Real-Time, Crowdsourcing-Enhanced Forecasting of Building Functionality During Urban Floods

Preprint. Under Review . Xie et al. | arXiv | 1 / 24 Real - Tim e , Cro w d s o u rci n g - Enhanced For ecasting of Building Funct ionality Durin g Urban Floods Lei Xie 1 , Peihui Lin 1 , N a iyu W ang 1 * , Paolo Gardoni 2 1 College of Civil Engine ering and Architecture, Zheji ang University , Hangzhou, Zhejiang 3 10058, China. 2 University of Illinois Urba na - Champaign, IL, USA * Correspondence: naiyuwang@zju.edu.cn ABSTRACT : Urban flood emergency response increasingly reli es on infrastructure impact forecasts rather than hazard variables alone. How ever , real - time predictions are unreliable due to biased rainfall, incomplete flood knowledge, and sparse observations. Conventional open - loop forecasting propagates impacts without adjusting the system state, causing errors dur ing c ritical decisions. This study presents CRAF (Cr owdsourcing - Enhanced Real - Ti m e A w a r e n e s s a n d F o r e c a s t i n g ) , a p h y s i c s - informed, closed - loop framew ork that converts s parse human - sensed evidence into rolling , decision - grade impact forecasts. By coupling physics - based simulation learning with crowdsourced observations, CRAF infer s system conditions from incomplete data and propagates them forward to produce multi - step, real - time predictions of zone - level building functionality loss without online retraining. This closed - loop design supports continuous state correction and forward prediction under weakly structured data w ith low - latency operation. Offlin e evaluation demonstrates stable gener alization across diverse storm scenarios. In operational deployment during T yphoon Haikui (2023) in Fuzhou, China, CRAF reduces 1 – 3 hour - ahead forecast errors by 84 – 95% relative to fixed rainfall - driven forecasting and by 73 – 80% relative to updated rainfall - driven forecasting, while li miting computation to 10 minute s per update cycle. These results show that impact - state alignment — rather than hazard refinement alone — is essential for reliable real - time decision support, providing a pathway toward operational digital twins for resil ient urban infrastructure systems. K EYWORDS : Building functionality; Crowdso urcing; Dee p - learning; Impact - based forecasting; Rolling updates; Spatiotemporal forecasting; Urban floods 1 INTRODUCTION Urban flooding, intensified by climate change and rapid urbanization, is occurring with increasing fr equency and severity worldwide, posing escalating risks to public safety , infrastructure performance, an d the continuity of essential services . Recent events highlight the magnitude of this challenge: the 2012 Beijing rainstorm caused 79 fatalities and RMB 1 1.64 billion in economic losses (Zhu et al. 2022) , while the 2021 floods in Germany resulted in 189 fatalities and damages exceeding €40 billion (Meyer a nd Johann 2025; Zander et al. 2023) . Beyond these direct losses, such events reveal systemic vulnerabilities in how urban infrastructure systems are monitored, assessed, and managed during rapi dly evolving hazard s. While a dvances in m eteorological and hydrological modeling have improved the prediction of hazard variables such as rainfall and inundation, emergency decisions are rarely made on hazards alone. Instead, operational actions — such as evacuation, shelter activation, and res ource allocation — depend on rel iable estimates of impacts, including los s of building functionality , accessibility disruptions, and service interruptions. From a systems and reliability perspective, the critical challenge is therefore not onl y to forecast hazards, but to estimate and continuously updat e the evolving impact state of the built environment in real time. T ranslating uncertain and rapidly changing hazards into reliable, rolling decision - relevant impact forecasts remains a central y et unresolved problem in flood risk management. Preprint. Under Review . Xie et al. | arXiv | 2 / 24 In practice, real - time impact forecasting is hindered by multiple sources of uncertainty . Precipitation forecasts during extreme events often exhibit substantial bias and variability , which propagate through hydrologic and hydraulic models and degrade down stream consequence predictions (Mazzoleni et al. 2017; Guan et al. 2023; Songchon et al. 2023 ). Flood - conditioning factors — such as drainage capacity , surface roughness, and antecedent conditions — are imperf ectly observed and difficult to calibrate during un folding events (Su et al. 2025) . Meanwhile, conventional monitoring netwo rks are designe d primarily fo r hazard variables and lack the spatial density and temporal responsiveness required to capture the heterogeneous progression of impacts across urban communities (Farahmand et al. 2023; Lee and T ien 2018; W ang et al. 2018a; W ang et al. 2018b; Panakkal and Padgett 2024) . As a result, impact forecasts f requently div erge fr om reality precisely during the narrow time windows when timely and reliable situational awareness is most critical. Crowdsourced obs ervations — including social media posts, citizen reports, and online news — provide a complementary source of real - time information that is abundant, low - cost, and often available with minimal latency during disaster evolution ( Fohringer et al. 2015; Rossi et al. 2018; Songchon et al. 2023 ). These human - sensed data streams can reveal localized evidence of flooding and d isruption that is not captured by traditional sensors ( Ta b l e 1 ) . Y et effectively translating these heteroge neous reports into o perational impact forecasting remains challenging. Crowdsourced data are inherently unstructured (text, images, and video), unevenly di stributed in space and time, and subject to reporting noise and ambiguity . These characteristics limit the ef fectiveness of classical data assimilation methods (e.g., Kalman - filter variants) that rely on structured measurements with well - characterized uncertainties ( Mazzo leni et al., 2017; Annis & Nard i, 2019; Songchon et al., 2023 ), as well as purely data - driven regression models that require dense, consistently labeled training data ( Yu a n e t a l . , 2 0 2 3 ; Safaei - Moghadam et al., 2024 ). Consequently , most existing approaches employ crowdsourcing to ref ine hazard estimations, r ather than to directly update system - level impact state forecasts that are most relevant to real - time decision - making ( Assumpç ão et al. 2021 ). Ta b l e 1 . Comparison of crowdsourcing - informed flood forecasting studies with CRAF Study Domain Physics - Tr a i n e d P r i o r (Offline) Crowd Assimilation (Online) Ta r g e t = Functionality/ Impact Methodol ogy Mazzoleni et al ., 2017 Flood ✗ ✓ ✗ (streamflow) Semi - Distributed Hydrological Model + Kalman fi lter Annis & Nardi , 2019 Flood ✗ ✓ ✗ (water depth/extent) FLO - 2D PRO + Ensemble Kalman filte r Restrepo - Estrada et al., 2018 Flood ✗ ✓ ✗ (streamflow) Social media → Rainfall proxy → Proba bilit y Dist ribut ed Model Songchon et al., 2023 Flood ✗ ✓ ✗ (water depth/extent) LISFLOOD - FP + Ensemble Kalman filte r Yu a n e t a l . , 2 0 2 3 Road flooding ✗ ✗ ✓ (road flooding risk) ML model (RF /AdaBoost ) trained on crowd labels Safaei - Moghadam et al., 2024 Road flooding ✗ ✗ ✓ (road flooding risk) ML model (RF /XGBoost/ SVC) trained on flood alerts This work (CRAF) Building functionality ✓ (encodes inter - ERZ correlation) ✓ (uses prior to fuse crowd) ✓ (building functionality loss) Graph - based situation awareness + spatiotemporal f orecasting (physics - supervised) Notes ： ✓ = present; ✗ = absent. Physics - trained prior encodes inter - zone i mpact correlations learned from large - scale simulations for online inference (beyond simply including physics - inspired features). Crowd assimil ation incorporates real - time human - sensed observations to update forecast s. Function ality/impact denotes operat ionally meaningful outcomes (e.g., building functionality , road operability) rather tha n hydraulic variables. Preprint. Under Review . Xie et al. | arXiv | 3 / 24 These limita tions sugge st that reliable real - time impact forecasting requires more than improved hazard prediction or isolated observations. Instead, it calls for a closed - loop, system - level framework in which impacts are treated as a lat ent state of the infras tructure network that is repeatedly estimated, corrected, and propagated as new evidence becomes available. Such a formulation enables continuous alignment between model predictions and evolving field conditions, thereby enhancing the reliability of forecasts used for operational decisions . In contrast, conventional