AIoT-Enabled Urban Platform for Flood Detection and Impact Mapping: Towards Real-Time Spatial Decision Support in Disaster Management

SkTahsinHossain1Emailtahsin.hossain@qut.edu.au

TanYigitcanlar1✉Emailtan.yigitcanlar@qut.edu.au

ZhaohuiLin2Emaillzh@mail.iap.ac.cn

PengjunZhao3Emailpengjun.zhao@pku.edu.cn

1QUT Urban AI HubQueensland University of Technology2 George Street4000BrisbaneQLDAustralia

2Institute of Atmospheric PhysicsChinese Academy of SciencesNo. 40 Huayanli Qijiahuozi, Chaoyang District100029BeijingChina

3School of Urban Planning and DesignPeking UniversityNo. 100 Zhongguancun North Street, Haidian District100871BeijingChina

Sk Tahsin Hossain ¹, Tan Yigitcanlar ^1,*, Zhaohui Lin ², Pengjun Zhao ³

¹ QUT Urban AI Hub, Queensland University of Technology, 2 George Street, Brisbane, QLD 4000, Australia; tahsin.hossain@qut.edu.au; tan.yigitcanlar@qut.edu.au;

² Institute of Atmospheric Physics, Chinese Academy of Sciences, No. 40 Huayanli Qijiahuozi, Chaoyang District, Beijing 100029, China; lzh@mail.iap.ac.cn

³ School of Urban Planning and Design, Peking University, No. 100 Zhongguancun North Street, Haidian District, Beijing 100871, China; pengjun.zhao@pku.edu.cn

^* Corresponding author

Abstract

Flooding is one of the most pervasive and destructive natural hazards, with its frequency and intensity expected to worsen under climate change. While advances in geospatial analytics, Internet of Things infrastructures, and artificial intelligence have enhanced urban data ecosystems, existing smart city platforms remain limited in their capacity to provide automated, real-time flood intelligence. Most platforms focus on delineating flood extent without extending to impact assessment, leaving critical gaps in disaster response and recovery. As a result, exposure analyses of residents, buildings, and infrastructure are often conducted manually, delaying emergency services, rescue operations, and longer-term recovery planning. This study introduces an Artificial Intelligence of Things-enabled urban platform designed for flood detection, impact mapping, and spatial decision support. The conceptual architecture integrates distributed sensing, satellite imagery, and pretrained deep learning models within the Esri ArcGIS ecosystem, operationalising the principles of platform urbanism for disaster management. Demonstration of the platform draws on the March 2022 flood event in South East Queensland, with a focus on selected suburbs in Logan and the Gold Coast. Using the Prithvi–Flood Segmentation model and harmonised Sentinel-2 imagery, the workflow automates the delineation of flood extent and links outputs with exposure analytics to identify affected suburbs, railway stations, roads, buildings, and residents. Future research should focus on fine-tuning pretrained models for local contexts and scaling the architecture to incorporate additional AI-driven modules—such as road extraction, infrastructure vulnerability assessment, or population displacement modelling—thereby extending the platform’s utility across multiple hazards and governance contexts.

Keywords:

flood detection

impact mapping

disaster management

artificial-intelligence-of-things (AIoT)

platform urbanism

smart cities

1. Introduction

Flooding is one of the most pervasive and destructive natural hazards worldwide (Rentschler et al., 2022; Devitt et al., 2023; Jonkman et al., 2024). Major flood events disrupt millions of lives, cause extensive property damage, and pose severe socio-economic challenges (Manzoor et al., 2022; Yu et al., 2022; Dong et al., 2025). Between 1990 and 2022, a total of 4,713 flood events were documented across 168 countries, impacting over 3.2 billion people, resulting in 218,353 fatalities, and causing economic losses exceeding 1.3 trillion US dollars (Liu et al., 2024). Alarmingly, flood frequency and intensity are on the rise—extreme rainfall events have become more common, a trend expected to worsen with climate change (Li et al., 2021; Wasko et al., 2021; Dike et al., 2022; Maity & Maity, 2022; Chen et al.,2023). This global context underscores the urgent need for improved flood monitoring and disaster management systems to mitigate impacts on communities.

In parallel with this escalating risk, urban environments are undergoing a digital transformation (Trilles et al., 2017; Javed et al., 2022; Yigitcanlar et al., 2024). Advances in remote sensing, Internet of Things (IoT) technologies, and geospatial analytics are making cities more data-rich and sensorised than ever before (Song et al., 2021; Rajkumar et al., 2024; Destefanis et al., 2025). Despite advances in digital infrastructure and urban analytics, many smart-city data platforms remain ill-suited to flood risk management: Tools are often reactive and fragmented, with conventional models static and reliant on historical data rather than multi-source, real-time integration (Bhanye, 2025; Hlal et al., 2025). Most rely on periodically updated, siloed datasets and are not equipped to capture the dynamic nature of flood events as they unfold across space and time (Raghavan et al., 2020; Bagheri & Liu, 2025). Although sensor networks and environmental monitoring tools are increasingly deployed, their integration into comprehensive platforms capable of automated processing and analysis remains limited (Sengupta, 2024).

Beyond flood detection itself, an equally pressing gap lies in the automation of impact assessment. While Artificial Intelligence (AI) has been applied to tasks such as flood mapping, prediction and exposure modelling, its use within live, city-scale decision systems is still underdeveloped (Mohamadiazar et al., 2024; Li et al., 2025; Liu et al., 2025; Mahmood et al., 2025). Current systems rarely extend to automatic impact assessment, often relying on human intervention to determine which parts of the built environment—such as railway stations, roads, buildings, or residents—are affected or at risk (Nemni et al., 2020; Gu et al., 2024; Sharma & Saharia, 2025). This dependency slows down situational awareness and can delay critical disaster response and recovery efforts. The absence of such capabilities means that cities often lack the capacity to generate timely, spatially grounded flood intelligence that can guide emergency interventions, allocate resources, and prioritise recovery efforts.

The Artificial Intelligence of Things (AIoT) offers a promising pathway to address these gaps. By combining AI-powered analytics with distributed sensing infrastructures, AIoT enables platforms to perform real-time detection, inference, and decision support across multiple urban domains (Bibri, 2023; Hossain et al., 2025). In the context of flood risk, an AIoT-enabled platform can integrate high-resolution imagery to detect inundation, map affected areas, and assess the exposure of infrastructure and populations in near real time (Sung et al., 2022; Bhanye, 2025). This approach aligns with emerging concept of platform urbanism, where modular digital platforms serve as coordination layers for sensing, data processing, and service delivery (Barns, 2020; Repette et al., 2021; Yigitcanlar et al., 2024). However, the operationalisation of AIoT platforms for flood intelligence remains limited, particularly in regions such as Australia where decentralised governance and localised risk dynamics present additional challenges.

Against this backdrop, this study asks: How can an AIoT-enabled urban platform integrate real-time geospatial data and AI analytics for dynamic flood detection, impact mapping, and spatial decision support in disaster management? This study addresses these gaps by introducing an AIoT-enabled urban platform architecture that integrates real-time geospatial data and AI analytics for flood detection and impact assessment, enabling operational spatial decision support in disaster response. The platform is demonstrated through a case study of the March 2022 flood event in Logan and the Gold Coast, Queensland, using satellite imagery, and pretrained AI models within the Esri ArcGIS ecosystem. Queensland was selected as the demonstration site because it represents one of the most flood-prone and climatically dynamic regions in Australia, making it an ideal testbed for evaluating AIoT-enabled disaster management systems times (Cook, 2017; Callaghan, 2021; Maleki & Eslamian, 2022). By advancing from flood extent detection to automated impact analytics, this study contributes to the development of new AIoT-based methods and technologies in disaster informatics, demonstrating their practical potential for enhancing flood monitoring, impact assessment, and climate-resilient urban management.

2. Literature Background

2.1. Flood Risk in Queensland

Flooding is Australia’s second-deadliest natural hazard after heatwaves (Kamruzzaman et al., 2018; Kankanamge et al., 2020), accounting for roughly 20% of all hazard-related deaths since 1900 and Queensland is considered as one of Australia’s most flood-prone states, with a history of frequent and damaging flood events associated with La Niña events, monsoonal troughs, east-coast lows and tropical cyclones, posing significant risks to infrastructure, ecosystems, and communities (Callaghan, 2021; Levin & Phinn, 2022; Robinson et al., 2024; ACS, 2025). A statewide analysis by the Queensland Reconstruction Authority (QRA) found that 60% of local councils in Queensland face high flood risk (QRA, 2023b), and the Climate Council reports that around 70% of Queenslanders have experienced at least one flood in the past five years—the highest exposure rate of any Australian state (Climate Council, 2024). This underscores how deeply embedded flooding is in Queensland’s urban and regional landscapes.

