PaperlayerView exact revision on arXiv

Source-linked AI summary

Empowering LLM Agents with Geospatial Awareness: Toward Grounded Reasoning for Wildfire Response

Yiheng Chen, Lingyao Li, Zihui Ma, Qikai Hu, Yilun Zhu, Min Deng, Runlong Yu

arXiv:2510.12061v2cs.AI

TL;DR

Wildfire response decisions must be made under uncertain, rapidly changing conditions, while existing approaches can lack context and LLMs remain weakly grounded in geography. The paper introduces GAL to retrieve and structure geospatial evidence for LLM reasoning; evaluations across California fires report stronger accuracy and temporal stability than physical and LSTM baselines for daily personnel and cost forecasts.

  • Problem

    Wildfire resource decisions face incomplete, uncertain information, while existing approaches can lack context, generalizability, and interpretability and LLMs remain weak in geospatial grounding.

    Source: §1 · pp. 1–2

  • Method

    GAL retrieves infrastructure, demographic, terrain, and weather data from wildfire detections and formats the evidence as a unit-annotated perception script, reinforced with historical analogs and daily change signals.

    Source: §1 · p. 2

  • Results

    Geospatially grounded agents consistently outperform physical and LSTM baselines for daily personnel and cost forecasts, with higher accuracy and stronger temporal stability across California fires.

    Source: §5 · p. 9

  • Takeaways & Limitations

    GAL grounds resource decisions in structured spatial evidence and historical analogs, supporting auditable recommendations from minimal wildfire inputs.

    Source: §5 · p. 9

    Evaluation covers a subset of large 2020 California wildfires, and public geospatial data noise, delays, and resolution mismatches constrain absolute accuracy.

    Source: §6 · p. 10

Abstract

from arXiv · show

Effective disaster response is essential for safeguarding lives and property. Existing statistical approaches often lack semantic context, generalize poorly across events, and offer limited interpretability. While Large language models (LLMs) provide few-shot generalization, they remain text-bound and blind to geography. To bridge this gap, we introduce a Geospatial Awareness Layer (GAL) that grounds LLM agents in structured earth data. Starting from raw wildfire detections, GAL automatically retrieves and integrates infrastructure, demographic, terrain, and weather information from external geodatabases, assembling them into a concise, unit-annotated perception script. This enriched context enables agents to produce evidence-based resource-allocation recommendations (e.g., personnel assignments, budget allocations), further reinforced by historical analogs and daily change signals for incremental updates. We evaluate the framework in real wildfire scenarios across multiple LLM models, showing that geospatially grounded agents can outperform baselines. The proposed framework can generalize to other hazards such as floods and hurricanes.

1 Introduction

Wildfire decisions must allocate resources amid uncertain, rapidly changing conditions, while existing statistical methods often lack context, generalizability, and interpretability. The paper introduces GAL to ground LLM agents in structured geospatial evidence and evaluates their wildfire-response reasoning.

1 Introduction

  • Wildfires spread across large areas under complex topographic and weather influences, making resource allocation difficult when information is incomplete and uncertain.Existing hotspot, fire-radiative-power, and time-series approaches provide useful signals but often lack context, generalizability, and interpretability.Source: §1 · p. 1
  • LLMs offer few-shot generalization and broad knowledge integration, but remain weak at geospatial grounding and reliable spatial reasoning.This leaves open whether structured geospatial data can support disaster-response decisions.Source: §1 · p. 1
  • GAL retrieves infrastructure, demographic, terrain, and weather attributes from hotspot coordinates and timestamps, then encodes them in a compact, unit-annotated perception script.Schema validation, unit normalization, bounded ranges, historical analogs, and daily change signals reinforce decision stability.Source: §1 · p. 2
  • The paper evaluates grounded LLMs across California wildfires, reporting improved interpretability and alignment with operational outcomes and establishing a cross-model benchmark.The stated contributions include reasoning analysis on real wildfire events and an open benchmark for geospatially grounded reasoning and decision accuracy.Source: §1 · p. 2

2 Related Work

Prior wildfire work spans physical, statistical, and remote-sensing approaches, while disaster-response LLMs bring flexible language reasoning but remain limited in geospatial perception. This work distinguishes itself by dynamically grounding LLM agents in structured Earth data.