open - loop approaches propagate impacts forward based solely on forcing inputs, allowing errors to accumulate and reducing their usefulness during critical respons e periods. Motivated by this need, this study introd uces CRAF (Crowdsourcing - Enhanced Real - Tim e A w a r e n e s s a n d Forecasting), a physics - informed framework that performs closed - loop impact - state alignment and forecasting, transforming sparse human - sensed evidence into rolling, decision - grade impact predictions ( Fig. 1 ). Using zone - level building functionality loss (ZFL) as a representative, decision - relevant metric, CRAF formulates real - time impact forecasting as a sequential state estimation and propagation problem, in which the evolving sys tem state is continuousl y reconstructed and updated as new observations emerge. The framework operates through three tightly coupled functional streams. Crowdsourced Impact Monitoring (CIM) extracts sparse, time - stamped impact cues from heterogeneous online sources. Situati onal A wareness (SA) assimilates these observations through a p hysics - trained prior to infer spatially coherent impact states across all zones, even under severe observation sparsity . Spatiotemporal Forecasting (STF) then propagates the calibrated state forward under rainfall forcing to generate rol ling, multi - step forecasts. By separating offline physics - sup ervised learning from online low - latency inference, CRAF enables repeated state correction and forecasting without retraining, thereby providing a stable, operationally viable, and data - efficient closed - loop capability for real - time impact forecasting. This st udy advan ces real - time flood impact forecasting throu gh three complementary developments. First, it int roduces a closed - loop impact - state forecasting paradigm that continuously aligns predictions with evolving field evidence, improving reliability under uncertainty . Second, it develops a phy sics - informed forecasting framework that integrates crowdsourced observations with simulation - based knowledge to enable robust state estimation when data are sparse, heterogeneous, and incomplete. Third, the frame work is demonstrated through operational deployment during a real flood event, demonstrating substantial reductions in short - horizon forecast errors during decisi on - critical periods. T ogether , these elements establish a practical foundation for real - time, impact - based decision support and more resi lient urban infrastructure systems. The re mainder of this paper is organized as follows. Section 2 presents the problem formulation and defines the impact variable and spatial units. Section 3 describes the datasets, physics - based simulations, and preprocessing procedures. Section 4 details the design and training of the CRAF modules. Secti on 5 evaluat es operational performance through a real - world case study . Section 6 concludes with key findings and future research directions. 2 PROBLEM FORMULATION AND FRAMEWORK 2.1 Spatial Unit and Predictive Impact V ariable To o p e r a t i o n a l i z e f l o o d - induced impacts for emergency management, we focus on building function ality loss as the target impact variable. Building functionality represents the ability of the building to maintain structurally safe occupancy , provide the inte nded services to the tenants (e.g., potable water and power), and have physical access under flooding, accounting for inundation, physical damage, utility disruptions, and accessibility constraint s ( Chavez et al ., 2025; Lin & Wa n g , 2 0 1 7 ; N o c e r a & G a r d o n i , 2019; Sun & Cha, 2022 ). Loss of functionality directly informs evacuation and shelter decisions and therefore serves as a decision - relevant indicator for emergency response ( Xie et al. 2025b; Guidotti et al. 2019; Lamadrid et al. 2025 ). Preprint. Under Review . Xie et al. | arXiv | 4 / 24 Fig. 1 Conceptual over view of the CRAF fra mework. Emergency actions are typical ly coo rdinated at aggregated spatial units rather than at the level of individual buildings. Accordingly , the study area is partitioned into Emergency Response Zones (ERZs), each comprising a cluster of geographically proximate buildings with similar flood - conditioning charact eristics. Let the ERZs be index ed b y ℳ = {1, … , 𝑀 } , where 𝑀 is the total number of ERZs. For ERZ 𝑚 ∈ ℳ containing 𝑁 ! buildings, the zone - level functionality loss (ZFL) at time 𝑡 , d enoted by 𝑧 ! " . ∈ [ 0, 1 ] # # , is defined as the average probability that buildings within the zone are in non - functional states requirin g evacuation or external assistance: 𝑧 ! " = . 1 𝑁 ! 2 2 𝑝 $ %" & ! # & ! '( ) " $ *+ . ∈ [ 0, 1 ] # # (1) Here 𝑝 $ %" & ! is the probability that building 𝑖 is in functionality state 𝐹 , at time 𝑡 , and ℒ denotes the set of non - functional states that mandate evacuation action (as detailed in Section 3.3 ). Ve c t o r i z i n g a c r o s s a l l E R Z s y i e l d s t h e s y s t e m - level impact state at time 𝑡 : 𝑍 " = ( 𝑧 + " , 𝑧 - " , … , 𝑧 . " ) ∈ [ 0,1 ] # . (2) This vector - valued represent ation defines the fundamental impact state to be sequentially estimated, updated, and propagated throughout the framewor k. 2.2 Impact Forecasting Problem Description The objective of this paper is to fore cast the spatiotemporal evolution of flood - induced impact states, represented by ZFL at ERZ resolution, in real time, accounting for errors and incompleteness in meteorological forcing, flood - conditioning at tributes, and observational evidence. Preprint. Under Review . Xie et al. | arXiv | 5 / 24 Let 𝐻 denote the look - back window and 𝐹 the forecast horizon. At each forecast initialization time 𝑡 , the framework ingests three categories of in puts: • Rainfall for cing. Observed rainf all over the look - back window , 𝑅 " / 0 1+2" , and forecast rainfall over the prediction horizon, 𝑅 " 1 +2" 1& , concatenated into a unified rainfall sequence 𝑅 " / 0 1+2" 1& . • Static ERZ - level attributes. Flood - conditioning features 𝑊 ∈ ℝ # . 34 # , where 𝑑 5 denotes the number of flood - conditioning features capturing spatial heterogeneity in flood susceptibility , and ERZ - level building attributes 𝐸 ∈ ℝ # . 34 $ , where 𝑑 6 denotes the number of building attributes describing the composition of the built environment. • Crowdsour ced observations. A c o l l e c t i o n o f crowdsourced posts 𝑃 7 % = {𝑃 8 ∣ 𝜏 ∈ 𝒯 " } , where 𝒯 " ⊆ [𝑡 − 𝐻 + 1, 𝑡 ] denotes the set of time indices within the look - back window where crowd observations are available. The forecastin g task is to map these heterogeneous inputs into a multi - step forecast trajectory of the im pact state: ℱ : I 𝑅 " / 0 1+2" 1& , 𝑊 , 𝐸 , 𝑃 7 % J → 𝑍 L " 1 +2" 1& (3) This formulation highlights two defining characteristics of real - time impact forecasting: (i) the current impact state is latent and must be inferred from sparse, noisy observations, and (ii) forecast trajectories must remain robust to rainfall - forcing uncertainty while supporting low - latency updates. Impact forecasting is therefore na turally viewed as a sequential state estimation and propagation process, rather than a one - shot direct prediction of future impacts. Under this formulation, open - loop methods propagate impact states using rainfall forcing alone, whereas closed - loop methods incorporate recurrent state correction when new impact observations become available. 2.3 Overview of the CRAF Framework CRAF i s des igned as an integrated three - module impact - state estimation and forecasting sy stem: CIM Module ( 𝒇 9:; ) — observation operator . From the collection of raw crowdsourced posts .𝑃 7 % = {𝑃 8 ∣ 𝜏 ∈ 𝒯 " } , the CIM module extracts time - and location - specific inundation cues from heterogeneous sources (text, images, or video) and converts them into sparse ERZ - level impact observations 𝑂 7 % = {𝑂 8 ∣ 𝜏 ∈ 𝒯 " } . At each time τ , a single observation is represented as 𝑂 8 = { ( 𝑚, 𝑧 ! 8 ) |𝑚 ∈ 𝑂 8 < } , where 𝑂 8 < ⊂ ℳ denotes the subset of ERZs for which reliable crowd evidence is available. The observation operat or is defined as: 𝑂 8 = 𝑓 =>. ( 𝑃 8 ) ， .𝜏 ∈ 𝒯 " (4) CIM does not attempt to produce spat ially complete impact fields; instea d, it provides sparse but informative measurements of the underlyi ng impact state. The CIM module pro duces deterministic ERZ - level impact observations and does not expl icitly model obs ervation noise or uncertainty . Instead, potential errors in the extracted observations are mitigat ed through a strict quality - control process , allowing th e resulting observations to be treated as high - confidence inputs in the subsequent data assimilation framework. SA Module ( 𝒇 ?