Major floods in Queensland date back to the 19th century – for example, the 1893 “Great Flood” in Brisbane saw river heights over 8 metres, a record that stood until modern times (Cook, 2017; Maleki & Eslamian, 2022). The January 1974 Brisbane flood—driven by Cyclone Wanda—peaked at 5.45 metres at the Brisbane River, caused 14 fatalities and inundated roughly 13,000 homes across 30 suburbs, and it catalysed a new mitigation regime, including the fast-tracked construction of Wivenhoe Dam in the late 1970s–80s (Cook, 2017).

Fig. 1

Satellite view of Rockhampton inundated by Fitzroy River flooding (2011). Large areas of the city and surrounding land were submerged, illustrating the scale of riverine flood impact in Central Queensland (NASA Earth Observatory, 2011).

In the past two decades Queensland has experienced multiple major floods. The most catastrophic event was the 2010–2011 Queensland floods, a La Niña-driven flooding that saw three-quarters of the state declared a disaster zone by January 2011 (Queensland Floods Commission of Inquiry, 2012; Bromhead, 2021). Torrential monsoonal rains, intensified by Cyclone Tasha, generated record river height peaks across nearly all river systems south of the Tropic of Capricorn. Flash flooding devastated the Lockyer Valley—most notably Toowoomba and Grantham on 10 January—while the Brisbane and Bremer Rivers overflowed into Ipswich and Brisbane. In total, 33–35 people lost their lives, about 35,000 homes and 3,572 businesses were inundated, and economic losses were estimated at A$14–20 billion (Queensland Floods Commission of Inquiry, 2012; Deloitte, 2016; Y. Hou et al., 2023). Critical infrastructure was also severely affected, with 19,000 km of roads and more than a quarter of the state’s rail network damaged, alongside multiple ports. Figure 1 shows an example of the inundation extent, with floodwaters surrounding the Fitzroy River in Rockhampton during the 2010–11 event. As illustrated in Fig. 1, floodwaters can be clearly observed inundating the urban core of Rockhampton and the surrounding floodplain along the Fitzroy River. The lighter cyan tones in the near-infrared composite image represent standing water, contrasting with the darker red tones of dry urban and vegetated areas, making the inundation extent visually distinguishable despite partial cloud cover.

Another landmark disaster occurred in February–March 2022, when an unprecedented rainstorm caused record flooding in South East Queensland (SEQ). Brisbane received 792.8 millimetres in three consecutive days—exceeding the 1974 record—while Gympie’s Mary River peaked at 23 metres, it’s highest in a century (Levin & Phinn, 2022; QRA, 2023a). At least 23 people died, and more than 20,000 homes and businesses were inundated (CDP, 2022). Some suburbs recorded over 400 millimetres in 24 hours, overwhelming drainage systems. The disaster unfolded in two waves, in late February and March, and is considered Queensland’s worst flood since 2011, surpassing earlier events in some locations (Grantham, 2023; Y. Hou et al., 2023). Figure 2 provides an overview of major historical flood events in Queensland, illustrating the widespread and recurring nature of inundation across the state.

Fig. 2

Major historical flood events in Queensland, Australia, showing the spatial extent of selected inundation events between 1974 and 2019.

Note

Data source by Queensland Government Spatial Catalogue (QSpatial) Portal (Prepared by the authors).

Aside from these significant events, Queensland has suffered other severe floods in recent years. In 2013, ex-Tropical Cyclone Oswald caused major flooding in Bundaberg and across central Queensland (Callaghan, 2021). In early 2019, an extreme monsoon trough stalled over northern Queensland: Townsville was inundated by record-breaking rainfall (over 1.2 metres in a week), and flooding spread across the Gulf Country. This 2019 event caused massive livestock losses (an estimated ~ 457,000 cattle perished in north-western Queensland) and damaged over 20,000 km of roads (Queensland Government, 2019). A summary of flood events from 2010 to 2025 is presented in Table 1, compiled from Queensland Government (2024).

Table 1
Flood and extreme-rainfall events in Queensland from 2010–2025 (Queensland Government, 2024).
Year	Event	Period	Hazard / Driver
2011	QLD Monsoonal Flooding	28 Feb – Mar 2011	Statewide flooding (La Niña)
2012	SEQ Heavy Rainfall and Flooding	23–26 Jan 2012	East-coast low/heavy rain
2013	TC Oswald – rainfall and flooding	21–29 Jan 2013	Ex-tropical cyclone; multi-basin flooding
2017	TC Debbie – rainfall and flooding	28 Mar – 6 Apr 2017	Cyclone landfall; riverine and coastal impacts
2021	Central/Southern/Western QLD Rainfall and Flooding	10 Nov – 3 Dec 2021	Spring storm sequences; riverine flooding
2022	North and Central-West Rainfall and Flooding	21 Apr – 12 May 2022	Inland flooding
2022	Southern QLD Flooding	6–20 May 2022	Successive troughs; riverine flooding
2022	SEQ Rainfall and Flooding	22 Feb – 5 Apr 2022	Multi-day “rain bomb”; extreme rainfall; major urban/riverine flooding
2023	South QLD Severe Storms and Rainfall	24 Dec 2023–3 Jan 2024	Severe convective storms; flash flooding
2024	North QLD Monsoon Trough	12–22 Jan 2024	Monsoon; widespread flooding
2024	TC Kirrily – associated rainfall and flooding	25 Jan – 26 Feb 2024	Cyclone rainbands; coastal and riverine flooding
2024	Western QLD Rainfall and Flooding	22 Mar – 20 Apr 2024	Monsoonal rain; riverine flooding
2024	Diamantina Rainfall and Flooding	1–7 Jul 2024	Inland basin flooding
2024	Central and Southern QLD Rainfall and Flooding	12–14 Aug 2024	Heavy rain; riverine flooding
2024–25	Southern Summer Rainfall and Flooding	9 Dec 2024–14 Jan 2025	Severe storms; intense rainfall and flash/riverine flooding
2025	North and Far North Tropical Low	29 Jan – 28 Feb 2025	Tropical low; widespread rainfall and flooding

These recurring flood events have had profound and wide-ranging social, economic, and infrastructural consequences across Queensland. Major disasters such as the 2011 and 2022 floods caused cumulative damages exceeding tens of billions of Australian dollars, disrupted critical supply chains, and displaced thousands of residents for extended periods (Deloitte, 2016; QRA, 2023a). Repeated inundation has also led to long-term community recovery challenges—particularly in low-lying suburbs of Brisbane, Ipswich, and Gympie—highlighting persistent exposure in both urban and regional settings. For instance, the 2019 North Queensland floods devastated the agricultural sector, with the loss of nearly half a million cattle and extensive damage to rural road networks, while the 2022 South East Queensland floods severely affected housing, transport, and energy infrastructure (Callaghan, 2021; Levin & Phinn, 2022). Collectively, these impacts demonstrate that Queensland’s flood history is not only hydrologically diverse but also socio-economically significant, reinforcing its suitability as a representative case for developing and testing AIoT-enabled disaster management systems.

These events are projected to intensify under climate change, as rising temperatures increase the frequency and magnitude of extreme rainfall events, particularly in coastal and subtropical zones (Callaghan, 2021; Li, et al., 2021; Robinson et al., 2024). Projection using climate system model by the Commonwealth Scientific and Industrial Research Organisation (CSIRO) and Bureau of Meteorology (BoM) indicates that extreme rain events that cause major floods will become more likely in coming decades (BoM, 2024; CSIRO, 2024). For instance, a recent case study for the Brisbane River projected that flood magnitudes for a given return period could increase significantly under high-emission climate scenarios (Saboia & Helfer, 2024). Queensland’s guidance now embeds climate projections in flood assessments: the QRA Flood Risk Management Framework requires studies to consider “potential future conditions, including climate change”, while updated Australian Rainfall & Runoff guidance and BoM–CSIRO projections show heavier short-duration rainfall—implying today’s 1%-AEP (1-in-100) thresholds will be exceeded more often by mid-century(Cook & Harrison, 2022); for Brisbane specifically, modelling indicates the present 1%-AEP level could be 1.2–2.5 metres higher by 2050 (QRA, 2021; BoM, 2024). QRA also highlights that rapid population growth on SEQ floodplains is increasing exposure, compounding climate-driven hazard (QRA, 2023b).

In sum, Queensland’s flood risk is expected to worsen under climate change. Heavier rainfall extremes, rising sea level, and possible shifts in cyclone behaviour all point to greater flood hazards for the state. This reinforces the importance of developing AIoT-enabled systems for flood detection, impact mapping and assessment, and disaster management. Within the broader context of platform urbanism, such systems can operate as digital coordination layers that integrate diverse geospatial data streams, apply advanced analytics, and generate actionable intelligence for decision-makers.

