Source-linked AI summary
A Topology-Guided GeoAI Agentic Framework for Urban Flood Screening
Yutian Tao, Haoan Feng, Xin Xu, Leila De Floriani
TL;DR
Urban flood screening can demand specialist modeling and data preparation, while threshold choices vary by study area. This framework combines deterministic geospatial analysis with recorded language-model decisions and human checkpoints; across five study areas, it reports agreement between Morse terrain structures and D8 routing.
Problem
Terrain-based screening can require specialist hydrological modeling, licensed software, and data preparation beyond the capacity of smaller organizations, and persistence thresholds vary by study area.
Source: §1 · p. 1Method
The framework uses deterministic public-data analysis to extract and rank Morse basins, while language-model decisions are restricted to recorded choice points and user review.
Source: §1 · pp. 1–2, §2.2 · p. 2, §2.3 · p. 2Results
Across five study areas, Morse-based terrain structures showed consistent agreement with independent D8 routing, while sensitivity analysis and self-review helped identify unstable rankings and unsupported statements.
Source: §4 · p. 4Takeaways & Limitations
The framework provides a traceable, low-cost first-pass screening with inspectable recommendations and user review of consequential decisions.
Source: §4 · p. 4The evaluation covers few study areas, uses U.S.-specific data services and simplified exposure, and lacks observed flood events as ground truth.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Open-world geographic artificial intelligence (GeoAI) agents can transform public geospatial data into accessible decision support for real-world challenges, but must balance autonomous reasoning with reliable spatial analysis and human oversight. We present a topology-guided GeoAI agentic framework for urban flood screening that takes a place name, drawn region, or uploaded boundary and returns ranked candidate basins with an interactive map and report. A deterministic controller retrieves public elevation, rainfall, soil, land-cover, exposure, and infrastructure layers, extracts terrain depressions with discrete Morse theory, and ranks basins with a relative flood-priority score. The language model acts only at recorded decision points: selecting a persistence threshold from a precomputed candidate table, drafting recommendations from computed evidence, and reviewing each recommendation against the same evidence. Three human-in-the-loop checkpoints let users revise the threshold, factor weights, or recommendation. In a College Park, Maryland case study, the selected threshold reduced 278 raw terrain minima to 70 basins, with 77% of these basins' lowest points also identified as sinks by a separately computed eight-direction (D8) flow-routing model. Across four further study areas, agreement ranged from 61% to 83%. By combining open data, explicit spatial reasoning, traceable evidence, and human oversight, the framework enables transparent first-pass flood screening while keeping consequential recommendations inspectable and user-controlled.
1 Introduction
Urban flood screening can require specialist modeling, software, and data preparation that smaller organizations may not be able to support. The framework addresses threshold selection and hydrological relevance with deterministic spatial analysis, recorded model decisions, and user oversight.
1 Introduction
- Terrain screening requires a per-area persistence threshold because terrain relief, DEM resolution, and noise vary, and that choice determines downstream basin counts and rankings.The threshold removes shallow basins treated as noise.Source: §1 · p. 1
- Persistence captures topological prominence but does not ensure physically plausible runoff, so the framework checks retained basins against D8 sinks and flow-accumulation areas.Both representations are derived from the same DEM.Source: §1 · p. 1
- The framework restricts the language model to recorded decision points and keeps geospatial computation deterministic, with three human-in-the-loop checkpoints.Decision points cover threshold selection, evidence-grounded recommendation, self-review, and question answering.Source: §1 · pp. 1–2
- The proposed framework combines public-data queries, deterministic terrain analysis, and language-model decisions restricted to recorded choice points.Its topology-guided method selects a persistence threshold, ranks basins with a relative flood-priority score, and checks results against D8 routing.Source: §1 · p. 2
2 GeoAI Agentic Framework
The framework turns a place name or boundary into ranked terrain basins and an evidence-grounded report using aligned public data and deterministic analysis. Language-model calls support selected decisions, while users can review and revise consequential choices.
