AIFloodSense: AI Flood-Risk Assistant

• AIFloodSense is a conversational AI assistant that integrates real-time hazard data, FEMA flood maps, and CDC vulnerability indices to democratize flood risk communication.
• It leverages GPT‑4’s function-calling to orchestrate diverse data streams from agencies like NWS and CDC, merging technical outputs into understandable narratives.
• The system translates complex flood-risk information into actionable guidance, aiding decision-makers and the public in emergency preparedness and spatial planning.

Searching arXiv for the named system and related papers to ground the article in current literature. arXiv search query: AIFloodSense OR "Towards Democratized Flood Risk Management" OR (Martelo et al., 2024) AIFloodSense is a prototype AI flood-risk assistant centered on GPT‑4 that links real-time hazard data, FEMA flood maps, and CDC social vulnerability information to produce natural-language guidance for decision-makers and the public (Martelo et al., 2024). In the source paper, it is presented not as a hydrologic or hydrodynamic solver, but as a conversational orchestration and interpretation layer above existing GIS, alerting, and vulnerability datasets. The system’s defining objective is the democratization of flood-risk communication: it translates flood zone codes, warning products, and social vulnerability indicators into actionable narratives without requiring specialized technical knowledge (Martelo et al., 2024). The name “AIFloodSense” is also used in later literature for distinct artifacts, including a global aerial imagery dataset for flood understanding (Simantiris et al., 19 Dec 2025); this suggests that the label has broadened beyond the original assistant into a wider family of AI-enabled flood-information systems.

1. Definition and conceptual scope

In its original formulation, AIFloodSense is a chat-based AI assistant built around GPT‑4’s function-calling capability and designed to connect real-time hazard data, flood maps, and social vulnerability information for interactive flood-risk communication (Martelo et al., 2024). Its operational role is interpretive rather than predictive in the numerical sense: no internal hydraulic or hydrologic model runs inside GPT‑4, and the assistant instead sits on top of existing FEMA, CDC, and National Weather Service information. This architectural choice places AIFloodSense within a class of AI systems that mediate between authoritative technical products and end users, rather than replacing the upstream scientific infrastructure.

The paper frames the core problem as a communication gap. Decision-makers often depend on domain experts to interpret flood-risk products, while members of the public struggle to understand flood warnings, floodplain labels, and their implications for insurance, construction, and evacuation (Martelo et al., 2024). AIFloodSense addresses this by using a LLM as a semantic router, explainer, and advisor. This suggests a reconfiguration of flood-risk management in which model interpretability and public engagement are treated as primary system objectives rather than secondary dissemination steps.

A broader research context sharpens this positioning. Other flood-AI systems target forecasting, inundation mapping, damage assessment, social sensing, or imagery interpretation rather than conversational explanation. Examples include global river flood forecasting with encoder–decoder LSTMs (Nearing et al., 2023), remote-sensing nowcasting via multi-sensor ensembles (Nazir et al., 29 Apr 2026), social-media flood relevance and topic fusion (Gong et al., 10 May 2025), city-scale mobile sensing and task scheduling (Qian et al., 2020), flood scene segmentation (R et al., 2022), and building-level flood damage assessment from SAR and optical imagery (Ho et al., 7 Jun 2025). AIFloodSense, in contrast, primarily addresses the last-mile interpretability problem.

2. System architecture and orchestration model

The original architecture consists of a front-end chat client, GPT‑4 (gpt-4-1106-preview) acting as controller and explainer, a function layer exposing four flood-specific tools, and visualization services that generate static and interactive maps (Martelo et al., 2024). Users submit natural-language questions; the client forwards the dialogue together with a schema describing available tools to the GPT‑4 API. GPT‑4 then decides whether to answer directly from pre-training or invoke one or more tools when fresh, location-specific data are required.

The function layer exposes exactly four specialized functions: get_flood_map, get_flood_data, get_svi_stats_and_tracts, and get_flash_flood_warnings (Martelo et al., 2024). These wrap calls to GIS, FEMA, CDC SVI, and NWS APIs and return structured JSON containing coordinates, flood-zone labels, SVI scores, timestamps, and related fields. GPT‑4 then integrates these outputs into a final response that includes explanations, suggestions, and mitigation guidance. In the paper’s terms, GPT‑4 performs three roles: semantic router or orchestrator, interpreter of technical outputs, and communicator or advisor.

This controller design is significant because it operationalizes function-calling as a domain-specific middleware rather than as a generic chatbot enhancement. The system is not simply retrieving documents; it is selecting tools, resolving ambiguities through follow-up questions, synthesizing heterogeneous outputs, and presenting a coherent explanation. A plausible implication is that AIFloodSense should be understood as a conversational GIS-and-warning interface whose novelty lies in workflow integration and interpretive compression.