2.2. Platform Urbanism for Flood Detection and Disaster Management

Platform urbanism frames cities as digitally mediated ecosystems, where data and analytics are coordinated through modular platforms that support decision-making across multiple domains (Leszczynski, 2020; Bibri, 2023). In flood management, this translates into “coordination layers” that integrate geospatial data from satellites and UAVs, continuous measurements from IoT sensors, and AI-driven analytics into a unified operational service (Iqbal et al., 2022; Boulouard et al., 2023). This approach reflects a shift away from static, siloed systems toward dynamic digital twins, where sensing, processing, and decision-support are tightly integrated to generate live situational awareness during crises (Shashi et al., 2023; Hlal et al., 2025). Within this paradigm, the convergence of AI and the IoT—or AIoT—has become especially significant (Samadi, 2022). These AIoT-enabled systems range from experimental research prototypes to operational platforms deployed by governments and industry. They harness networks of sensors, data from satellites and drones, numerical weather prediction models, and machine learning models to monitor floods in real time and even forecast flood events before they occur (Sung et al., 2022; Bhanye, 2025).

IoT-based sensor networks are the backbone of modern flood monitoring, using distributed water-level, rainfall, flow, pressure and camera sensors to stream high-frequency measurements via cellular, radio or LPWAN links into cloud (or hybrid) platforms for analysis and alerting (Hadi et al., 2020; Sengupta, 2024). Compared with conventional manual gauges, IoT sensors enable real-time, automated data collection across multiple sites. For instance, in Colima, Mexico, an IoT hydrometric network operated reliably through extreme events (Tao et al., 2024). Similar architectures typically leverage LoRaWAN, NB-IoT or GSM to connect remote sensors. In mountainous Indonesia, an AIoT early-warning system linked LoRa nodes and SIM900 GSM modules to a cloud server for continuous monitoring and automated SMS/app alerts (Sung et al., 2022). In Queensland, the QRA has promoted “alternative flood warning infrastructure” (radar level, ultrasonic depth and LiDAR rainfall sensors) to complement legacy gauges and transmit data in real time to cloud analytics (Joe, 2024).

Layered on IoT networks, AI and machine-learning models analyse streaming hydrometric data to generate alerts and short-lead forecasts; recent reviews show recurrent/ensemble approaches (e.g., LSTM variants) improve short-term hydrological prediction by capturing non-linear dynamics (Waqas & Humphries, 2024). At operational scale, Google’s Flood Hub exemplifies physics-informed ML that assimilates weather, satellite-rainfall and hydrological history to provide riverine flood forecasts up to seven days ahead and disseminate warnings to at-risk communities (Horn-Muller, 2023; Nearing et al., 2024). Beyond numerical prediction, AI supports imagery-led flood mapping and impact assessment: platforms and studies combine satellite/UAV and even crowdsourced photos to delineate inundation and estimate depth, with the Flood Analytics Information System (FAIS) integrating tweets, cameras and gauges for rapid situational awareness (Donratanapat et al., 2020; Samadi, 2022). Operational services such as Copernicus Emergency Management Service Rapid Mapping and NASA’s near-real-time flood products supply event-scale inundation layers that can be overlaid with building, road and population datasets to quantify exposure. Regionally, decision-support dashboards like the Iowa Flood Information System and Australian council platforms (e.g., LiXiA) fuse sensor feeds with radar/satellite layers and model outputs to support field-ready response.

These systems illustrate the benefits of platform-based architectures: they integrate diverse data sources, deliver interactive dashboards, and enhance situational awareness for decision-makers. Yet, most remain primarily focused on forecasting inundation patterns. While valuable, these outputs are only the first step in disaster management. In practice, recovery and response operations require more targeted intelligence—such as identifying which buildings, roads, or communities are directly affected, even that would be affected in the next hours or days (Safapour et al., 2021; Ogie et al., 2022). This kind of automated impact mapping is still underdeveloped in existing platforms, which often stop short of transforming raw flood maps into detailed assessments of built-environment exposure.

Automation is therefore a critical frontier for flood platforms. Rather than relying on manual overlays or case-by-case analyses, there is a growing need for systems that can continuously ingest satellite imagery and IoT data, automatically delineate floodwaters, and directly map their impacts on infrastructure and communities. Addressing this gap, the proposed Queensland platform embeds pretrained deep learning models for building footprint detection within an AIoT-enabled geospatial workflow. By automating the detection of floods, mapping of inundation, and identification of affected structures in near real time, the system operationalises the principles of platform urbanism for disaster recovery—shifting from observation to actionable intelligence and ensuring that the right information reaches responders when and where it is needed most.

2.3. Conceptual Framework for the AIoT-enabled Platforms

The AIoT provides a unifying architecture for integrating intelligent analytics with distributed sensing networks, enabling real-time monitoring and decision support (Pise et al., 2022; K. M. Hou et al., 2023). Following established frameworks in the literature, the conceptual design of the proposed platform is organised into three interdependent layers (Hossain et al., 2025).

The sensing layer incorporates diverse data sources, ranging from radar and optical satellite imagery (e.g., Sentinel-1 SAR) and UAV-based aerial surveys to in-situ IoT devices such as rainfall sensors and water-level monitors (Kuguoglu et al., 2021; Bibri et al., 2024). These heterogeneous inputs are transmitted via communication protocols like Zigbee, LoRaWAN, or 5G (Anthony, 2024; Muhammad et al., 2024). The edge/fog layer provides a first stage of geospatial preprocessing and rapid analytics, filtering noise, detecting anomalies, and enabling near real-time responsiveness even when cloud access is limited (Song et al., 2021; Jevremovic et al., 2023). The cloud/platform layer then aggregates these multi-source streams, applying advanced deep learning techniques for tasks such as flood extent extraction and object-level mapping. This layer also manages feedback loops and supports dissemination through decision-support dashboards (Pise et al., 2022; Bibri & Huang, 2025).

Extending this conceptual model, the present study develops a platform that operationalises AIoT principles within the ArcGIS ecosystem. Pretrained deep learning model from the ArcGIS Living Atlas is embedded into geospatial workflows in ArcGIS Pro, with outputs dynamically linked to dashboards in ArcGIS Experience Builder. This automated pipeline—from data ingestion and feature extraction to inundation mapping and exposure analysis—translates raw data into actionable insights for urban flood management.

The platform is demonstrated through the case of the February–March 2022 SEQ floods, one of the region’s most severe recent disasters. Logan and the Gold Coast were selected as representative sites because they experienced extensive inundation during this event, with both urban and peri-urban communities heavily affected. High-resolution satellite imagery was processed through pretrained deep learning models within the ArcGIS ecosystem to detect flood extent, which was then intersected with building footprints to identify affected structures. This case study is used for demonstration purposes only, illustrating the workflow’s capacity to automate flood detection and impact mapping; the modular design remains extensible to other regions, hazards, and datasets.

Fig. 3

Conceptual framework for AIoT-enabled platform architecture, adapted and extended from Hossain et al. (2025).

3. Platform Architecture for AIoT-Enabled Flood Detection and Impact Mapping for Spatial Decision Support

3.1. Conceptual Architecture of the Proposed AIoT-enabled Platform

Figure 4 presents the conceptual architecture of the proposed AIoT-enabled platform for flood detection and impact mapping, designed within the ArcGIS ecosystem. The workflow begins with the ingestion of data layers from ArcGIS Online or the ArcGIS Living Atlas, including suburb or neighbourhood boundaries and high-resolution satellite or aerial imagery. The suburb boundaries are first dissolved to generate a single administrative boundary for the Area of Interest (AOI). This unified polygon acts as the clipping mask to extract the corresponding imagery for the AOI, ensuring that flood detection is conducted consistently across the study region.

Flood extent is then derived using the Prithvi–Flood Segmentation deep learning model, accessed directly from the ArcGIS Living Atlas and executed through the Classify Pixels Using Deep Learning tool in ArcGIS Pro. This model produces a raster surface delineating inundated areas. To support geospatial editing and spatial analysis, the raster output is converted into polygons, which are then exported as Feature Layers. Copies of these layers are saved locally and also published as Hosted Feature Layers within ArcGIS Online. Once hosted, they are automatically linked to a Web Map, where their attributes can be queried, updated, and styled. Dashboards built in ArcGIS Experience Builder are connected to these hosted layers and refresh dynamically to visualise the flood extent in near real time. Further technical details of the deep learning model (Prithvi–Flood Segmentation) are provided in Sub-section 3.2, while requirements for the input imagery are outlined in Sub-section 3.3.

Fig. 4

Conceptual architecture of the AIoT-enabled platform for flood detection and impact mapping, designed within the ArcGIS ecosystem.

To extend analytical functionality, the platform incorporates a set of impact-mapping workflows (lower section of Fig. 2). These workflows leverage ArcGIS spatial analysis tools—primarily Select Layer by Location and Dissolve—to assess the exposure of urban assets. Specifically:

Suburb-level analysis: Flood extents are intersected with suburb boundaries to generate suburb-specific flood polygons, which are updated into corresponding Hosted Layers.

