Skip to content

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

🕒 Published (v1): 2025-11-13 23:50 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper presents a planning support system that wraps a novel spatially constrained self-organizing map (RepSC-SOM) with an LLM-based agentic interface, enabling non-technical urban planners to generate demand-oriented spatial planning regions for disaster risk management. The agent orchestrates geocoding, dataset selection, algorithm configuration, and iterative user refinement through natural language. A case study on flood risk in Jacksonville, FL demonstrates end-to-end functionality.

Problem

Standard administrative planning units (census tracts, zip codes) are misaligned with disaster vulnerability patterns because they were designed for unrelated purposes (e.g., mail delivery, population counting). Data-driven regionalization methods exist but require geospatial expertise, limiting adoption by planners. No prior system combined flexible spatial clustering with an agentic LLM interface enabling iterative, non-technical exploration.

Method

The system has two integrated components:

RepSC-SOM — a three-stage Embedding-Clustering-Refining pipeline over a raster grid: 1. Embedding: autoencoder projects input socioeconomic/environmental features into latent space; SOM neuron count and initialization derived from semivariogram-based geographic threshold. 2. Clustering: iterative BMU assignment with two-stage filtering — candidate neurons first filtered by Haversine distance within a geographic threshold, then matched by feature similarity. 3. Refining: region-growing post-processing merges spatially fragmented clusters guided by target region count, feature similarity, and spatial contiguity constraints.

Agentic orchestration layer: a central planning agent interprets natural language inputs (study area, disaster type), calls geocoding and metadata-filter tools against the Florida Geographic Data Library, presents candidate features to the user, invokes RepSC-SOM with user-selected parameters, visualizes results on an interactive map, and supports iterative human-in-the-loop refinement through conversational feedback.

Key Contributions

  • RepSC-SOM: extends traditional SOM with semivariogram-based adaptive geographic filtering and region-growing refinement for spatially coherent, demand-oriented regions.
  • Agentic planning support system with tool-calling (geocoding, geoprocessing, metadata filtering, visualization) orchestrated through a conversational LLM agent.
  • Human-in-the-loop design allowing planners to adjust features, constraints, and region counts mid-process without coding.
  • Integration of heterogeneous local socioeconomic and environmental data (Florida Geographic Data Library) for hazard-specific regionalization.

Results

  • Qualitative demonstration only via case study (Jacksonville, FL, flooding risk).
  • No quantitative benchmarks, ablation studies, or comparisons against baseline regionalization methods are reported in this paper.
  • Platform successfully generates interactive regionalization outputs from user-specified disaster type and study area through the full pipeline.

Limitations

  • No quantitative evaluation; results are demonstrated rather than measured against baseline methods or ground truth.
  • Case study limited to a single city (Jacksonville, FL) and single hazard type (flooding); generalizability untested.
  • Dependence on Florida Geographic Data Library limits immediate portability to other regions without data source extension.
  • Raster grid representation may lose sub-grid spatial detail.
  • NeurIPS 2025 UrbanAI Workshop paper — scope and length constrained; system described at high level without ablations.

Relevance to Agentic AI / LLM Agents

This work is a concrete instantiation of an LLM agent as a domain-specialist orchestrator: the agent maintains dialogue context, selects and calls heterogeneous geospatial tools, and coordinates a multi-step analytical pipeline in response to open-ended user intent — a standard agentic pattern applied to a non-software domain. It demonstrates the human-in-the-loop refinement paradigm where the agent mediates between a complex ML algorithm and a non-technical user, which is directly relevant to broader work on agentic AI lowering barriers to expert-system use. The architecture — central planning agent + specialized tool calls + memory — maps cleanly onto emerging agentic frameworks. Of particular interest to the LLM agents community is the demonstration that agentic wrappers can make geospatial ML accessible without code, validating tool-augmented agents beyond software engineering tasks.