2.1 System Overview and Data Acquisition
- Runs begin with a place name, drawn polygon, or uploaded boundary and retrieve aligned public data; elevation is required, while missing optional layers are omitted or replaced by documented defaults.Substitutions are recorded in the evidence bundle and report; numerical stages are deterministic for fixed inputs.Source: §2.1 · p. 2
2.2 Topology-Guided Basin Extraction
- The controller builds candidate persistence thresholds from distribution percentiles and a default, then the language model selects one using the candidate table and basin-count changes before TTK extracts retained basins.Both the candidate-based estimate and achieved basin count are recorded.Source: §2.2 · p. 2
2.3 Basin Prioritization
- Each retained basin receives a relative flood-priority score from six normalized factors, with default entropy weights or user-selected presets and custom weights.A sensitivity check compares the top-ranked basin across balanced, life-safety, infrastructure-protection, and entropy weighting schemes; the score is not a calibrated flood probability or inundation estimate.Source: §2.3 · p. 2
2.4 Screening-Level Runoff and Consistency Checks
- Screening-level peak runoff supports recommendation generation but is excluded from the priority score; D8 checks compare Morse minima with D8 sinks and contributing areas with flow-accumulation areas.The area comparison uses Pearson correlation and R^2 after log transformation.Source: §2.4 · p. 3
2.5 Recommendation and Human Oversight
- Separate language-model calls draft recommendations from the evidence bundle and review them for numerical consistency, unsupported certainty, and omitted data limitations.Users can revise the threshold, factor weights, or recommendation at three checkpoints; failures do not modify completed numerical results.Source: §2.5 · p. 3
3 Evaluation
The evaluation tests terrain simplification, cross-model spatial consistency, and ranking robustness, using College Park as an end-to-end case study and four additional areas for spatial checks. In College Park, 77% of retained basin minima also coincided with D8 sinks, and catchment 8 ranked first under all four tested weighting schemes.
3 Evaluation
- The College Park case used a user-specified priority to protect the university campus and nearby residential streets from flooding.Data retrieval returned ten categories, including a 192×235 USGS 3DEP elevation grid and a 100-year rainfall intensity of 79 mm/h.Source: §3 · p. 3
- The first TTK pass found 278 raw minima, and the selected 0.508 m, 70th-percentile threshold had an estimated 83 retained minima.The evaluation covered terrain simplification, cross-model spatial consistency, and robustness of the final ranking and recommendation.Source: §3 · p. 3
- 77% of the 70 retained College Park basins had minima that were also D8 sinks, with sink agreement summarized across four study areas in Figure 2.The remaining 23% lay mostly on flat areas where D8 assigns a downslope neighbor to a cell retained as a Morse minimum.Source: §3 · p. 3, Figure 2 · p. 4
- Catchment 8 ranked first under entropy weighting with a 0.249 km2 contributing area and 0.61 relative flood-priority score, remaining first under all four tested weighting schemes.It lay within FEMA’s Special Flood Hazard Area and 6.8 m from the nearest mapped water feature.Source: §3 · p. 3
- Figure 3 compares D8 flow-accumulation area with Morse contributing area on log-scaled axes, with a log-space linear fit and per-panel statistics.The evaluation includes College Park and four additional areas for spatial consistency checks.Source: §3 · p. 3, Figure 3 · p. 4
4 Conclusion
The framework combines public geospatial data, deterministic spatial analysis, language-model reasoning, and human review for preliminary urban flood screening. Across five study areas, it showed agreement between Morse terrain structures and D8 routing, while the evaluation remains limited in scope and lacks observed flood events as ground truth.
4 Conclusion
- Across five study areas, the framework showed consistent agreement between Morse-based terrain structures and independent D8 routing.Sensitivity analysis and self-review helped identify unstable rankings and unsupported statements.Source: §4 · p. 4
- The framework is designed to make preliminary flood-risk screening more accessible and transparent through traceable evidence and user-inspectable outputs.Users can inspect results through an interactive map, report, and LLM interface; human checkpoints keep consequential decisions reviewable.Source: §4 · p. 4
- The evaluation is limited to a small number of study areas, U.S.-specific data services, simplified exposure representation, and no observed flood events as ground truth.The authors identify broader geographic evaluation, more complete data, multiple LLMs, and observed or simulated flood information as future work.Source: §4 · p. 4