Railway station impact check: Using a 300 metres buffer zone, railway stations are classified into categories—Already Affected, At Risk, Not Affected—and the results are stored in the railway Hosted Layer.

Building impact check: Building footprints within or adjacent to flooded areas are flagged with the same categorical status, enabling asset-specific exposure mapping.

Road network impact analysis: Road segments intersecting flood zones are automatically labelled and updated in the road network Hosted Layer, supporting rapid transportation impact

Population exposure: Residential areas intersecting flood polygons are analysed to estimate potentially affected populations, recorded in the corresponding Hosted Layer.

All analytical outputs are published as Hosted Feature Layers, visualised via Web Maps, and delivered through Experience Builder dashboards. This design ensures a fully connected, modular workflow: from data ingestion and AI-based flood detection to asset-level impact analysis and real-time visualisation.

3.2. Prithvi - Flood Segmentation: A Pretrained Deep Learning Model for Flood Detection in ArcGIS Living of the World

Prithvi–Flood Segmentation (Prithvi-100M-sen1floods11) is a state-of-the-art deep learning model co-developed by NASA and IBM to automate post-event flood delineation from multispectral satellite imagery (Li et al., 2023; Esri, 2025; Hsu et al., 2025). Distributed via the ArcGIS Living Atlas (updated July 2025), the model enables pixel-level classification of inundated versus non-inundated areas, with an additional class for clouds/no-data. Its primary purpose is to support disaster response and recovery by producing rapid, repeatable flood extent maps that can be directly integrated into GIS-based impact assessments.

Prithvi-100M-sen1floods11 employs a Vision Transformer (ViT) encoder trained with a Masked Autoencoder (MAE) self-supervised strategy, adapted for semantic segmentation. The encoder is followed by a lightweight convolutional head to generate pixel-level classifications into three classes:

♣ 0 = no water

♣ 1 = water/flood

♣ -1 = no data/clouds

This design enables robust feature extraction from multi-band imagery while preserving spatial precision. The figure below (Fig. 5) illustrates the model pipeline: raw multispectral imagery is ingested and flattened, passed through the ViT encoder, decoded via convolutional layers, and compared against labelled ground-truth maps using weighted binary cross-entropy loss to refine segmentation predictions.

Fig. 5

Workflow of the Prithvi-100M-sen1floods11 flood segmentation model Esri (2025).

The Prithvi–Flood Segmentation model ingests a single six-band composite multispectral scene (raster, mosaic dataset, or image service) derived from Harmonised Landsat-8/9 (HLSL30) or Harmonised Sentinel-2 (HLSS30) level-2 products, ordered as Blue, Green, Red, Narrow NIR, SWIR-1, SWIR-2. At inference, inputs must follow this band order; cloud or missing data are captured via the ‘no data/clouds’ class. The model returns a classified raster (as above) suitable for immediate overlay with exposure layers (e.g., buildings, roads, population) for impact analysis. Given the transformer backbone and high-resolution inputs, GPU acceleration is recommended (compute capability ≥ 6.0). Although designed for global application, light local fine-tuning as has suggested by Esri (2025).

Band mapping (by sensor product; required band order at inference):

♣ Sentinel-2 / HLSS30: B2 (Blue), B3 (Green), B4 (Red), B8A (Narrow NIR), B11 (SWIR-1), B12 (SWIR-2)

♣ Landsat-8 / HLSL30: B2 (Blue), B3 (Green), B4 (Red), B5 (NIR), B6 (SWIR-1), B7 (SWIR-2)

The Prithvi–Flood Segmentation model was fine-tuned using the Sen1Floods11 dataset, a benchmark collection of 446 labelled image chips (512 × 512 pixels) drawn from 14 biomes, 357 ecoregions, six continents, and 11 major flood events. This dataset includes both Sentinel-1 SAR and Sentinel-2 optical imagery; however, for this release, the model was fine-tuned on Sentinel-2 data using a six-band composite. Labels distinguish between three classes: no water, flood water, and no data or clouds. While the original Prithvi foundation model was pre-trained on multi-temporal sequences, the flood segmentation variant was optimised for single-timestamp segmentation. This design choice simplifies data ingestion from routine harmonised Landsat and Sentinel products while still demonstrating the model’s capacity to generalise across temporal contexts.

In terms of performance, the model achieved strong results after 100 epochs of fine-tuning on Sentinel-2 imagery using an NVIDIA V100 GPU. Reported accuracy included:

♣ No water: IoU (Intersection-over-union) 96.90%, Accuracy 98.11%

♣ Flood water: IoU 80.46%, Accuracy 90.54%

♣ Aggregate: mIoU 88.68%, mean accuracy 94.37%, overall accuracy 97.25%

Importantly, the model also demonstrated robustness when tested on an unseen flood event in Bolivia, achieving a mean IoU of 86.7% and mean accuracy of 93.1%. These results indicate that the model generalises well across diverse geographic and hydrological conditions, making it suitable for global application as well as targeted regional studies such as Queensland. It should be noted that the model defines ‘water/flood’ purely on pixel segmentation of the six-band image composite, rather than applying a fixed terrestrial water-depth or surface-elevation threshold, since such a threshold is not specified in the model documentation (Jakubik et al., 2023). In this study, we adopted this model as it is integrated into the ArcGIS ecosystem (via ArcGIS Living Atlas) and therefore aligns directly with our platform’s architecture. Moreover, its open-access release by NASA/IBM and documented high generalisability across geographic regions (Li et al., 2023) provide a solid foundation for our Queensland application.

4. Case Demonstration of the Proposed AIoT-Enabled Platform Architecture

To demonstrate the functionality of the proposed AIoT-enabled platform, the workflow was applied to the task of flood detection and impact mapping using case study data from the February–March 2022 SEQ floods. This event, often referred to as the “rain bomb,” was one of the state’s most severe recent disasters, with record-breaking rainfall leading to extensive inundation across Brisbane, Logan, the Gold Coast, and surrounding catchments (Grantham, 2023). For demonstration purposes, a subset of suburbs in Logan and the Gold Coast was selected. These areas were chosen as they represent some of the worst-affected urban communities during the 2022 floods. The focus here is on methodological demonstration rather than comprehensive coverage, with scalability to broader regions considered in later discussions.

The case demonstration employs the Prithvi–Flood Segmentation deep learning model as described in the previous section. This model is designed for global flood mapping and integrates seamlessly with the ArcGIS ecosystem, enabling its deployment within an operational geospatial workflow. By combining outputs from the model with authoritative administrative boundaries and feature datasets (e.g., suburbs, road networks, buildings, railway stations, population grids), the workflow supports automated assessment of flood impacts on multiple aspects of the built environment.

For clarity, the platform architecture is divided into two major components. The core architecture is presented in Fig. 8, which illustrates the end-to-end pipeline for data ingestion, deep learning-based flood detection, and integration into dashboards for real-time decision support. This pipeline includes key stages such as boundary preprocessing, imagery cropping, model execution, raster-to-polygon conversion, feature layer publishing, and dashboard visualisation. To showcase the extended analytical capacity of the platform, five Analytical Modules are then introduced.

Figure 9 presents Module A, which illustrates the workflow for identifying suburbs where flooding has occurred. Figure 10 presents Module B, which demonstrates how the platform evaluates the exposure of railway stations. Using a spatial proximity threshold (300 metres in this study), stations are classified into categories of “Already Affected”, “At Risk”, or “Not Affected”. Figure 11 brings together the remaining analytical modules. Module C shows the procedure for determining which buildings fall within flooded areas, while Module D applies the same workflow to road networks, enabling the identification of inundated transport corridors. Finally, Module E focuses on residents, estimating population exposure by overlaying census-based demographic layers with the mapped flood extents. Together, these modules demonstrate how the proposed AIoT-enabled platform translates raw flood extent data into actionable intelligence for disaster response and recovery.

4.1. Pre-Requisites for Platform Execution

4.1.1. Web Map Setup in ArcGIS Online

For the initial implementation, the platform workflows were executed within ArcGIS Pro (version 3.4). Running the first instance in the desktop environment offered two key advantages. First, it provided an efficient and controlled setting to validate the Prithvi–Flood Segmentation model and ensure the end-to-end workflow—from flood extent extraction to exposure analysis of suburbs—was functioning correctly (Fig. 6). Second, ArcGIS Pro allowed careful preparation of the project environment, including the configuration of layer symbology, legends, and classification schemes.

Fig. 6

Initial execution of the platform within ArcGIS Pro (left), where workflows were validated and symbology configured, and subsequent publishing as an ArcGIS Online Web Map (right), which serves as the operational hub connected to interactive dashboards in ArcGIS Experience Builder.

Once the workflows were successfully tested in ArcGIS Pro, the outputs were published as a web map in ArcGIS Online to serve as the operational hub for the platform (Fig. 6). Hosting the data in the online environment is critical because it enables seamless connectivity with ArcGIS Experience Builder, where the results are disseminated through interactive dashboards and information panels. These interfaces automatically refresh as the underlying web map is updated with new data or analyses and prediction, ensuring decision-makers always have access to the most current situational picture.