The architecture also supports map generation. Static maps, when retrieval succeeds, use color-coded FEMA flood zones, and interactive web maps are implemented through Leaflet and OpenStreetMap overlays (Martelo et al., 2024). GPT‑4 references these maps through links and textual descriptions. This makes the assistant not only a textual explainer but also a guide to map-based evidence.

3. Data sources and information fusion

AIFloodSense integrates four principal data streams: real-time flood alerts, hydrological and meteorological information embedded in agency products, flood zone and property-level map data, and Social Vulnerability Index data (Martelo et al., 2024). The real-time warning stream comes from the U.S. National Weather Service API through get_flash_flood_warnings(location), returning event type, description, timestamps, and area descriptions. GPT‑4 converts these fields into concise local summaries.

Flood zone and property-level information are accessed through get_flood_map(latitude, longitude, zoom) and get_flood_data(address) (Martelo et al., 2024). The former generates static flood maps and interactive web maps with zones such as A, AE, V, X, B, C, and D. The latter returns FEMA FIRM panel metadata, flood-zone classification, SFHA_TF, zone subtype, political or community fields, and geocoded coordinates. GPT‑4 then explains the implications for risk, insurance, and floodplain regulation.

Social vulnerability is drawn from the CDC/ATSDR 2020 SVI dataset at census-tract level via get_svi_stats_and_tracts(state_abbr, county, theme, op, threshold) (Martelo et al., 2024). The function supports themes RPL_THEME1, RPL_THEME2, RPL_THEME3, RPL_THEME4, and RPL_THEMES, comparison operators, and thresholds on the interval $0$–$1$. It returns tract counts, max/min/average SVI, tract identifiers, scores, polygon boundaries, and an interactive map. The paper reports examples including New Orleans tracts with SVI > 0.9, Galveston Island tracts with SVI > 0.85, and low SVI values in Lake County, Colorado.

The paper explicitly notes that the prototype does not directly ingest raw rainfall, discharge, or weather-model grids; these inputs are only implicitly embedded through NWS alert products and pre-existing FEMA studies (Martelo et al., 2024). It nevertheless outlines canonical hydrologic components such as rainfall–runoff transformation via a unit hydrograph,

$Q(t) = \int_{0}^{t} P(\tau)\, u(t-\tau)\, d\tau,$

channel routing, and stage–discharge relationships. These equations are presented as computational background rather than implemented modules. AIFloodSense therefore occupies an interface layer over precomputed agency science.

The fusion step occurs inside GPT‑4. It merges flood zones, current alerts, historical impact narratives available through model pre-training, and SVI profiles into user-level explanations (Martelo et al., 2024). This multimodal synthesis is central to the system’s practical utility. A plausible implication is that the assistant’s value depends less on any single data source than on the semantic alignment of heterogeneous information products into one localized narrative.

4. Interpretive logic, risk communication, and user interaction

AIFloodSense translates technical flood terminology into plain-language guidance while retaining regulatory and probabilistic meaning (Martelo et al., 2024). For example, the assistant explains that Zone AE corresponds to a high-risk 1% annual chance floodplain and may imply elevation requirements relative to Base Flood Elevation. It also clarifies return-period misconceptions by reframing the “100-year flood” as a 1% annual exceedance probability and notes that over a 30-year mortgage the probability of at least one such event is approximately $26\%$: $P = 1 - (1-0.01)^{30} \approx 0.2603.$ This interpretive step is explicitly positioned as a correction to common misunderstanding.

The assistant also translates water-level and warning terminology. The paper gives examples in which GPT‑4 interprets stage relative to flood stage, such as a river level of 100.8 ft being above a flood stage of 100 ft while still classified as minor flooding (Martelo et al., 2024). It similarly renders SVI percentiles in national comparative terms, such as describing a tract as being in the top 10% nationally for vulnerability.

Actionability is a central feature. In the Baton Rouge construction case, when the site is identified as Zone AE, GPT‑4 recommends elevating the building above Base Flood Elevation, using flood-resistant foundations and materials, raising utilities and electrical systems, implementing green infrastructure and stormwater controls, ensuring NFIP-compliant insurance and code adherence, and preparing emergency response and evacuation plans (Martelo et al., 2024). These recommendations are not generic flood advice detached from data; they are tied to the property-level attributes returned by the tools.

The interaction model supports both top-down and bottom-up exploration. The paper distinguishes a top-down flow from regional alerts to neighborhoods to specific tracts, and a bottom-up flow from a single address to general risk and mitigation (Martelo et al., 2024). This bidirectionality matters because planners, emergency managers, and residents approach flood-risk information with different entry points. AIFloodSense is designed to accommodate both expert workflows and personal safety queries within one conversational framework.

This interaction paradigm aligns with a wider movement in flood informatics toward broader public participation. Related work has used social media to infer flood relevance and geographic patterns (Gong et al., 10 May 2025, Arthur et al., 2017), vision systems to infer street-level flood depth from public imagery (Fuad et al., 17 Mar 2026), and aerial or drone imagery for segmentation and post-disaster interpretation (R et al., 2022, Mou et al., 14 Jan 2025, Simantiris et al., 19 Dec 2025). AIFloodSense differs in that it does not primarily sense floods from raw media; it interprets institutional data products for public-facing use.

