Enhancing Demand-Oriented Regionalization with Agentic AI and Local Heterogeneous Data for Adaptation Planning

AI summary

Overview

• Research area: Agentic AI for urban planning and geospatial decision support, specifically demand-oriented regionalization for disaster risk and climate adaptation.
• Technical level: Intermediate. The paper is a system description and case-study demonstration; it assumes some familiarity with clustering, self-organizing maps, and geospatial data, but presents no equations or benchmark tables.
• Scope (one sentence): The paper introduces an agentic AI planning support system built on the RepSC-SOM regionalization framework, demonstrated on flooding risk in Jacksonville, Florida, that lets users generate, explore, and refine demand-oriented planning regions through natural-language, human-in-the-loop interaction.

What This Paper Is About

Conventional planning units such as census tracts, zip codes, and neighborhoods were designed for purposes like mail delivery or population data collection, so they often fail to reflect hazard exposure, social vulnerability, or the specific demands of local communities. Planners who lack geospatial or coding expertise have few practical ways to build alternative spatial divisions from fine-grained local data. The paper's goal is to make data-driven regionalization usable by those planners by pairing a spatially constrained self-organizing map algorithm with AI agents that suggest input features, guide spatial constraints, and support iterative exploration.

Key Contributions

1. A planning support system with agentic AI for demand-oriented regionalization. A central planning agent orchestrates the workflow through a conversational interface, turning open-ended planning questions into end-to-end analyses. It geocodes the user's location, selects and configures geospatial datasets (for example, the Florida Geographic Data Library), visualizes features as interactive map layers, summarizes the dataset, collects user preferences on feature selection and number of regions, calls the regionalization algorithm, and places the resulting partitions on the map.
2. The RepSC-SOM framework (Representative-initialized, Spatially Constrained Self-Organizing Map). This extends a traditional self-organizing map with representative-based initialization, adaptive geographic filtering, and region-growing refinement to produce spatially coherent regions that capture localized vulnerabilities and risk profiles.
3. A transparent, human-in-the-loop workflow. Every step of the regionalization is designed to be interpretable, letting users adjust configurations, intervene in the process, provide feedback on outputs, change input features, and explore alternative configurations. AI agents respond to that input by updating the regionalization so final regions reflect both computational rigor and user knowledge.
4. A demonstration case study. The platform is demonstrated on flooding-related risk in Jacksonville, Florida, showing interactive selection of features based on study area and disaster type, generation of regions via RepSC-SOM, and iterative refinement of the output.

Main Findings

• Agentic orchestration lowers the barrier to advanced spatial analysis: The architecture lets non-technical users run machine learning-based geospatial decision support by specifying only a study area and a hazard of interest in natural language.
• Regions can be demand-oriented rather than administratively fixed: By integrating heterogeneous local socioeconomic and environmental data, the system produces spatially contiguous regions that capture localized vulnerabilities and hazard-specific risk profiles, rather than inheriting the rigidity of administrative or census boundaries.
• The regionalization pipeline is a transparent three-step Embedding–Clustering–Refining process: Input features selected by users are projected into higher-dimensional latent space with an auto-encoder; the SOM is initialized according to a geographic threshold derived from the semivariogram of the input feature data, which also determines the number and initial states of neurons; grid cells are then assigned to Best Matching Units with candidate neurons first filtered by a geographic threshold using the Haversine distance, and neurons updated with their spatial neighbors; finally, region growing merges initial regions guided by the user's expected number of regions, feature similarity, and spatial constraints to improve compactness and reduce fragmentation.
• The case study demonstrates feature suggestion by disaster type: For Jacksonville, Florida and flooding, the LLM-based system dynamically selected and presented candidate features spanning socioeconomic indicators, environmental conditions, and infrastructure vulnerability.
• Adaptability to dynamic conditions: The authors argue the interactive, human-in-the-loop design allows regions to flexibly respond to changing hazard patterns and evolving urban conditions, supporting resilient and equitable adaptation strategies.
• No quantitative evaluation is reported. The paper contains no benchmark comparisons, accuracy metrics, runtime figures, or counts of regions; the demonstration is qualitative and screenshot-based.

Methodology in Plain English

The authors build a conversational planning tool around three ideas working together.

First, an AI agent handles the logistics. A user types in a place and a hazard. The agent finds the location, pulls relevant geospatial datasets, shows the data on a map, describes what the data contains, and asks the user which features to use and how many regions to create. When the user answers, the agent runs the regionalization and maps the result, keeping track of the conversation so it can respond to further requests.