4.1.2. Image Acquisition and Preprocessing

A key requirement of the platform is the preparation of harmonised multispectral imagery, as the Prithvi–Flood Segmentation model is specifically designed to operate on six-band composite rasters (Blue, Green, Red, Narrow NIR, SWIR-1, SWIR-2). These composites are derived from either the Harmonised Landsat-8 (HLSL30) or Harmonised Sentinel-2 (HLSS30) products. While the model can also process Sentinel-2 and Landsat level-2 products, its best performance is achieved with the HLS harmonised datasets due to their improved atmospheric correction, spatial co-registration, and standardised band alignment (Esri, 2025).

This study uses HLS Sentinel-2 MSI Surface Reflectance Daily Global 30 metres v2.0 (HLSS30), LP DAAC DOI 10.5067/HLS/HLSS30.002 (Masek et al., 2021). HLSS30 provides global 30-m Nadir BRDF-Adjusted Reflectance (NBAR) in Cloud-Optimised GeoTIFF (COG), tiled on the MGRS grid with daily revisit (S2A/B/C), including atmospheric correction, cloud/cloud-shadow masking, spatial co-registration, illumination/view-angle normalisation, and spectral bandpass harmonisation. For Prithvi–Flood Segmentation inference, the six bands listed above are stacked in the specified order.

Fig. 7

Example Python code for downloading HLSS30 imagery for the March 2022 Queensland flood event using GEE.

To operationalise imagery preparation, a custom Python workflow was employed using the Google Earth Engine (GEE) API and the geemap library. This script was used to search, filter, and download of HLSS30 imagery for the study’s Region of Interest (ROI), as defined in ArcGIS (parts of Logan City and Gold Coast City). An excerpt from the workflow is shown in Fig. 7 where the HLSS30 ImageCollection is filtered to the ROI.

The downloaded imagery was subsequently hosted in ArcGIS Online and linked to the platform’s processing pipeline, ensuring that the imagery layer could be seamlessly ingested by the Prithvi–Flood Segmentation model within the ArcGIS Pro environment. This manual download–upload loop served the purposes of demonstration, but it can be automated in future implementations by directly sourcing data through the ArcGIS Living Atlas or by connecting the workflow to APIs for routine ingestion, thereby strengthening its alignment with the AIoT sensing layer. By setting these imagery pre-requisites, the platform ensures that inputs are both technically compatible with the deep learning model and operationally scalable within the ArcGIS ecosystem, forming a critical foundation for the subsequent flood detection and impact analysis workflows.

4.2. Core Architecture of the Demonstrated Platform

The core workflow of the proposed AIoT-enabled platform, as presented in Fig. 8, operationalises flood detection through a sequence of geoprocessing steps in ArcGIS Pro 3.4, coupled with integration of pretrained AI models from the ArcGIS Living Atlas. The process begins with the ingestion of suburb boundaries from ArcGIS Online for Logan and Gold Coast. For demonstration purposes, 62 suburbs were selected—34 from Logan and the remainder from the Gold Coast—covering both directly inundated areas and adjacent suburbs at risk. These boundaries were first processed using the Dissolve tool, which merges them into a single administrative boundary polygon. This step is critical for efficiency, as the Prithvi Flood segmentation model can process large continuous extents more effectively than fragmented inputs, ensuring a coherent flood extent is generated across the entire AOI. The dissolved boundary was then saved locally as a shapefile for downstream use.

Next, the dissolved polygon was used to spatially constrain the imagery through the Clip Raster tool. High-resolution multispectral data was sourced from ArcGIS Online, specifically the HLS Sentinel-2 Multi-spectral Instrument Surface Reflectance Daily Global 30m (HLSS30 v2.0) dataset, which meets the Prithvi–Flood Segmentation model’s six-band input requirements (Blue, Green, Red, Narrow NIR, SWIR-1, SWIR-2). The clipped raster, exported as a GeoTIFF, was passed to the Classify Pixels Using Deep Learning tool, where the Prithvi_FloodSegmentation.dlpk model from the ArcGIS Living Atlas was applied. The model was executed with the following configuration:

♣ Model Definition: Prithvi_FloodSegmentation.dlpk

♣ Arguments: Padding = 56, Batch Size = 4, Test Time Augmentation = True, Predict Background = True

♣ Environment: GPU processor for accelerated execution

The model output was a classified raster with three classes: no water, water/flood, and no data/clouds. To prepare this for further analysis, the raster was reprocessed using the Reclassify tool, where all NODATA pixels were retained as “NODATA” and water/flood pixels were assigned the value 1. This ensured that only inundated areas were retained for vectorisation.

The reclassified raster was then converted into a polygon using the Raster to Polygon tool. This step is essential for downstream workflows, as vector data is required for spatial analysis, attribute editing, and integration with hosted feature services. The resulting polygon shapefile was enriched with attributes using the Add Fields (Multiple) tool, where fields for Suburb, City, Area, and FloodC (Check) were added.

To enable seamless publishing, the enriched shapefile was processed through the Make Feature Layer and Save To Layer File tools, creating a feature layer compatible with ArcGIS Online Web Maps. Unlike shapefiles, which cannot be directly read by hosted services, feature layers maintain schema consistency and allow for dynamic updating within the web environment. To streamline updates, the workflow employed a Delete Rows operation on the previously hosted flood extent layer, clearing outdated records while preserving the feature schema. The newly generated flood polygons were then uploaded via the Append tool, effectively refreshing the existing hosted layer without duplicating services. This approach enables the platform to continuously ingest new flood extent outputs for different suburbs or events, while ensuring schema and service continuity in line with the AIoT principle of modular, reusable pipelines.

Fig. 8

Core architecture of the AIoT-enabled platform demonstrating the end-to-end workflow from flood detection to feature layer updating within the ArcGIS ecosystem.

Together, these steps constitute the core architecture of the platform, providing an automated pipeline from data ingestion to hosted outputs in ArcGIS Online. Once published, the hosted feature layer is connected to web maps and ArcGIS Experience Builder dashboards, where results auto-refresh as updates are appended. In our demonstration setup (64 GB system RAM, NVIDIA A40-4Q GPU), the segmentation step using the Prithvi–Flood Segmentation model completed in under 10 minutes for the AOI. The following section extends this core architecture into five Analytical Modules, which operationalise impact assessment across suburbs, railway stations, buildings, road networks, and residents.

4.3. Architecture of Analytical Module A: Identifying Suburbs Affected by Flood Events

Fig. 9

Architecture of Module A for identifying suburbs affected by flood events.

Module A represents the first analytical extension (Fig. 9) occurred and quantify inundated areas at the suburb level. The process begins with an Input Function that ingests suburb boundaries from the hosted ArcGIS Online web map, alongside another Input Function retrieves the Detected Flood Extent produced through the core architecture. These two datasets are then processed using the Select Layer by Location tool under two conditions. In the first, the parameters are set as Input Features = Suburb Boundary, Selection Features = Flood_Extent, Selection Relationship = Intersect. A Calculate Field tool is subsequently applied, populating the attribute field FloodC with the value “Flooding Event”. In the second run, the same tool is used but with the Invert Spatial Relationship option enabled, allowing identification of suburbs with no overlap with the flood extent. For these, the FloodC field is updated to “No Flooding Event”. This dual operation ensures that all suburbs in the study area are classified as either impacted or unaffected.

The results of this step serve as a precondition for downstream workflows. Using the updated attribute table, the Select Layer by Attribute tool isolates only those suburbs where FloodC = “Flooding Event”. These selected features are exported via the Export Features tool, converted into a temporary layer with Make Feature Layer, and finalised using Save to Layer File. At this stage, a dedicated layer file is created that contains only suburbs where flooding has been detected. Although a new file is generated for the initial execution, subsequent workflows overwrite the existing hosted layer to maintain consistency within the web map environment.

In the next stage, the saved layer file of flooded suburbs is intersected with the Detected Flood Extent (ingested a second time within this module) using the Intersect tool. This step clips the flood polygon by suburb boundaries, generating separate flood extents for each suburb. Attribute enrichment follows, with the Calculate Field tool used to populate Suburb and City fields with values drawn from the local administrative boundary dataset. To ensure analytical efficiency, the outputs are passed through the Dissolve tool so that each suburb is represented by a single polygon capturing the total extent of inundation.

Finally, flood impact is quantified using the Calculate Geometry Attributes tool, which computes the flooded area per suburb (Area field, property = Area (geodesic), units = Square Kilometres). The resulting temporary output is again prepared as a feature layer (via Make Feature Layer and Save to Layer File) and updated into the ArcGIS Online hosted web map. This ensures that dashboards built in ArcGIS Experience Builder automatically refresh to display suburb-level flood extent and area statistics in real time.