2.1 Traditional Wildfire Modeling

  • Physical simulations, statistical and time-series models, and remote sensing support wildfire prediction, but can be slow and lack contextual awareness.Examples include mechanistic fire-spread models, fire–atmosphere systems, historical-pattern approaches, and satellite-based detection and forecasting.Source: §2.1 · p. 2

2.2 LLMs for Disaster Response

  • LLM disaster-response studies explore geographic descriptors, maps, and geospatial data, yet models still struggle to perceive and reason reliably about hazard conditions.The cited work spans decision support, social-media location descriptors, and map-integrated analysis.Source: §2.2 · p. 2
  • Wildfire LLMs can describe behavior qualitatively but struggle with quantitative forecasts such as fire radiative power; related work also studies evacuation prediction.WildfireGPT was introduced for response tasks and later evaluated for quantitative fire-spread prediction (Xie et al., 2024).Source: §2.2 · pp. 2–3
  • Unlike prior studies, GAL dynamically retrieves structured Earth data and organizes it into a stable, unit-annotated representation for actionable outputs across evolving disaster contexts.The retrieval occurs at inference time through live databases.Source: §2.2 · p. 3

3 Methodology

The methodology links wildfire detections to response recommendations through a geospatial retrieval layer and a structured LLM reasoning process. GAL synchronizes environmental context with daily targets, formats it for stable reasoning, and uses historical analogs and validation to shape outputs.

3.1 Framework Design

  • The framework turns hotspot locations and detection dates into fire footprints, retrieves spatial attributes, and feeds a compact perception script to an LLM for daily personnel and cost recommendations.The retrieval includes infrastructure, demographics, terrain, and weather from a PostGIS–raster database.Source: §3.1 · p. 3

3.2 Data Preparation

  • Each fire–day pairs one incident-date perception script with daily personnel and cost targets, synchronizing dynamic inputs and joining static layers to that day's footprint.Fire-days with missing or inconsistent ground-truth records are excluded; absent geospatial fields are marked NA, and short weather gaps use a one-to-two-day rolling window to avoid future leakage.Source: §3.2 · p. 4

3.3 Geospatial Awareness Layer (GAL)

  • GAL clusters hotspot detections into daily fire footprints, retrieves aligned infrastructure, demographic, terrain, and weather attributes, and consolidates event features with temporal change signals.It derives cluster and global summaries, including activity measures and rolling temporal anchors, to support incremental decisions.Source: §3.3 · pp. 4–5

3.4 GAL-based LLM Reasoning

  • The reasoning module uses a fixed-slot, unit-normalized script, historical analogs as soft bounds, and rubric-guided reasoning to produce validated structured recommendations.Incremental mode also uses yesterday’s outputs and change cues; validation checks the JSON schema, units, and bounded ranges, with targeted reprompts after failures.Source: §3.4 · p. 5

4 Experiments

Across five held-out California wildfires, geospatially grounded LLMs achieved the lowest error in every evaluation column. Ablations and case analyses examined GAL and RAG effects, failure patterns, and model reasoning cues.

4.1 Dataset

  • The evaluation covered 14 California wildfires from 2020, holding out five events and using nine as the historical corpus and non-LLM-baseline training set.The split maintained spatial and temporal disjointness between training and evaluation.Source: §4.1 · pp. 5–6

4.2 Benchmark Models

  • The benchmark compared regular and reasoning LLMs with Physical and LSTM baselines under shared prompts, RAG, and output schemas, without fine-tuning.Models included Gemini and GPT families, including compact variants.Source: §4.2 · p. 6

4.3 Evaluation Protocol

  • For each fire, models forecast daily personnel and budget from same-day detections and GAL features; evaluation reported MAE and RMSE.RMSE highlights larger errors associated with operational under- or over-allocation.Source: §4.3 · p. 6