Second, the regionalization algorithm works on a raster grid of the study area and proceeds in three stages. In Embedding, the chosen input features are compressed through an auto-encoder into a latent space that captures interactions among them; at the same time the self-organizing map's neurons are initialized based on a geographic threshold computed from the data's semivariogram, which sets how many neurons there are and where they start. In Clustering, each grid cell is assigned to a "best matching unit" — but only after candidate neurons are filtered by that same geographic threshold using Haversine distance, so a cell can only match to nearby neurons. The winning neuron's weights are then updated using the cells assigned to it. In Refining, the preliminary clusters are turned into initial regions and merged through region growing, governed by the number of regions the user asked for, how similar the features are, and spatial constraints, which makes the regions more compact and less fragmented.

Third, the user stays in the loop. Results are shown on an interactive map, and the user can give feedback or change settings, with the agent updating the regionalization accordingly. The authors illustrate the whole flow with flooding risk in Jacksonville, Florida.

Why This Matters

The work targets a persistent mismatch in urban governance: the boundaries planners are stuck with were not designed for climate risk or social vulnerability, and rebuilding them requires expertise that most planning offices do not have. By putting an agentic conversational layer in front of a spatially constrained clustering method, the paper points toward planning units that can be generated on demand for a specific hazard and a specific community.

Real-world applications:

• Disaster preparedness and response: Defining intervention zones for flooding or other hazards that match actual exposure and vulnerability patterns rather than administrative lines.
• Resource allocation and prioritization: Identifying priority areas for investment or protective measures where risk concentrates.
• Climate adaptation planning: Producing adaptive planning units that can be regenerated as hazard patterns and urban conditions change.
• Community engagement and participatory planning: Using shared, transparent, user-adjustable regions as reference points for stakeholder participation and decision making.

Industry relevance: the approach speaks to government planning and emergency management agencies, GIS and urban analytics vendors, and insurers or infrastructure operators who need defensible, localized risk segmentation without in-house machine learning specialists. It also illustrates a broader pattern for the geospatial software industry: wrapping specialized spatial algorithms in agentic interfaces so domain experts, not just coders, can drive them.

Future Directions

• Quantitative evaluation. The paper reports no accuracy, compactness, stability, or comparison metrics against conventional units or alternative regionalization methods; establishing such benchmarks is an open task.
• Validation across hazards and regions. The demonstration covers a single case — flooding in Jacksonville, Florida — so how the agentic feature suggestion and RepSC-SOM refinement generalize to other hazards, cities, and data environments is untested here.
• Rigorous user studies with planners. The claim that the platform lowers barriers for non-technical planners is asserted rather than measured; usability and decision-quality studies would test it.
• Agent reliability and data governance. Questions remain about how the agent handles missing, inconsistent, or fine-grained local data, how transparent its feature suggestions are to users, and how users can audit or override its reasoning.

Target Audience

Urban planners and planning support system researchers, GIScientists and geospatial data scientists interested in constrained clustering, and AI researchers working on agentic or LLM-driven decision support in the built environment. It is also relevant to emergency management and climate adaptation practitioners who want to understand how demand-oriented planning units could be generated, and to graduate students in urban planning, geography, or computer science looking for an applied example of human-in-the-loop AI in spatial analysis. Readers seeking quantitative benchmarks or a production-ready tool will not find them here; the contribution is architectural and demonstrative.

Authors’ abstract

Conventional planning units or urban regions, such as census tracts, zip codes, or neighborhoods, often do not capture the specific demands of local communities and lack the flexibility to implement effective strategies for hazard prevention or response. To support the creation of dynamic planning units, we introduce a planning support system with agentic AI that enables users to generate demand-oriented regions for disaster planning, integrating the human-in-the-loop principle for transparency and adaptability. The platform is built on a representative initialized spatially constrained self-organizing map (RepSC-SOM), extending traditional SOM with adaptive geographic filtering and region-growing refinement, while AI agents can reason, plan, and act to guide the process by suggesting input features, guiding spatial constraints, and supporting interactive exploration. We demonstrate the capabilities of the platform through a case study on the flooding-related risk in Jacksonville, Florida, showing how it allows users to explore, generate, and evaluate regionalization interactively, combining computational rigor with user-driven decision making.