In summary, Module A operationalises the suburb-level detection of flood impacts by systematically classifying suburbs as flooded or not, generating disaggregated inundation polygons, and calculating flood areas. This design both facilitates immediate visualisation in dashboards and provides a consistent input for subsequent modules that extend analysis to critical infrastructure and residents’ exposure.

4.4. Architecture of Analytical Module B: Assessing Railway Stations as Affected or At Risk

Module B extends the analytical capacity of the platform by evaluating the exposure of railway stations to flooding, as an example. Figure 10 illustrates the architecture of this workflow, which classifies each station as Already Affected, At Risk, or Not Affected.

The workflow begins with an Input Function that ingests railway station locations from the ArcGIS Online web map and another that retrieves the Detected Flood Extent generated by the core architecture. These datasets are processed using the Select Layer by Location tool under multiple conditions.

First, stations within 300 meters of the flood extent are identified (Relationship = Within a distance; Search Distance = 300 m). The Calculate Field tool is then applied, updating the attribute field Flood_U with the value “At Risk”. Second, stations directly intersecting the flood extent are selected (Relationship = Intersect), and their status is updated to “Already Affected”. Finally, the Invert Spatial Relationship option is used with the Within a distance (300 m) setting to capture all remaining stations outside both the flooded area and the buffer zone. These are labelled as “Not Affected” in the Flood_U field.

Unlike earlier modules, this workflow directly updates the hosted railway station feature layer within the ArcGIS Online web map. Since the attribute table of the hosted layer is modified in place, there is no need to export, create temporary shapefiles, or re-ingest layers. This ensures immediate reflection of updates across the connected dashboards in ArcGIS Experience Builder, supporting near real-time situational awareness. Importantly, while this demonstration adopted a 300-meter buffer to represent risk, this threshold can be adjusted for different contexts or criticality levels, making the workflow adaptable to varied disaster management needs.

4.5. Extension to Analytical Modules C–E: Buildings, Roads, and Residents

The same conceptual approach applied in Module B is extended across three additional modules (Fig. 11). These workflows systematically assess flood exposure for buildings (Module C), road networks (Module D), and residents (Module E), each adapted to the spatial characteristics of the respective asset.

Fig. 10

Architecture of Module B for assessing railway stations as affected or at risk.

Fig. 11

Architectures of Modules C–E for detecting affected buildings, roads, and exposed residents.

For buildings, the Select Layer by Location tool is used to determine whether structures intersect the flood extent or fall within the designated buffer zone, classifying them as Already Affected, At Risk, or Not Affected. For road networks, a parallel workflow evaluates whether road segments overlap with inundated areas or lie within proximity buffers. This enables the identification of impassable routes, potentially risky segments, and unaffected roads, supporting operational decisions such as closures, detours, and restoration priorities.

Finally, for residents, population point data from the Australian Bureau of Statistics (ABS, 2021) was intersected with the flood extent and proximity buffer to classify communities into the three exposure categories. To support this demonstration, the resident data was not used at the household level but rather randomly distributed across existing residential buildings, purely for illustrative purposes. This synthetic distribution allowed us to test the capability of the workflow to detect and label potentially exposed populations without compromising privacy or using identifiable records. It is important to emphasise that the approach was undertaken solely for research demonstration.

All other spatial datasets—including railway stations, buildings, and the road network—were sourced from the Queensland Government Spatial Catalogue (QSpatial) Portal, ensuring consistency with authoritative, publicly available sources.

4.6. Interactive Multi-Module Dashboard for Flood Detection and Impact Assessment

Fig. 12

Interactive dashboards of the AIoT-enabled platform for flood detection and impact assessment in Logan and Gold Coast, Queensland. Views represent: (a) suburb-level exposure, (b) railway stations, (c) buildings, (d) road networks, and (e) residents.

The modular workflows developed in this study—comprising the Core Architecture and Analytical Modules A through E—are unified within a single operational environment. These workflows, designed in ArcGIS ModelBuilder, are linked through preconditions so that each process executes sequentially, ensuring a smooth transition from flood detection to impact assessment. Rather than operating as isolated models, the integrated design supports a continuous pipeline that ingests imagery, runs the Prithvi–Flood Segmentation model, and extends to automated exposure analysis for suburbs, railway stations, buildings, road networks, and residents.

To make the results accessible and actionable, the outputs are published to ArcGIS Online and visualised through a five-page dashboard built in ArcGIS Experience Builder. This dashboard functions as the front-end of the platform, translating technical workflows into an interactive environment for planners, emergency responders, and policymakers. Each page corresponds to a specific analytical module: (a) suburb-level exposure, (b) railway stations, (c) buildings, (d) road networks, and (e) residents. Figure 12 illustrates the dashboard views, demonstrating how the platform transforms raw geospatial data and deep learning outputs into intuitive, decision-ready intelligence.

By consolidating complex geoprocessing into a streamlined, interactive interface, the dashboard embodies the principle of AIoT-enabled disaster intelligence. It bridges the gap between automated back-end analysis and real-time front-end decision support, offering a practical tool for disaster management and post-flood recovery planning in Queensland and beyond.

4.7. Limitations and Scalability of the Platform

The proposed AIoT-enabled platform demonstrates strong potential for advancing flood detection and impact mapping, but like any emerging system, it operates within certain limitations. One key consideration is the availability and timeliness of imagery. For this study, we utilised Harmonised Sentinel-2 (HLSS30) imagery accessed via GEE, which is fully compatible with the Prithvi–Flood Segmentation model. However, in real-world scenarios, especially immediately after a flood event, suitable cloud-free images may not always be available in the public domain. This can delay rapid flood assessment at the very moment when speed is most critical. Subscription-based services (e.g., Planet, ICEYE, Maxar) or locally maintained aerial repositories can overcome this constraint, though they may introduce additional financial or technical requirements.

Another limitation is tied to the event-driven nature of floods, which makes scheduling workflows less straightforward than for continuous monitoring systems. Automated scheduling (e.g., daily or weekly runs) may either generate redundant outputs during dry periods or miss the onset of a flood. A more effective strategy is to link the platform with hydrological and meteorological monitoring or flood forecasting systems, so that workflows are triggered when water levels or rainfall thresholds are exceeded. This coupling would reduce computational overhead and ensure the platform activates only when operationally relevant.

At the same time, the platform’s scalability offers exciting opportunities. Its modular design in ArcGIS ModelBuilder allows new analytical modules to be integrated without disrupting existing pipelines. Beyond flood segmentation, the architecture can incorporate pretrained models from the ArcGIS Living Atlas, such as the building footprint extraction model, as well as road and land-cover extraction models. Integrating these alongside Prithvi–Flood Segmentation outputs would enable near real-time updates of both the built environment and hazard exposure layers, offering a more holistic picture of urban vulnerability. Additionally, the architecture is designed to accommodate the integration of meteorological station observations (e.g., rainfall, gauge water levels) and forecast model outputs (e.g., weather and hydrological predictions). Such data streams could enable the platform to shift from post-event mapping toward proactive preparedness by triggering the segmentation and impact-analysis workflows based on exceedance of rainfall or water-level thresholds.

Finally, because the platform is grounded in the principles of AIoT, it can evolve into a broader multi-source decision-support ecosystem. Additional data sources—such as IoT river gauges, rainfall sensors, UAV imagery, and predictive meteorological and hydrological models—can be integrated into the workflow, enriching situational awareness and extending the system from post-event analysis towards proactive resilience planning and adaptive governance.

5. Conclusion

This paper has presented an AIoT-enabled urban platform for flood detection and impact mapping, designed to support real-time spatial decision-making in disaster management. The demonstrated workflow moves beyond conventional flood extent mapping by incorporating analytical modules that quantify impacts on the built environment and residents. By linking flood extents to suburbs, buildings, transport networks, and population datasets, the platform enables decision-makers not only to identify where flooding has occurred but also to assess who and what is at risk—a crucial step for prioritising emergency response, coordinating recovery, and planning long-term resilience.

The platform’s architecture exemplifies the convergence of AIoT with geospatial analytics. Prithvi–Flood Segmentation model provides the automated detection of inundation, while ArcGIS ModelBuilder orchestrates a modular workflow that integrates detection with impact mapping. The result is an operational framework where sensing, processing, and decision support are tightly coupled, consistent with emerging paradigms of platform urbanism. Importantly, the platform demonstrates that AIoT is not only about data collection but also about delivering actionable intelligence that supports residents and communities in times of crisis.

A key contribution of this study is the demonstration that damage and exposure assessments can be automated in near real time. By identifying flooded suburbs, detecting at-risk rail stations, mapping affected buildings and roads, and estimating resident exposure, the system provides the types of insights traditionally generated through slow, manual assessments. This capacity is particularly relevant for disaster recovery and humanitarian response, where timely information on displaced populations, damaged infrastructure, and affected households can accelerate relief operations and guide resource allocation. With further integration of household-level records from local governments, the platform could be extended to support resident-level evacuation planning and welfare checks, representing a major advance in disaster informatics.

