learning reveals emergent
1061 flood risk profile of urban areas. arXiv e-prints, arXiv-2309.
1062 Zhou, H., Liu, L., Lan, M., Zhu, W., Song, G., Jing, F., ... & Gu, X. (2021). Using
1063 Google Street View imagery to capture micro built environment characteristics in
1064 drug places, compared with street robbery. Computers, Environment and Urban
1065 Systems, 88, 101631.
1066 Zhou, Y., et al. (2023). "Integrated dynamic framework for predicting urban flooding
1067 and providing early warning." Journal of Hydrology 618: 129205.
1068 Zhou, H., Liu, L., Wang, J., Wilson, K., Lan, M., & Gu, X. (2024a). A Multiscale Assessment
1069 of the Impact of Perceived Safety from Street View Imagery on Street Crime. Annals of the
1070 American Association of Geographers, 114(1), 69-90.
1071 Zhou, R. Z., Hu, Y., Zou, L., Cai, H., & Zhou, B. (2024b). Understanding the disparate
1072 impacts of the 2021 Texas winter storm and power outages through mobile phone location
1073 data and nighttime light images. International Journal of Disaster Risk Reduction, 103,
1074 104339.

30
Declaration of interests

☒ The authors declare that they have no known competing financial interests or personal relationships
that could have appeared to influence the work reported in this paper.

☐ The authors declare the following financial interests/personal relationships which may be considered
as potential competing interests:

of
ro
-p
re
lP
na
ur
Jo