@ ) — spatial impact state completion . The SA module transforms sparse observations into a spatially complete and physical ly coherent impact state estimate. Given observations 𝑂 8 and static ERZ attr ibutes ( 𝑊 , 𝐸 ) , the SA module infers a dense ZFL snapshot 𝑌 8 = 𝑓 AB ( 𝑂 8 , 𝑊 , 𝐸 ; 𝐺 + ) ∈ [ 0,1 ] # . (5) where 𝐺 + = ( 𝑉 # , 𝐴 + ) is an ERZ graph whose adjacency matrix encodes int er - zone impact co - variation learned from physics - based simulations. For observed zones 𝑚 ∈ 𝑂 8 < , the in ferred state satisfies ( 𝑌 8 ) ! = 𝑧 ! 8 ; for unobserved zones, values ar e inferred through graph - based propagation. Its parameters are learned offline from physics - generated ZFL trajectories under simulated sparse - observation conditions. STF M odule ( 𝒇 ?CD ) — spatiotemporal impact state propagation . The STF module propagates calibrated impact states forwa rd in time under rainfall forcing. Given rainfall inputs 𝑅 " / 0 1+2" 1& and a history of impact states composed of inferred states 𝑌 7 % and predict ed states at unobserved times, the STF module produces rolling multi - step forecasts: 𝑍 L " 1 +2" 1& = 𝑓 AE& I 𝑅 " / 0 1+2" 1& , X 𝑍 L 7 % & Y 𝑌 7 % Z ; 𝐺 - . J (6) Preprint. Under Review . Xie et al. | arXiv | 6 / 24 where 𝒯 " F ⊆ [𝑡 − 𝐻 + 1, 𝑡 ] denotes time steps without crowd observations, ∥ denotes temporal concatenation, and 𝐺 - = ( 𝑉 # , 𝐴 - ) is a graph encoding correlations among ERZ - level ZFL temporal trajectories. The STF module is trained offline using physics - generated ZFL sequences and corresponding rainfall forcing, ensuring consistency between spatial inference and temporal propagation . 2.4 Offline Physics - Supervised Learning and Online Forecast Operation CRAF integrates offline physics - supervised learning with onl ine impact - state inferenc e to ena ble rolling real - time forecasts. Prior to deployment, the SA and STF modules a re traine d offline usi ng large ensemb les of physics - based flood and building - function ality simulations. These simulations generate spatiotemporally correlated ZFL trajectories under diverse rainfall and flood - conditioning scenarios, allowing the models to learn (i) persistent spatial co - variation of impacts across ERZs and (ii) temporal evolution patterns of functionali ty loss conditioned on rainfall forcing. Once trained, the parameters of both modules are fi xed and remain unchanged during real events. During a flood event, CRAF operates entirely in online inference mode ( Fig. 2 ). The CIM module continuously ingests heterogeneous public posts and converts them into sparse ERZ - level impact observations. These obser vations are assimilated by the pre - trained SA module to infer a spatially complete estimate of the curr ent impact state, which is then propagated forward by the pre - trained STF module to generate rolling multi - step forecasts. As new crowd evidence becomes available, impact states and forecasts are updated sequentially without retrai ning. This repeated CIM→SA→STF cycle closes the loop between sensing, state estimation, and rainfall - conditioned propagation, enabling continuous realignment of the latent impact state with evolving real - world condition s while maintaining low - latency operation. Fig. 2 The application flowchart of CRAF framework for dynami c forecast updates . 3 DATA DESCRIPTI ON AND PREPARATION FOR MODEL TRAINING This section describes the data requi red to instantiate and train the CRAF framework. Consi stent with its system - oriented design, CRAF relies on a limited set of broadly available data categories rather than location - specific or bespoke inputs. These data support Preprint. Under Review . Xie et al. | arXiv | 7 / 24 representation of static exposure and flood - conditioning characteristics, dynamic meteorological forcing, physics - based supervisory labels for of fline learning, and real - time observational evidence for online impact - state updating. To i l l u s t r a t e t h e p r a c t i c a l i m p l e m e n t a t i o n , t h e framework is instantiated for a representa tive flood - prone urban community in Fuzhou, China. While specific datasets are drawn from this case study , the framework itself is not tied to any particula r city , se nsor network, or simulation platform. 3.1 General Data Requ irements of the CR AF Framework Application of the CRAF framework requires four categories of data, corresponding directly to the modeling components introduced in Sect ion 2. (1) ERZ - level building attributes Building and infrastructure data are required to characterize the composition of the built environment within each ERZ. At a minimum, thes e attri butes include building typology or archetype proportions, which enable aggregation of building - level im pacts into ZFL. These attri butes are treated as static features and represented as an ERZ - level matrix 𝐸 ∈ ℝ . 34 $ . (2) Meteor ological for cing Ti m e - varying meteorological inputs — primarily rainfall — drive impact - state propagation in the forecasting module. Rainfall is represented as hourly sequences spanning a historical look - back window 𝐻 and a forecast horizon 𝐹 , denoted as 𝑅 ∈ ℝ G0 1& H3) ' , where 𝐻 + 𝐹 is the sequence length and 𝑁 I is the number of rainfall inputs. Data may or iginate from gauges, radar products, or numerical weather prediction outputs. (3) Flood - conditioning variables Flood - conditioning variables capture static spatial heterogeneity governing flood accumulation and impact patterns, including topography , drainage characteristics, and proximity to waterways. These variables are aggregated at the ERZ level to form a condit ioning matrix 𝑊 ∈ ℝ . 34 # . (4) Physics - based simulation outputs Offline training of the SA a nd STF modules relies on physics - based flood and impact simulations to gener ate supervisory labels. These simulations provide internally consistent spatiotemporal trajectories of ERZ - level functionality loss under diverse rainfall scenarios. The framework does not depend on a specific hydrodynamic or damage model; any approach capable of producing coherent flood depth and impact estimates may be used. Notably , o ffline training does not require historical flood events and can use synthetic o r design rainfall scenarios that adequately span local intensity – duration – frequency characteristics. While less realistic than observed data, these scenarios provide a practical alternative for regions with limited historical records. To g e t h e r , t h e s e d a t a c a t e g o r i e s d e f i n e t h e m i n i m u m information required to deploy CRAF in a new urban setting. The specific datasets used in the Fuzhou case study are summarized in Ta b l e 2 . 3.2 Te s t b e d C o m m u n i t y a n d D a t a I n s t a n t i a t i o n The urban area of Fuzhou, China, is used to demonstrat e data preprocessing, offline training, and real - time deployment of the CRAF framework. As shown in Fig. 3a , a residential building inventory of 553 potentially impacted structures in the study area is classified into four dominant archetypes: high - rise, multi - story , detached villa, and overlay villa. Buildings are gr ouped into 𝑀 =50 ERZs based on spatial proxi mity and flood - conditioning characteristics (Y ang et al. 2021, Zhou et al. 2024) , yielding resp onse units consistent with emergency management practice ( Fig. 3b ). For each ERZ, building archetype proportions are aggregated to form the ERZ - level attribute matrix E ( Fig. 3c ). Meteorologi cal forcing is derived from multiple historical typhoon events to cover a range of rainstorm intensities for offline training, yielding 𝑁 I =752 rainstorm sequences of 24 h length; the demonstration event (i.e., T yphoon Haikui, 2023) is designated as the independent test set to avoid information leakage. Flood - conditioning variables include elevation - derived features and proximity indicators related to river networks and drainage outlets. These variables are initially computed at high spatial resolution and subsequently aggr egated to the ERZ level t o form the conditioning matrix W . Preprint. Under Review . Xie et al. | arXiv | 8 / 24 Ta b l e 2 . Summary of data sources for the Fuzhou case study Data categor y Indicator Symbol Resolution Sources ERZ - specific building attributes Proportions of archetype 𝐸 ∈ ℝ !"# ! - Amap (lbs.ama p.com) Meteorol ogical variables Hourly rain fall sequences 𝑅 ∈ ℝ $%& '( ") " 1 hour Rainfall seq uences of 9 hi storical typh oon events* at 2122 meteorological stations from Fujian Provincial Flood Information Release Syste m ( https://slt.fujian. gov . cn/ ) Flood conditioning variables Digital Ele vation M od el (DEM) 𝑊 ∈ ℝ !"