5. Evaluation, performance, and observed failure modes

The original evaluation uses five prompt categories with six prompts each: get_flood_map, get_flood_data, get_svi_stats_and_tracts, get_flash_flood_warnings, and general flood-related conceptual questions (Martelo et al., 2024). Two iterations were conducted, with the second used as the main evaluation after prompt refinement. Scoring employed a 1–5 scale across three dimensions: relevance, error resilience, and context understanding, each subdivided into two criteria.

The results show a clear pattern. General flood understanding achieved 5.0/5 across all criteria, indicating that GPT‑4 excelled in conceptual explanations (Martelo et al., 2024). Flash flood alerts scored roughly 4.3/5, with strong performance in retrieving NWS alerts and summarizing them. Flood-data queries achieved about 4.1–4.2/5, while flood-map usage scored around 4.0/5. The SVI function was also around 4.0/5 overall and was the most problematic specialized tool because of token-limit issues and invalid queries when prompts lacked required county information.

Some refined prompts reached perfect scores of 5/5 for SVI-related use cases, such as Galveston Island, Humphreys County, and Skagway (Martelo et al., 2024). This matters because it indicates that the weaker average performance was not due to an intrinsic inability to interpret vulnerability data, but to tool and prompt design sensitivity. The paper therefore emphasizes prompt refinement and constrained function design as engineering necessities rather than optional polish.

The response-time analysis reveals a pronounced difference between direct conceptual answers and tool-mediated responses. Function calls averaged about 36 s, whereas responses without function calls averaged about 12 s (Martelo et al., 2024). Flash flood alerts were the fastest function-using category at about 5–25 s, while SVI queries were the slowest, with an average around 85 s and extreme cases up to 198 s. These latencies were attributed to large-dataset filtering and token constraints.

Several failure modes are documented. GPT‑4 sometimes hallucinated references to interactive maps not actually returned by get_flood_data, and after static map retrieval failures it sometimes under-explained flood map symbology instead of turning the failure into an explanatory moment (Martelo et al., 2024). Long SVI outputs could cause truncation or loss of earlier conversational context because of the 4096-token output limit. These issues indicate that the limiting factor in such systems is not only model knowledge but also tool contract design, output compression, and error-aware prompting.

6. Implementation, extensibility, and related research trajectories

Deployment requires a GPT‑4 endpoint with function-calling support, access to FEMA flood map and FIRM services or national equivalents, a weather or hydrological warning source such as the National Weather Service, the CDC/ATSDR SVI dataset or analogous regional vulnerability indices, GIS infrastructure for map generation, and a web or mobile chat interface with links to interactive maps (Martelo et al., 2024). The backend must implement and secure the four functions, handle GIS operations and caching, and pre-filter large datasets before they are passed to the LLM.

The paper recommends several best practices: narrowing function outputs, enforcing clear parameter requirements such as mandatory state and county for SVI queries, introducing paging or aggregation to avoid token overload, and using explicit system instructions to prevent hallucinated map availability (Martelo et al., 2024). It also proposes regional adaptation by replacing FEMA, NWS, and CDC resources with national or regional analogues and adjusting geo-reasoning to local coordinate systems and administrative units.

Future enhancements mentioned in the paper include integrating GPT‑4 vision capabilities, connecting to real-time gauge data and forecasting systems, and introducing conversation memory across sessions (Martelo et al., 2024). These developments would move AIFloodSense closer to richer multimodal flood-intelligence systems. This suggests a convergence with adjacent research lines: continuous remote-sensing nowcasting across multiple sensors (Nazir et al., 29 Apr 2026), building-level flood damage assessment using SAR, optical, and risk priors (Ho et al., 7 Jun 2025), and crowd-sourced street-level depth estimation from social media imagery (Fuad et al., 17 Mar 2026).

The name “AIFloodSense” itself has already diversified in later publications. One 2025 paper introduces “AIFloodSense” as a global aerial imagery dataset with 470 images from 230 flood events across 64 countries and supports classification, semantic segmentation, and visual question answering (Simantiris et al., 19 Dec 2025). That usage differs materially from the GPT‑4 assistant of 2024. A plausible implication is that the term has evolved from denoting a single assistant architecture into a broader label for AI-driven flood-information infrastructures spanning conversational support, perception datasets, and decision-support tooling.

Taken together, AIFloodSense occupies an identifiable place in the flood-AI landscape. Its original contribution lies in making agency-produced flood data interpretable, conversational, and socially contextualized (Martelo et al., 2024). It does not replace flood models; it renders their outputs operationally legible. In that respect, its significance lies less in numerical innovation than in the construction of a translation layer between technical flood-risk systems and human decision-making.