Looking forward, several opportunities remain to extend the platform’s capabilities. Fine-tuning the Prithvi–Flood Segmentation model on Queensland-specific imagery would enhance local accuracy, particularly in heterogeneous coastal and inland environments. Beyond flood detection, integration with additional pretrained AI models available in the ArcGIS Living Atlas would enable continuously updated baselines of the built environment against which flood impacts can be dynamically assessed. In addition, coupling the platform with hydrological and meteorological simulation or forecasting models represents a key area for future exploration. Such integration would enable proactive flood prediction and scenario modelling, transforming the system from a post-event detection framework into a forward-looking decision-support tool for early warning and preparedness.

Future iterations of the platform should also incorporate urban topographic features and drainage-related infrastructure—such as rainwater pipelines, parks, and open green spaces—to better represent runoff dynamics and natural water retention capacity. Including these parameters would allow more precise simulation of flood propagation and urban drainage efficiency within AI-driven analyses. Furthermore, imbedding the urban hydrometeorological prediction products with combined AI and numerical model techniques will provide more early warning information, and further enhance the capability of platform. This would expand the platform from a single-hazard tool to a multi-layered urban intelligence system capable of supporting a wide range of resilience and recovery planning tasks.

To ensure resilient and trustworthy deployment, the platform can also incorporate security safeguards as it scales. These include encrypted data transfer, role-based access controls in ArcGIS Online, and audit logging to protect the integrity of hosted layers. At the AIoT level, sandboxed execution of models and fallback workflows can mitigate risks from corrupted imagery or adversarial inputs, while privacy-preserving protocols ensure responsible handling of sensitive exposure data. Together, these measures will enable secure, scalable adoption of the platform in operational disaster management, ensuring that decision-makers receive timely, actionable, and reliable intelligence when it matters most.

References

ACS (2025) Floods. Australian Climate Service (ACS). https://www.acs.gov.au/pages/6212cfadf42c4c44becef7182ead8bb6

Anthony B (2024) Decentralized AIoT based intelligence for sustainable energy prosumption in local energy communities. Cities 152:105198

Bagheri A, Liu G (2025) Climate change and urban flooding: assessing remote sensing data and flood modeling techniques. Environ Reviews 33:1–14

Barns S (2020) In: Barns S (ed) Making sense of platform intermediation. Springer, Platform Urbanism, pp 99–125

Bhanye J (2025) Flood-tech frontiers: smart but just? Discover Global Soc 3(1):59

Bibri S (2023) The metaverse as a virtual model of platform urbanism. Smart Cities 6(3):1345–1384

Bibri S, Huang J (2025) Artificial intelligence of things for sustainable smart city brain and digital twin systems. Environ Sci Ecotechnol 26:100591

Bibri S, Huang J, Jagatheesaperumal S, Krogstie J (2024) The synergistic interplay of artificial intelligence and digital twin in environmentally planning sustainable smart cities. Environ Sci Ecotechnology 20:100433

BoM (2024) State of the climate 2024. Bureau of Meteorology (BoM). https://www.bom.gov.au/state-of-the-climate/

Boulouard Z, Ouaissa M, Ouaissa M, Krichen M, Almutiq M, Algarni M (2023) Streamlining river flood prevention with an integrated AIoT framework. 20th ACS/IEEE International Conference on Computer Systems and Applications

Bromhead H (2021) Disaster linguistics, climate change semantics and public discourse studies. Lang Sci 85:101381

Callaghan J (2021) Weather systems and extreme rainfall generation in the 2019 north Queensland floods compared with historical north Queensland record floods. J South Hemisphere Earth Syst Sci 71(1):123–146

CDP. (2022) 2022 Australian flooding. Center for Disaster Philanthropy (CDP). https://disasterphilanthropy.org/disasters/2022-australian-flooding/

Chen G, Ling J, Lin Z, Xiao Z, Li C (2023) Role of tropical-extratropical interactions in the unprecedented 2022 extreme rainfall in Pakistan. Atmos Res 291:106817

Climate Council (2024) State of Queensland: Disaster ground zero. https://emergencyleadersforclimateaction.org.au/reports_resources/state-of-queensland-disaster-ground-zero/

Cook M (2017) The Australia day floods, January 1974. Environment & Society Portal, vol 15. Rachel Carson Center for Environment and Society

Cook M, Harrison E (2022) Ten years on: Brisbane’s compounding flood risk. In: Lukasiewicz A, O’Donnell T (eds) Complex Disasters: Compounding, Cascading, and Protracted. Springer, pp 101–121

CSIRO (2024) Australia's climate, global climate and projected future changes. https://www.csiro.au/en/research/environmental-impacts/climate-change/State-of-the-Climate/Report-at-a-Glance

Deloitte (2016) The economic cost of the social impact of natural disasters. https://www.preventionweb.net/publication/economic-cost-social-impact-natural-disasters

Destefanis T, Guliyeva S, Boccardo P, Fissore V (2025) Advancing flood detection and mapping. Remote Sens 17(17):2943

Devitt L, Neal J, Coxon G, Savage J, Wagener T (2023) Flood hazard potential reveals global floodplain settlement patterns. Nat Commun 14(1):2801

Dike V, Lin Z, Fei K, Langendijk, Gaby, Nath, Debashis (2022) Evaluation and multimodel projection of seasonal precipitation extremes over central Asia based on CMIP6 simulations. Int J Climatol 42:7641

Dong B, Xia J, Wang X (2025) Comprehensive flood risk assessment in highly developed urban areas. J Hydrol 648:132391

Donratanapat N, Samadi S, Vidal J, Sadeghi Tabas S (2020) A national scale big data analytics pipeline to assess the potential impacts of flooding on critical infrastructures and communities, vol 133. Environmental Modelling & Software, p 104828

Esri (2025) Prithvi - flood segmentation. https://www.arcgis.com/home/item.html?id=29dc90c33daf402caa9293c2088d1057

Grantham S (2023) #bnefloods: An analysis of the Queensland Government media conferences during the 2022 Brisbane floods. The Australian Journal of Emergency Management, 38(2), 42–48

Gu J, Xie Z, Zhang J, He X (2024) Advances in rapid damage identification methods for post-disaster regional buildings based on remote sensing images. 14(4):898

Hadi M, Yakub F, Fakhrurradzi A, Hui C, Azizan A (2020) Designing early warning flood detection and monitoring system via IoT. Earth Environ Sci 479(1):012016

Hlal M, Baraka Munyaka J, Chenal J, Azmi R, Adraoui M (2025) Digital twin technology for urban flood risk management. Remote Sens 17(17):3104

Horn-Muller A (2023) Google's AI-enabled flood forecasting goes global. https://www.axios.com/2023/05/22/googles-ai-flood-forecast

Hossain T, Yigitcanlar T, Nguyen K, Xu Y (2025) Platform urbanism for resident safety. Urban Clim 61:102445

Hou K, Diao X, Shi H, Ding H, Zhou H, de Vaulx C (2023) Trends and challenges in AIoT/IIoT/IoT implementation. Sensors 23(11):5074

Hou Y, Wei Y, Wu S, Li J (2023) Mapping the social, economic, and ecological impact of floods in Brisbane. Water 15(21):3842

Hsu C, Li W, Wang S (2025) Geospatial foundation models for image analysis. Int J Geogr Inf Sci 39(9):2096–2125

Iqbal U, Riaz B, Barthelemy M, J., Perez P (2022) Artificial intelligence of things (AIoT)-oriented framework for blockage assessment at cross-drainage hydraulic structures. Australasian J Water Resour, 1–11

Jakubik J, Fraccaro P, Borges D, Muszynski M, Weldemariam K, Zadrozny B, Ganti R, Mukkavilli K (2023) Prithvi-100M-sen1floods11: Flood mapping model (IBM-NASA geospatial). Model card, Hugging Face

Javed A, Shahzad F, Rehman S, Zikria Y, Razzak I, Jalil Z, Xu G (2022) Future smart cities. Cities 129:103794

Jevremovic A, Kostic Z, Perakovic D (2023) Energy-efficient edge intelligence. Mobile Networks and Applications

Joe (2024) The QRA’s alternative flood warning infrastructure (rain and level gauges. Queensland%20Reconstruction%20Authority%20,and%20the%20Bureau%20of%20Meteorology. https://aqxsolutions.com.au/2024/11/22/flood-iot-qra-guidelines/#:~:text=The%20

Jonkman S, Curran A, Bouwer L (2024) Floods have become less deadly. Nat Hazards 120(7):6327–6342

Kamruzzaman M, Deilami K, Yigitcanlar T (2018) Investigating the urban heat island effect of transit-oriented development in Brisbane. J Transp Geogr 66:116–124