# # 30m Geospatial Data Cloud ( www .gsc loud.cn ) Aspect 30m Tr a n s f o r m a t i o n f r o m D E M d a t a b a s e d o n ArcGIS softwa re Curvature 30m Slope 30m To p o g r a p h i c w e t n e s s index (TWI) 30m Proximity to river network - Vi s u a l i n t e r p r e t a t i o n f r o m G o o g l e r e m o t e sensing imagery Proximity to drainage outlet - Fuzhou W ater Resources Bureau Supervisory labels ZFL t rajectories 𝑍 ∈ ℝ $%& '( "! Hourly per ERZ Physic - based simulation (detailed i n Section 3.3 ) * Notes: The 9 typhoon even ts — Soudelor (2015), Meranti (201 6), Megi (2016), Nesat ( 2017), Haitang (2017), Lekima (2019), Lupit (202 1), Doksuri (2023), and Hai kui (2023) — provides 72 - hour rainfall se quences (24 h pre - landfall to 48 h post - landfall). Sequences are segmented using a 24 - hour sliding window wit h 12 - hour stride, retai ning only subsequences with accumulated rainfall e xceeding 350 mm. Fig. 3 (a) Building inventory in Fuzhou study area; (b) spatial distribution of ERZs; (c) archetype composition of each ERZ. Preprint. Under Review . Xie et al. | arXiv | 9 / 24 3.3 Physics - Based Simulation for Offline Supervisory T raining Physics - based simulations are employed to generate supervisory data for offline training of the SA and STF modules. In the testbed implementatio n, a two - stage simulation workflow is adopted ( Fig. 4 ). First, urban flood inundation is simulated using a coupled 1D – 2D hydrodynamic framework that integrates the Storm W ater Management Model (SWMM) with the W eighted Cellular Automata 2D model (WCA2D) , resolving both drainage network flow and surface inundation under historical rainfall sequences ( Ta v a k o l i f a r e t a l . , 2 0 2 1 ; Z e n g e t a l . , 2 0 2 2 ). Second, building functionality is assessed through a three - level fragility analysis, progressing from Needs - Related Components (NRCs) to individual dwelling units and fina lly to the entire residential building, thereby establishing dep th - to - functionality mapping criteria. Simulated water depths are then translated into probabilistic building functionality s tates — considering physical damage, uti lity dis ruption and transpor tation accessibility — generating a seven - level functionality state (I – VII) probability vector f or each building (Xie et al., 2025a) . ERZ - level functionality is subsequently derived by aggregating building - level probabilities of non - functional states ( ℒ ={III, … , VII}) within each ERZ, resulting in ZFL trajectories following the formulation in Section 2 . Repeating this process across all histo rical rainstorm scenarios yields an ensemble of spatiotemporally correlated ZFL trajectories, which provide structured and physically consistent supervisory labels for offline training of both the SA and STF modules. Importantly , within CRAF , physics - based simulation serves solely as a s ource of prior knowledge for model training and is not used for real - time forecasting during deployment. Fig. 4 Flowchart of the physics - based si mulation workflow for generating ZFL t rajectories. Preprint. Under Review . Xie et al. | arXiv | 10 / 24 4 DEVELOPMENT OF CRAF FRAMEWORK This section details the technical implemen tation of CRAF, focusing on the architectures of its constituent modules and the associa ted learning strategy . Although CRAF operates as a un ified observation – state – forecast process during deployment, its modules are designed under a strict separation between offli ne learning and online inference. The SA and STF modules are trained offline using physics - based flood and functionality simulations (Section 3.3) to learn persistent spatial and temporal impact dependencies, while real - time operation is limited to forward inference and sequenti al state updating. The following subsections describe each module and its r ole in the end - to - end system. 4.1 Crowdsour ced Impact Monitori ng (CIM): Observation Operator The CIM module serves as the online observation operator of CRAF. Its role is to convert noi sy, heterogeneous crowdsourced data into sparse but quality - controlled ERZ - level observations of ZFL. As illustrated in Fig. 5 , CIM is implemented as a modular, multimodal perception and infe rence framework, rather than a monolithic end - to - end network. First, flood - related content is collected from near – real - time social media and web platforms (in cluding Weibo, TikTok, Tence nt Docs, and va rious news portals) using keyword - driven crawlers ( Table 3 ). Second, a relevance filtering step removes non - flood - related posts and retains only entries containing credible spatiotemporal and inundation cues. Textual relevance is determined using a fine - tuned BERT - based model ( Kaur & Kaur, 2023 ), while content from other sources unde rgoes m anual verification. Third, fine - grained flood attribute mining derives time, location, and water depth c ues for each post. Specifically, location in text data is derived from check - ins, if available, or inferred via Point of Interest (POI) - based regular expression (regex) matching. Images and videos are ma nually interpreted to identify flooding scenarios and landm arks, which are then georeferenced by cross - check ing with base maps or street - view data to derive coordinates. Flood depth cues in texts are also extracted via regex matching. Meanwhile, images and videos are processed through a physics - informed, rule - based mapping layer that translates qualitative inundation cues into quantitative flood - depth estimates using a hierarchical reference framework ( Table 4 ). By combining location and flood depth, these point - based estimates are spatially interpolated via a constrained region - growing algorithm ( Lin et al., 2023 ). Finally, water depths are mapped to ERZ - level ZFL using the same depth – functionality relationships adopted in the physics - based simulations ( Section 3.3 ), ensuring consistency between offline supervisory labels and online observation s. To main tain re liability in th e infer ence pi peline, the CIM mod ule i ncorporates manual inspection and validation, correcting mis labeled or ambiguous entries and filtering out outliers or low - credibility report s before flood attribute extraction. By ensuring that only high - quality observations enter the pipeline, CIM avoids explicit error modeling, which would require large sample sizes and is impractical for sparse or unevenly distributed social media data, making manual quality control a key step for accura te ERZ - level functionality inference. Ta b l e 3 . Summary of flood - related keywords Category Keywords Related to r ainfall rainstorm ( 暴雨 ) ， heavy rain ( 大雨 ) ， Related to t yphoon typhoon ( 台风 ) ， Haikui ( 海葵 ) Related to f looding flooding ( 洪涝 ) ， inunda tion ( 淹 ) ， waterlogging ( 积水 ) ， waist - deep ( 齐腰深 ) ， knee - deep ( 齐膝深 ). Preprint. Under Review . Xie et al. | arXiv | 11 / 24 Ta b l e 4 . Flood - depth hierarchical mapping strategy Reference object Level Part Flooding depth estimation Reference Human body (average human height of 1.7m) A Ankle 0.1 (Chaudhary et al. 2020; Y an et al. 2024) B Calf 0.3 C Knee 0.45 D Thigh 0.64 E Wa i s t 0.85 F Chest 1.28 G Neck 1.49 Shared bicycle A Center of th e wheel 0.3 Self - measured B To p o f t h e w h e e l 0.5 C Saddle 0.6 Car A Center of th e tire 0.33 (Hao et al. 2022; Songchon et al. 2023; Y an et al. 2024) B To p o f t h e t i r e 0.66 C Door handle 0.8 Notes: The f looding depth es timatio n repres ents st atisti cal aver ages bas ed on ex pert kno wledge. Fig. 5 CIM module archit ecture and its hi erarchical flood a ttributes extraction mechanism. 4.2 Situational A wareness (SA): Physics - Informed Spatial State Completion The SA module performs spat ial impact - state completion, inferring a spatially complete and physically coherent ZFL snapshot from sparse E RZ - level observations. This step provides consistent initial conditions for downstream spatiotemporal forecasting and mitigates spatial artifacts arising from uneven observation coverage. Preprint. Under Review . Xie et al. | arXiv | 12 / 24 SA operates on an ERZ graph 𝐺 + = ( 𝑉 # , 𝐴 + ) , where nodes represent ERZs and edges encode persistent inter - zone impact dependencies learned offline fr om physics - based simulations. Using simulated ZFL trajectories (Section 3.3), pairwise inter - ERZ depen dencies are quantified and thresholded to construct a sparse adjacency structure that serves as a physics - informed structural prior during online inference. SA is implemented as a multi - layer Graph Attention Network (GAT) performing masked node regression over the ERZ graph as illustrated in Fig. 6(a) . Each ER Z node is associated with a feature vector comprising: (i) the observe d ZFL value 𝑧 # # when available, otherwise an inverse - distance - weighted pr ior 𝑧 \ ; (ii) aggregated building archetype proportions E ∈ ℝ # . 34 $ ; and (iii) selected flood - conditioning variables 𝑊 < ∈ ℝ # . 34 # ∗ , including DEM, curvature, distances to rivers and drainage outlets . In the GAT, stacked multi - head attention layers propagate infor mation via attention - weighted message passing, enabling inference a t unobserved nodes conditioned on both local observations and physics - informed spatial dependencies. Observed nodes are clam ped to reported values, whil e unobserved nodes are estimated through learned attention mechanisms. Offline training uses physics - generated ZFL snapshots with synthetically masked observations to emulate realistic crowd - data spar sity ( Fig. 6(b) ). The training objective minimizes reconstruction error exclusively on masked nodes, promoting robust spatial inference under observation scarcity. Separate SA models are trained for different obs ervation densit ies and reused unchanged during real - time de ployment. Fig. 6 SA module (a) architecture and (b) trai ning strategy Preprint. Under Review . Xie et al. | arXiv | 13 / 24 4.3 Spatiotemporal For ecasting (STF): Rainfall - Conditioned State Propagation The STF module performs rainfall - conditioned temporal state propagation, generating rolling multi - step forecasts of ERZ - level ZFL by advancing the completed impact state forward in time. As shown in Fig. 7 , STF couples rainfall forcing encoding with a spatiotemporal graph neural forecasting backbone. Given the rainfal l sequence 𝑅 " / 0 1+2" 1& , a temporal encoder extra cts forcing features over the look - back and forecast hor izon. These features are fused with the historical ZFL state sequence, consisting of SA - completed snapshots at observed times and model - imputed states at unobserved times. Th e fus ed representation is processed by stacked spatiotemporal convolution (ST - Conv) blocks, which combine temporal gated convolutions (e.g., GLU - based caus al convolutions) with graph - based spatial aggregation to capture cross - ERZ propagation dynamics ( Yu et al. 2018 ) . Spatial coupling is defined on an ERZ graph 𝐺 - = ( 𝑉 # , 𝐴 - ) , where 𝐴 - encodes persistent inter - zone co - variation learned f rom physics - generated training trajectories. The output layer applies two temporal gated convolutions and a 2 - D convolution to produce single - step forecasts. STF is trained offline using physics - generated rainfall – impact pairs ( R , Z ) under a sliding - window sampling strategy ( Fig. 8 ). A 13 - hour window with a 1 - hour stride is applied to each 24 - hour sequence, with zero - padding at the start. This yields 24 training pairs per sequence, for a total of 18,048 samples. The model is optimized for one - step - ahead prediction, w hile multi - step forecasts during both training and deployment are obtained via autoregressive rollout, consistent with real - time rolling - forecast operation. Section 4 establishes CRAF as a modular , tightly integrated, physics - informed impact forecasting system. System performance is evaluated next from two complementary perspectives: controlled physics - based simulation experiments to establish baseline behavior and gene ralization ( Section 5 ), followed by real - world application during an operational flood event to demonstrate practical ef fectiveness under uncertainty ( Section 6 ). Fig. 7 STF module architecture. Preprint. Under Review . Xie et al. | arXiv | 14 / 24 Fig. 8 Sliding - window train ing and autoregres sive rollout st rategy for STF . 5 RESULTS I - OFFLINE VERIFICATION USING PHYSICS - BASED SIMULATIONS This section presents an offline verification of the CRAF framework using physics - based simulations. The objective is to assess the reliability , robustness, and cross - event generalization of the closed - loop forecasting process under controlled conditions, isolating the intrinsic behavior of the system prior to real - time deployment. By evaluating spatial state reconstruction and temporal propagation against known ground tr uth, these experiments establish whether the framework can provide stable and physicall y consistent forecasts independent of crowdsourced noise and operational uncertainties. All numerical exp eriments are conduc ted using the physics - generated flood and building - functionality trajectories described in Section 3.3 . Model training, validation, and testing rely exclusively on simulated data with known ground truth, enabling syst ematic assessment of spat ial inference and temporal propagation behavior without confounding observational noise. 5.1 Ve r i f i c a t i o n o f S A M o d u l e U n d e r S p a r s e Observations This experiment evalu ates the ability of SA module to reconstruct spatially complete ZFL fields from sparse and uneven ERZ - level observations. Physics - generated ZFL snapshots serve as ground truth. T o emulate realistic observation scarcity , ERZ - level observations are synthetically masked at observation ratios of 10%, 20%, 30%, 40%, and 50%, with observed ERZs sele cted un iformly a t rando m for each snapshot. Masked ERZs are treated as unobserved and inf erred solel y through the physics - trained situational prior . The physics - generated dataset designates T yphoon Haikui as the hold - out test set, with the remaining samples split into training and validation (7:3) for hyperparameter calibration. Performance is quantified using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), comp uted exclusively over masked ERZs. Ta b l e 5 summarizes reconstruction errors across training, validatio n, and testing under progressively sparse observations. Mean MAE ranges from 0.024 to 0.051 and RMSE from 0.048 to 0.093 across all Preprint. Under Review . Xie et al. | arXiv | 15 / 24 observation ratios. Importantly , errors remain low even when only 10 – 20% of zones are observed , as also shown in Fig. 9 , indicating that the learned spatial dependencies provide reliable inference under severe information scarcity . This beha vior suggests that the physics - informed prior supplies a stable structural constraint, enabling consistent state estimation across storms rather than event - specific fitting. Such robustness is critical for operational settings where observations are incomplete and uneven ly distributed. Ta b l e 5 . MAE an d RMSE of the SA model under sparse ERZ - level observations Fig. 9 SA model performance on the test set (T yphoon Haikui) acr oss varying ERZ - level observations. 5.2 STF Module Te m p o r a l S t a b i l i t y a n d F o r c i n g Consistency of Spatiot emporal Forecasting This experiment examines the temporal behavior of the STF module under mult i - step autoregressive rollout, to assess (i) the forecast stability over increasing lead times, and (ii) whether the forecast evolution is genuinel y conditioned on meteorological fo rcing rather than dominated by autoregressive persistence. Figs. 10 summarizes the performance of the STF module on the held - out T yphoon Haikui, presenting both temporal and spatial error patterns. Fig.10(a) shows the MAE evolution over a 24 - hour iterative forecast window , indicating that errors increase smoothly during early lead times and stabilize at longer horizons. It also compares simulated and predicted ZFL values at multiple ty pical forecast horizons, showing that strong linear agreement is maintained across lead times without drift or mean collapse. Although Prediction dispersion increases gradually with lead time, as expected, no systematic bias amplification is observed. Fig.10(b) displays the corr esponding spatial distr ibution of forecast errors across ERZs, with most zones maintaining low errors and a 24 - h average below 1.52%. To g e t h e r , t h e s e r e s u l t s d e m o n s t r a t e t h a t f o r e c a s t errors grow gradually and remain bounded over extended lead times, indicating stable multi - step propagation without numerical drift or bias amplification. The abs ence of error escalation and the preservation of spatial structure suggest that the model maintains physically consist ent dynamics during iterative rollout, a prerequisite for dependable rolling forecasts in real - time operations. From a reliab ility standpoint, this stab ility ensures that forecast qu ality degrades predictably rather than abruptly , supporting continuous situational awareness for emer gency decision - making. To v e r i f y t h a t S T F f o r e c a s t s a r e g e n u i n e l y conditioned on meteorological forcing, a rainf all - removal ablation (STF - NR) is conduc ted in which rainfall inputs are excluded while all other model Error