Kankanamge N, Yigitcanlar T, Goonetilleke A, Kamruzzaman M (2020) Determining disaster severity through social media analysis. Int J Disaster Risk Reduct 42:101360

Kuguoglu B, van der Voort H, Janssen M (2021) The giant leap for smart cities. Sustainability 13(21):12295

Leszczynski A (2020) Glitchy vignettes of platform urbanism. Environ Plann D: Soc Space 38(2):189–208

Levin N, Phinn S (2022) Assessing the 2022 flood impacts in Queensland combining daytime and nighttime optical and imaging radar data. Remote Sens 14(19):5009

Li C, Dike V, Lin Z, Gao X (2021) Projected changes in precipitation extremes over jiulongjiang river basin in coastal Southeast China. Front Earth Sci 9:771592

Li W, Lee H, Wang S, Hsu C, Arundel S (2023) Assessment of a new GeoAI foundation model for flood inundation mapping Proceedings of the 6th ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery, Hamburg, Germany

Li Z, Zhou Z, Wang H, Li X, Jia H (2025) Artificial intelligence-incorporated prediction for urban flooding processes in the past 20 years, vol 192. Environmental Modelling & Software, p 106525

Liu Q, Du M, Wang Y, Deng J, Liu J (2024) Global, regional and national trends and impacts of natural floods, 1990–2022. Bull World Health Organisation 102(6):410–420

Liu Z, Coleman N, Patrascu F, Yin K, Li X, Mostafavi A (2025) Artificial intelligence for flood risk management. Int J Disaster Risk Reduct 117:105110

Mahmood M, Wang X, Aziz F, Dogulu N (2025) Urban flood modelling, vol 191. Environmental Modelling & Software, p 106507

Maity S, Maity R (2022) Changing pattern of intensity–duration–frequency relationship of precipitation due to climate change. Water Resour Manage 36(14):5371–5399

Maleki M, Eslamian S (2022) History of world extreme floods. Flood Handbook. CRC

Manzoor Z, Ehsan M, Khan M, Manzoor A, Abioui M (2022) Floods and flood management and its socio-economic impact on Pakistan. Frontiers in Environmental Science, pp 10–2022

Masek J, Ju J, Roger J, Skakun S, Justice C (2021) HLS Sentinel-2 Multi-spectral Instrument Surface Reflectance Daily Global 30m v2.0

Mohamadiazar N, Ebrahimian A, Hosseiny H (2024) Integrating deep learning, satellite image processing, and spatial-temporal analysis for urban flood prediction. J Hydrol 639:131508

Muhammad A, Shangguan D, Rasool G, Khan A, Mukhtar M (2024) A localized evaluation of surface water quality using gis-based water quality index along Satpara Watershed Skardu Baltistan, Pakistan. ISPRS Int J Geo-Information 13(11):393

NASA Earth Observatory (2011) Flooding around Rockhampton, Queensland. NASA Earth Observatory. https://earthobservatory.nasa.gov/images/48429/flooding-around-rockhampton-queensland?

Nearing G, Cohen D, Dube V, Gauch M, Matias Y (2024) Global prediction of extreme floods in ungauged watersheds. Nature 627(8004):559–563

Nemni E, Bullock J, Belabbes S, Bromley L (2020) Fully convolutional neural network for rapid flood segmentation in synthetic aperture radar imagery. Remote Sens 12(16):2532

Ogie R, James S, Moore A, Dilworth T, Amirghasemi M, Whittaker J (2022) Social media use in disaster recovery. Int J Disaster Risk Reduct 70:102783

Pise A, Almuzaini K, Ahanger T, Farouk A, Nuagah S (2022) Enabling artificial intelligence of things (AIoT) healthcare architectures and listing security issues. Computational Intelligence and Neuroscience, 2022(1), 8421434

QRA (2021) Queensland flood risk management framework. https://www.qra.qld.gov.au/QFRMF

QRA (2023a) 2021-22 Southern Queensland floods - recovery operation. https://www.qra.qld.gov.au/2021-22-Southern-Queensland-Floods

QRA (2023b) Queensland statewide assessment of flood risk factors. http://www.qra.qld.gov.au/flood-risk/SAFRF

Queensland Floods Commission of Inquiry (2012) Queensland and Brisbane 2010/11 floods. https://knowledge.aidr.org.au/resources/flood-queensland-2010-2011/

Queensland Government (2019) The north-west Queensland Monsoon event of 26 January – 9 February 2019. https://www.publications.qld.gov.au/ckan-publications-attachments-prod/resources/156cf3d2-7709-41d1-bdcc-69abd2006603/nw-monsoon-survey-report.pdf?ETag=f700f2309345ae9b23b795ce5a0eb3c4

Queensland Government (2024) History of floods - Queensland flood history. Queensland Government. https://www.qld.gov.au/emergency/dealing-disasters/disaster-types/flood/flood-aware/queensland-flood-history

Raghavan S, Simon B, Lee Y, Tan W, Kee K (2020) Data integration for smart cities. Computational Science and Technology, Singapore

Rajkumar N, Viji C, Latha P, Vennila V, Shanmugam S, Pillai N (2024) The power of AI, IoT, and advanced quantum based optical systems in smart cities. Opt Quant Electron 56(3):450

Rentschler J, Salhab M, Jafino B (2022) Flood exposure and poverty in 188 countries. Nat Commun 13(1):3527

Repette P, Sabatini-Marques J, Yigitcanlar T, Sell D, Costa E (2021) The evolution of city-as-a-platform. Land 10(1):33

Robinson C, Barnes M, Narsey S, Reeder M (2024) The meteorology of the 2019 North Queensland floods. Q J R Meteorol Soc 150(761):1971–1987

Saboia M, Helfer F (2024) Design rainfalls under climate change scenarios in South East Queensland, Australia, vol 55. Urban Climate, p 101919

Safapour E, Kermanshachi S, Pamidimukkala A (2021) Post-disaster recovery in urban and rural communities. Int J Disaster Risk Reduct 64:102535

Samadi S (2022) The convergence of AI, IoT, and big data for advancing flood analytics research. Front Water, Volume 4–2022

Sengupta S (2024) IoT-based flood detection and management systems in urban areas. Risk Assess Manage Decisions 1(2):301–313

Sharma N, Saharia M (2025) DeepSARFlood: Rapid and automated SAR-based flood inundation mapping using vision transformer-based deep ensembles with uncertainty estimates. Sci Remote Sens 11:100203

Shashi, Centobelli P, Cerchione R, Ertz M, Oropallo E (2023) What we learn is what we earn from sustainable and circular construction. J Clean Prod 382:135183

Song T, Cai J, Chahine T, Li L (2021) Towards smart cities by internet of things (IoT). J Knowl Econ 12(2):1–17

Sung W, Devi I, Hsiao S (2022) Early warning of impending flash flood based on AIoT. EURASIP Journal on Wireless Communications and Networking, 2022(1), 15

Szukalski B (2024) Overview of pretrained AI models in ArcGIS Living Atlas. https://community.esri.com/t5/arcgis-living-atlas-blog/overview-of-pretrained-ai-models-in-arcgis-living/ba-p/1367078

Tao Y, Tian B, Adhikari B, Zuo Q, Luo X, Di B (2024) A review of cutting-edge sensor technologies for improved flood monitoring and damage assessment. Sensors 24(21):7090

Trilles S, Calia A, Belmonte Ó, Torres-Sospedra J, Montoliu R, Huerta J (2017) Deployment of an open sensorized platform in a smart city context. Future Generation Comput Syst 76:221–233

Viswambharan V, Singh R (2020) Introducing pretrained geospatial deep learning models. https://www.esri.com/arcgis-blog/products/arcgis/imagery/introducing-ready-to-use-deep-learning-models

Viswambharan V, Singh R (2025) Pretrained nodels in ArcGIS. https://www.esri.com/arcgis-blog/products/arcgis/geoai/task-and-generalized-pretrained-models

Waqas M, Humphries UW (2024) A critical review of RNN and LSTM variants in hydrological time series predictions. MethodsX 13:102946

Wasko C, Nathan R, Stein L, O'Shea D (2021) Evidence of shorter more extreme rainfalls and increased flood variability under climate change. J Hydrol 603:126994

Yigitcanlar T, Xia B, Cortese TT, Sabatini-Marques J (2024) Understanding city 4.0. Sustainability, 16(1), 326

Yu Q, Wang Y, Li N (2022) Extreme flood disasters. Water 14(8):1211

Statements & Declarations

Funding:

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Competing Interests:

The authors have no relevant financial or non-financial interests to disclose.

Author Contributions: Sk Tahsin Hossain: Data collection, processing, investigation, analysis, and writing – original draft; Tan Yigitcanlar: Supervision, conceptualisation, writing - review & editing; Zhaohui Lin, Pengjun Zhao: Writing - review & editing. All authors have read and agreed to the final version of the manuscript.