metric MAE RMSE O bs . r atio 10% 20% 30% 40% 50% 10% 20% 30% 40% 50% T rain set 0.051 0.032 0.033 0.032 0.03 3 0.09 3 0.061 0.060 0.058 0.05 8 V alidation set 0.04 2 0.02 7 0.02 5 0.024 0.025 0.083 0.055 0.048 0.04 9 0.048 Te s t set ( Hai kui ) 0.044 0.030 0.030 0.030 0.0 3 0.087 0.06 2 0.060 0.059 0.058 Preprint. Under Review . Xie et al. | arXiv | 16 / 24 components are held fixed. As shown in T abl e 6, removing rainfall forcing results in an order - of - magnitude degradation in both MAE and RMSE across all horizons and produces unstable error growth. This sharp deterioration confirms that forecast skill is governed primarily by physically meaningful rainfall – impact relationships rather than autoregressive persistence. Consequently , the STF module preserves causal consistency with external forcing, a necessary condition for reliable forecasting under evolving me teorological conditions. Fig. 10 (a) T emporal evolution of prediction errors (MAE) over a 24 - hour iterative forecas t window; (b) Spatial distribution of MAE across ERZs, with 24 - hour average values. Preprint. Under Review . Xie et al. | arXiv | 17 / 24 Ta b l e 6 . Effect of removing rainfall forcing on multi - step ZFL forecast accuracy under the LOSO pro tocol. Error metric Model Forecast horizons (hours ahead) +6 h +12 h +18 h +24 h MAE STF 0.0024 0.0077 0.0131 0.0155 STF - NR 0.3712 0.5245 0.4095 0.2661 RMSE STF 0.0102 0.0191 0.0247 0.0293 STF - NR 0.4804 0.6031 0.5138 0.3810 Overall, the offline verification demons trates that CRAF provides stable, physicall y co nsistent, and transferable spatial – temporal forecasting behavior across storms. The fr amework maintains reliable state reconstruction under sparse observations and stabl e multi - step propagation under dynamic forcing, establishing a dependable foundati on for subsequent real - time, closed - loop deployment . Section 6 extends this evaluation to a clo sed - loop setting, examining its performance under real - world operational c onstr aints. 6 RESULTS II - ONLINE DEMONS TRATION UNDER REAL - WORLD CONDITIONS 6.1 Demonstration Set up and Operational Context This section evaluates the op erational behavior of the CRAF framework during a real flood event, assessing end - to - end system reliability under realist ic conditions characterized by uncertain rainfall forcing, sparse observations, and time - critical decision constraints. The demonstrat ion is conducted using the severe rainstorm associated with T yphoon Haikui (5 – 6 September 2023) in Fuzhou, China. This event is held out entirely from training and validation and serves as an independent operational test. Its rapid rainfall i ntensification, substantial numerical weather prediction bias, and uneven availability of real - time observations create a representative stress scenario rather than a best - case setting. Evaluation is framed fro m a decision - support perspective. Emergency managers operate on hourl y update cycles and must issue evacuation advisories, activate shelters, and allocate resources within short lead times. Accordingly , the analysis emphasizes (i) the reliability of situa tional awareness under sparse crowdsourced evidence, (ii) the ability of rolling forecasts to correct rainfall - driven bias through impact - state updating, and (iii) forecast stability within short horizons (1 – 3 h) that align with practical decision windows. Results are interpreted in terms of operational behavior and s ystem r obustness rather than location - specific performance . 6.2 Rainfall Forecast Uncertainty and Impact Forecasting Challenges Accurate impact forecasting during urban flood events is fundamentally constrained by the quality of rainfall forcing. Fig.1 1(a) compares observed hourly rainfall during T yphoon Haikui with the operational numerical weather p rediction u sed for forecasting. The forecast substantially underestimates both rainfall intensity and cumulative accumulation, producing 162.6 mm compared to the obs erved 447.5 mm over the event period. When forecasts undere stimate rainfall, hydrologic and hydraulic models propagate this bias downstream, leading to systematic underestimation of flood extent and associated functionality loss. Consequently , rainfall - driven impact forecasts alone may f ail to capture both the timing and magnitude of escalating disruptions during the most critical respons e window . Rather than attempting to correct meteorol ogical inputs — which is rarely feasible within operational time constraints — CRAF pe rforms co rrection directly in th e impact space by assimilating real - time human - sensed observations ( Fig. 1 1(b) ) . This strategy shifts uncertainty management from upstream hazard estimates to downstream consequences, where decisi ons are ultimately made. The following analyses demonstrate how this impact - space updating stabilizes situational awareness and improves forecast reliability under severe rainfall forcing errors. Preprint. Under Review . Xie et al. | arXiv | 18 / 24 Fig. 11 Rainfall and waterlogging scenarios: ( a) te mporal comparison of f orecasted and observed hourly rainfall and crowd activity; (b) representative waterlogging observati ons at different time steps. 6.3 Situational A wareness Enhancement Under Sparse Crowd Observations Accurate s hort - horizon impact forecasting hinges on the quality of the curr ent impact - state estimate. During Ty p h o o n H a i k u i , t h i s r e q u i r e m e n t i s e s p e c i a l l y a c u t e : rainfall forcing is severely underpredicted, while crowdsourced observations — though inform ative — are sparse, uneven, and temporally clustered. As shown in Fig. 12(a) , ERZ - level observation coverage during the first major crowdsourcing peak (22:00 on September 5 to 02:00 on September 6) sp ans only 16% – 42%, leaving large portions of the urban system unobserved at precisely the moment when impacts are escalating. If initialized directly from biased rainfall forecasts and r aw sparse observations, any downstream forecasting — regardless of model sophistication — would b e struct urally mis aligned with real ity . The SA module addresses this failure mode by expli citly perform ing current impact - state alignment prior to forecasting. Leveraging a physics - trained spatial prior , SA propagates information from obs erved ERZs (shaded) to unobserved zones (unshaded), reconstructing a spatially complete and physically coherent ZFL field that reflects system - level impact conditions rather than isolated reports. As illustrated in Fig. 12(b) , inferred ZFL patterns evolve smoothly across space, capturing coherent escalation dynamics instead of fragmented observational artifa cts. Importantly , SA does not merely “fill gaps” in observations. By encoding inter - ERZ impact co - variation learned from physics - based simulations, it constrains inference to physically plausible spatial dependencies, preventing underestimation under s parse coverage. In doing so, SA transforms heterogeneous crowdsourced evidence into a decision - grade impact state that reconciles human sensing with physical process constraints. From an operational standpoint, this alignment step is Preprint. Under Review . Xie et al. | arXiv | 19 / 24 foundational. W ithout spatial completion, downstream forecasts would inherit compounded bias from rainfall underestimation and incomplet e observati ons, rapi dly eroding decision confidence. By reconstructing a calibrated, system - consistent impact state prio r to propagation, SA provides a reliable initialization that anchors subsequent forecasts to evolving ground truth. This capabil ity transforms sparse and heterogeneous crowd reports into actionable situational awareness and is essential for dependable real - time impact forecasting under precisely t he conditions where rainfall - driven approaches are most vulnerable. Fig. 12 Spatiotemporal distribution of ZFL across four time slices. 6.4 Rolling Forecast Updating and Error Reduction During Critical Decision W indows Building on the calibrated impact states produced by the SA module, rolling multi - step forecasts are evaluated under four clearly defined information regimes, with the first two serving as baseline scenarios: 1. STF - OL - FR (Open - Loop, Fixed Rainf all - driven Forecasting): represents a one - time forward propagation of the spatiotemporal forecast over the entire horizon using fixed rainfall forcing, without any rolling updat es or assimilation of impact observations; 2. STF - OL - UR (Open - Loop, Updated Rainfall - driven Forecasting): at each cycl e the spatiotemporal forecasting model is driven by updated rainfall (from numerical weather predictions and ground meteorological station observations), but wit hout assimilating impact observations; 3. CRAF (Closed - Loop CIM+SA+STF Forecasting): STF is reinitialized at each update cycle using SA - completed impact states that assimilate crowdsourced observations , in combination with updated rainfall forcing ; and 4. Obse rved ZFL : the reference impact state derived from aggregated crowd evidence, serving exclusively as the evaluation benchmark. Figs. 13 compare rolling forecasts produced by STF - OL - FR and CRAF across successive update cycles during T yphoon Haikui. Across successive update cycles, open - loop forecasts ( STF - OL - FR ) Preprint. Under Review . Xie et al. | arXiv | 20 / 24 systematically underestimate both the magnitude and timing of functionality loss due to rainfall forcing bias, producing trajectories that ar e internally consistent yet operationally misleading. In contrast, closed - loop CRAF forecasts remain closely couple d to the observed impact evolution. By assimilating crowdsourced evidence before each rollout, CRAF repeatedly realigns the system state with current conditions, preventing error accumulation and maintaining forecas t credibility . This behavior y ields consi stent 1 – 3 h - ahead error reductions of 84.4 – 95.1% relative to the fixed - forcing baseline, demonstrating stable closed - loop dynamics rather than isolated corrections. Fig. 13 Comparison of d ynamic fore casting performance between STF - OL - FR baseline (fixed rainfall - driven one - time forecasting without crowd assimilation) and CR AF models across multiple initialization time step s. Preprint. Under Review . Xie et al. | arXiv | 21 / 24 Fig. 14 present s prediction err or (MAE ) comparisons between STF - OL - UR and CRAF . When rainfall inputs are updated but impact observations are not assimilat ed with STF - OL - UR, substantial errors persist. Even unde r identical rainfall forcing, CRAF reduces short - horizon MAE by 72.8 – 79.6%, and the error distribution becomes markedly more concentrated. This contraction reflects not only improved accuracy but also enhance d stability and predictability — properties that directly translate to greater operational trust. T ogether , these results show that reliability gains arise primarily from dynam ic state alignment rather than rainfall refinement alone. Given tha t time - sensitive emergency actions (e.g., evacuation advisories, shel ter activation, and resource deployment) are typically initiated within narrow 1 – 3 h decision windows, this dual merit of CRAF carries decisive practical weight. In stark cont ras t, o pen - loop rainfall - driven forecasts can deviate rapidly under forcing uncertainty , undermining trust and operational usefulness. By closing the loop between s ensing and forecasting, CRAF continuously realigns predictions with evolving ground truth at low late ncy . This ability to stabilize forecasts under uncertainty — rather than merely propagate ha zards — defines the practical value of closed - loop impact forecasting and distinguishes CRAF from conventional op en - loop approaches. During the Haikui even t, CRAF co mpleted a full update cycle — including crowd ingestion, situational inference, and multi - step forecasting — in approximately 10 minutes per update on a standard GPU. This latency comfortably supports hourly or sub - hourly rolling updates required in operational emergency response. To ensure reliability, crowdsourced data assimilation is restricted to the event escalation phase, as post - peak reports m ay underestimate residual functional ity loss. Together, these design choices refl ect practical deployment constraints and ensures timely yet stable real - time operation. Fig. 14 Prediction error comparisons between STF - OL - UR and CRAF . Preprint. Under Review . Xie et al. | arXiv | 22 / 24 7 DISCUSSION AND CONCLUSIONS This study introduced CRAF , a physics - informed, closed - loop impact forecasting framework designed to operate under two persistent challenges in real - world flood response: uncertain hazard forcing and sparse, heterogeneous observati ons. Through controlled o ffl i n e verification and an operational deployment, the results demonstrate that rel iable short - horizon impact forecasting depends less on refining hazard inputs than on continuously aligning the evolving impact state with real - world evidence. This finding reframes impact forecasting as a state estimation problem rather than solely a hazard prediction problem . A c e n t r a l i n s i g h t i s t h a t i m p a c t - state initialization dominates forecast s kill under forcing uncertainty . When ra infall f orecasts are bia sed, o pen - loop impact propagation rapidly di verges from reality , even if internally consistent. CRAF mitigates this failure mode by assimilating sparse crowdsourced observations through a physics - trained situational prior, producing spatially coherent impact state s before forecasting. This alignment step is foundational rather t han auxiliary: without it, downstream forecas ts remain structurally misaligned rega rdless of mod el sophisticat ion. By explicitly closing t he loop between sensing, situational inference, and forecasting, CRAF departs from conventional open - loop hazard - driven approaches and delivers stable gains within decision - relevant horizons. In the demonstrated event, CRAF reduced 1 – 3 hour - ahead impact forecast errors by approxi mately 84.4% - 95.1% relative to the fixed rainfall - driven baseline and 72.8% – 79.6% relative to the updated rainfall - driven baseline, while maintaining 10 - minute update latency suitable for real - time operation. The r obustness of th e framewo rk stems from a deliberate separation between offline physics - supervised learning and online inference and updating. Physics - based simulations are used of fline to learn transferable spatial and temporal dependencies, while real - time operation is restricted to state updating and forward propagation. This design avoids instability associated with online retraining while enabling rapid, reliable updates und er operational constrain ts. By prioritizing situational alignment over hazard perfection, CRAF offers a robust foundation for impact - based early warning systems and operational digital twins that must remain coupled to the real world unde r uncertainty . Several limitations warrant acknowledgment. The effectiveness of situational correction depends on the availability of timely observational evidence during event es calation, and the physics - based simulations used for offline training reflect region - and sy stem - specific modeling assumpt ions. Moreover, actual impact s inferred from CIM and SA remain reliable and independent of rainfall forecasts, but STF’ s forecasted impact s are influenced by rainfall inputs, wi th large deviations po tentially accumulating errors in longer - term rolling forecasts. Future work will ext end the framework across additional hazards, regions, and sensing modalities, explore adaptive weighting of observational uncer tainty , and inves tigate strategi es to mitigate rain fall - dependent forecast errors, such as adaptive parameter updating or inverse es timation of rainfall forcing. ACKNOWLEDGEMENT This project is supported by Zhejiang Department of Science & T echnology through Grant No. 2024C03255 . REFERENCE S Annis, A., and F . Nardi. 2019. “Integra ting VGI and 2D hydraulic models into a data assimilation framework for real time flood forecasting and mapping.” Geo - spatial Information Science, 22 (4): 223 – 236. T aylor & Francis. https://doi.org/10.1080/10095020.2019.1626135. Assumpção, T . H., I. Popesc u, A. Jonoski, and D. Solomatine. 2021. “Citizen observations contributi ng to flood modelling: opportuni ties and c hallenges. ” J Flood Risk Manag , 22 (2): 1473 – 1489. https://doi.org/10.5194/hess - 22 - 1473 - 2018. Chaudhary , P ., S. D’Aron co, J. P . Leitão, K. Schindler, and J. D. W egner . 2020. “W ater level predicti on fr om soci al media ima ges with a multi - task ranking approach.” ISPRS Journal of Photogrammetry and Remote Sensing , 167: 252 – 262. https://doi.org/10.1016/j.isprsjprs.2020.07.003. Chavez, J. - P. , J . M u r c i a - Delso, and F . López - Almansa. 2025. “Probabilistic framework for post - earthquake functionality assessment of seismically isolated hospital buildings using Bayesian networks.” Engineering Structures , 328: 1 19633. https://doi.org/10.1016/j.engstruct.2025.119633. Farahmand, H., Y . Xu, and A. Mostafavi. 2023. “A spatial – Preprint. Under Review . Xie et al. | arXiv | 23 / 24 temporal graph deep learning model for urban flood nowcasting leveraging heterogeneous community features.” Sci Rep , 13 (1): 6768. https://doi.org/10.1038/s4 1598 - 023 - 32548 - x. Fohringer , J., D. Dransch, H. Kreibich, and K. Schröter . 2015. “Social media as an information source for rapid flood inundation mapping.” Nat. Haz ar ds Earth Syst. Sci. , 15 (12): 2725 – 2738. https://doi.org/10.5194/nhess - 15 - 2725 - 2015. Guan, M., K. Guo, H. Y an, and N. W ri ght. 2023. “Bottom - up multileve l flood hazard mapping by integr ated in undation modelling in data sca rce citi es.” Journal of Hydr ology , 617: 1291 14. ht tps://doi.org/10.1016/j.jhydrol.2023.1291 14. Guidotti, R., Gardoni, P ., and Rosenheim, N. , (2019) "Integration of physical infrastructure and social systems in communities" reliability and resilience analysis" Reliability Engineering & Sys tem Safety , 185, 4 76 - 492. Hao, X., H. L yu, Z. W an g, S. Fu, and C. Zhang. 2022 . “Estimating the spatial - temporal distribution of urban street ponding levels from surveillance videos based on computer vision. ” Wa t e r R e s o u r M a n a g e , 36 (6): 1799 – 1812. https://doi.org/10.1007/s1 1269 - 022 - 03107 - 2. Kaur K, Kaur P . 2023. “BER T - CNN: Improvin g BER T for Requirements Classificati on usi ng CNN”. Proc edia Computer Science , 218: 2604 - 2611. Lamadrid, A. J., M. Escal eras, D. Mits ova, A. - M. Esnard, and A. Sapat . 2025. “Househo ld evacu ation de cisions and relationship to infrastructure disruption using evidence from Hurricane Irma.” Environment Systems and Decisions , 45 (3): 32. https://doi .org/10.1007/s10669 - 025 - 10019 - 0. Lee, C., and I. T ien. 2018. “Pr obabilistic Framework for Integrating Multiple Data Sources to Estim ate D isaster and Failure Events and Increase Situational A wareness.” ASCE - ASME J. Risk Uncer tainty Eng. Syst., Part A: Civ . Eng. , 4 (4): 04018042. American Society of Civil Engineers. https://doi.org/10.1061/AJRUA6.0000995. Lin, L., C. T an g, Q. Liang, Z. W u, X. W ang, an d S. Zhao. 2023. “Rapid urban flood risk mapping for data - scarce environments using soci al sens ing and region - stable deep neural network.” Journal of Hydrology , 617: 128758. https://doi.org/10.1016/j.jhydrol.2022.128758. Lin, P ., and N. W ang. 2017. “Stochast ic post - disaster functionality recovery of community building portfolios I: Modeling .” Structural Safety , 69: 96 – 105. https://doi.org/10.1016/j.strusafe.2017.05.002. Mazzoleni , M. , M. V erlaan, L. Alfonso, M. Mo nego, D. Norbiato, M. Ferr i, and D. P . Sol omatine. 2017. “Ca n assimilation of crowdsourced data in hydrological modelling improve f lood pred iction?” Hydr o l. Earth Syst. Sci. , 21 ( 2): 839 – 861. Coperni cus GmbH. https://doi.org/10.5194/hess - 21 - 839 - 2017. Meyer , H., and G. Joh ann. 2 025. “ Insight s From Homeowners on the Impact of Flood Risk Communication on Adaptive Behavior at the Property Level From the 2021 Flood Event in Germany .” J Flood Risk Management , 18 (1): e70038. https://doi.org/10.1 11 1/jfr3.70038. Nocera, F . , and Gardoni, P ., (2019). "A Ground - up appr oach to estimate the likelihood of business interruption," International Journal of Disaster Risk Reduction, 41, 101314 Panakkal, P ., and J. E. Padgett. 2024. “More eyes on the road: Sensing flooded roads by fusing real - time ob servations from public data sources.” Reliabil ity Enginee ring & System Safety , 251: 1 10368. https://doi.org/10.1016/j.ress.2024.110368. Restrepo - Estrada, C., S. C. de Andrade, N. Abe, M. C. Fava, E. M. Mendiondo, and J. P . de Albuquerque. 2018. “Geo - social media as a proxy for hydrometeorological data for streamflow estimation and to improve flood monitoring.” Computers & Geosciences , 1 11: 148 – 158. https://doi.org/10.1016/j.cageo.2017.10.010. Rossi, C. , F . S. Acerbo, K. Ylinen, I. Juga, P . Nur mi, A. Bosca , F. T a r a s c o n i , M . C r i s t o f o r e t t i , a n d A . A l i k a d i c . 2 0 1 8 . “Early detection and infor mation extraction for weather - induced floods using social media streams.” International Journal of Disaster Risk Reduction , 30: 145 – 157. https://doi.org/10.1016/j.ijdrr .2018.03.002. Safaei - Moghadam, A. , A. Hoss einzadeh, and B . Minske r . 2024. “Predicting real - time roadway pluvial flood risk: A hybrid machine l earning approach coupling a graph - based flood spreading model, historical vulnerabilities, and W aze data.” Journal of Hydr ology , 637: 131406. https://doi.org/10.1016/j.jhydrol.2024.131406. Songchon, C., G. Wright, and L. Beevers. 2023. “The use of crowdsourced social media data to improve flood forecasting.” Journal of Hydr ology , 622: 129703. https://doi.org/10.1016/j.jhydrol.2023.129703. Su, J., J. W . Pedersen, I. M. Ringgaard, and M. A. D. Larsen. 2025. “Incorporating compound flood events in the weather service v alue chain.” npj Nat. Hazar ds , 2 (1): 26. https://doi.org/10.1038/s44304 - 025 - 00077 - 4. Sun, M., and E. J. Cha. 2022. “Probabilistic Building Functionality Estimation Framework for Buildings Subjected to Hurricane W inds.” J. Struct. Eng. , 148 (10): 04022147. https://doi.org/10.1061/(ASCE)ST .1943 - 541X.000341 1. Ta v a k o l i f a r , H . , H . A b b a s i z a d e h , S . N a z i f , a n d E . Shahghasemi. 2021. “Development of 1D – 2D Urban Flood Simulation Model Based on Modified Cell ular Automata Approach.” J. Hydr ol. Eng. , 26 (2): 04020065. Preprint. Under Review . Xie et al. | arXiv | 24 / 24 https://doi.org/10.1061/(ASCE)HE.1943 - 5584.0002036. Wa n g , R . - Q., H. Mao, Y . W a ng, C. Rae, and W . Shaw . 2018a. “Hyper - reso lution monitoring of urban flooding with social media and crowdsourc ing data.” Comput Geosci - uk , 11 1: 139 – 147. https://doi.org/10.1016/j.cageo.2017.1 1.008. Wa n g , Y . , A . S . C h e n , G . F u , S . D j o r d j e v i ć , C . Z h a n g , a n d D . A. Savić. 2018b. “An integrat ed fr amework for high - resolution urban flood modelling considering multiple information sources and urban features.” Envir onment al Modelling & Software , 107: 85 – 95. https://doi.org/10.1016/j.envsoft.2018.06.010. Xie, L., P . Lin, N. W ang, Y . Xu, and B. R. Ellingwood. 2025a. “Real - time probabilistic spatiotemporal forecasting of building funct ionality in flood emergencies: a deep learning approach to facilitate community evacuation planning.” Structure and Infrastru cture Engineering, 1 – 21. Informa UK Limited. Xie, L., N. W ang, P . Lin, a nd H. Mahmoud . 20 25b. “Spatiotemporal Forecasting of Functionality States of Community Building Po rtfolios under Flood Evolution.” J. Struct. Eng. , 151 (2): 04024218. American Soc iety of Civil Engineers. https:/ /doi.org/10.1061/JSENDH.STENG - 13672. Ya n , Z . , X . G u o , Z . Z h a o , a n d L . T a n g . 2 0 2 4 . “ A c h i e v i n g f i n e - grained urban flood percept ion and spatio - temporal evolution analysis based on s ocial media.” Sustainable Cities and Society , 101: 105077. https://doi.org/10.1016/j.scs.2023.105077. Ya n g , X . , S . L i u , C . J i a , Y. L i u , a n d C . Yu . 2 0 2 1 . “V ulnerability assessment and management planning for the ecological environment in urban wetlands.” Journal of Envir onmental Management , 298: 1 13540. https://doi.org/10.1016/j.jenvman.2021.1 13540. Yu , B . , H . Y i n , a n d Z . Z h u . 2 0 1 8 . “ S p a t i o - Te m p o r a l G r a p h Convolutional Ne tworks: A Deep Le arning Framework f or Tr a f f i c F o r e c a s t i n g . ” a r X i v. Yu a n , F . , C . - C. Lee, W . Mobley , H. Farahmand, Y . Xu, R. Blessing, S. Dong , A. Mostaf avi, and S. D. Brod y . 2023. “Predicting road flooding risk with crowdsourced reports and fine - grained traffic data.” Computational Urban Science , 3 (1): 15. https://doi.or g/10.1007/s437 62 - 023 - 00082 - 1. Zander , K. K., D. Nguyen, M. Mir babaie, and S. T . Garnett. 2023. “A ware but not prepared: understanding situat ional awareness during the century flood in Germany in 2021.” International Journal of Disaster Risk Reduction , 96: 103936. https://doi.org/10.1016/j.ijdrr .2023.103936. Zeng, Z. , Z. W ang, and C. Lai . 2022. “Simulat ion Performance Evaluation and Uncertai nty Analysis on a Coupled Inundation Model Combini ng SWMM and WCA2D.” Int J Disaster Risk Sci , 13 (3): 448 – 464. https://doi.org/10.1007/s13753 - 022 - 00416 - 3. Zhou, S., W . Jia, M. W ang, Z. Liu, Y . W ang, and Z. W u. 2024. “Synergistic assessment of multi - scenario urban waterlogging through data - driven decoupli ng anal ysis in high - density urban areas: A case s tudy in Shenzhen, China.” Journal of Envir onmental Management , 369: 122330. https://doi.org/10.1016/j.jenvman.2024.122330. Zhu, H., P . Obeng Oforiwaa, and G. Su. 2022. “Real - time urban rainstorm and waterlogging disaster detection by We i b o u s e r s . ” Nat. Hazards Earth Syst. Sci. , 22 (10): 3349 – 3359. https://doi.org/10.5194/nhess - 22 - 3349 - 2022.