4.4 Model Performance

  • LLMs achieved the lowest error in every evaluation column; Physical and LSTM baselines never ranked first.Complex fires had higher uncertainty, yet grounded LLMs remained competitive; smaller reasoning models often matched or surpassed larger counterparts.Source: §4.4 · p. 6

4.5 Ablation Studies

  • GAL improved forecast alignment and spending trajectories, while RAG reduced error in most cases, with results varying by model and event.Without GAL, SCU personnel forecasts substantially overstated staffing after the fire subsided; GPT-o3-mini improved personnel forecasts across all four RAG-tested fires.Source: §4.5 · p. 7, Figure 3 · p. 8

4.6 Failure case analysis

  • GAL case analyses show text-only agents overallocated personnel in a low-exposure region and underallocated costs in high-risk terrain.Without GAL, SCU staffing was overestimated by a factor of 2–3, while CZU peak-day costs were underestimated by more than a factor of four.Source: §4.6 · pp. 8–9, Figure 6 · p. 9

4.7 Output Reasoning Analysis

  • Terrain typically dominated personnel predictions, while cost-feature importance varied by model, with GPT-5 shifting from terrain to resource cues.A post-hoc regressor analyzed features extracted from LLM reasoning traces and grouped into six categories.Source: §4.7 · p. 9

5 Discussion and Conclusion

The study introduces GAL to ground wildfire resource decisions in structured spatial evidence and reports improved forecasting across California fires. Its findings connect structured inputs, historical analogs, and feature attribution to performance, stability, and operational priorities.

5 Discussion and Conclusion

  • Geospatially grounded agents consistently outperform physical and LSTM baselines in forecasting daily personnel and cost across California fires.The study also reports higher accuracy and stronger temporal stability.Source: §5 · p. 9
  • Smaller reasoning models can match larger ones when inputs are normalized and semantically organized, while spatial and temporal cues enhance robustness.Historical analogs provide scale priors that moderate overreaction to noisy or incomplete signals.Source: §5 · p. 9
  • Personnel forecasts depend primarily on spread potential and exposure, whereas cost estimates emphasize accessibility and logistical demand.The authors describe these feature patterns as reflecting real-world operational priorities.Source: §5 · p. 9
  • GAL is presented as a generalizable interface for hazards beyond wildfire, including floods, hurricanes, and earthquakes.Source: §5 · p. 9

6 Limitations

The evaluation covers a subset of large 2020 California wildfires, and broader validation requires additional data connectors and local calibration. Public-dataset quality and the descriptive, post-hoc nature of feature attribution also bound the findings.

6 Limitations

  • Evaluation is limited to a subset of large 2020 California wildfires; validation across other regions, years, or hazards requires additional data connectors and local calibration.Source: §6 · p. 10
  • Public geospatial datasets may contain detection noise, temporal delays, or resolution mismatches, and data quality constrains absolute accuracy despite mitigation steps.Source: §6 · p. 10
  • Feature-importance analysis is post-hoc and descriptive rather than causal; controlled studies are needed to establish causal mechanisms in operational settings.Source: §6 · p. 10

Appendix

The appendix documents the reasoning prompts and wildfire events, adds a qualitative case study and expanded benchmarks, and reports configuration efficiency statistics.

Appendix

  • The appendix provides the full system and user prompts, event summaries, a qualitative case study, extended benchmarks, and token-usage and API-cost statistics.Source: Appendix · p. 12

A Prompt Design

  • The output defines daily personnel as total assigned headcount and daily budget as new cost incurred that day in USD.Source: A Prompt Design · pp. 12–13
  • Resource-estimation guidance accounts for continuing needs, cumulative costs, crew fatigue, and uncertainty rather than equating no detected hotspots with full extinguishment.Source: A Prompt Design · p. 13
  • The prompt asks for updated daily personnel and budget estimates using current conditions, changes from prior analysis, and fire and operational context.Its context includes cumulative costs and personnel-days, rolling trends, fire intensity, affected areas, and resource access.Source: A Prompt Design · pp. 12–13

B Wildfire Events and Statistics

The study uses fourteen large California wildfire events from late July through early October 2020, dividing them between historical retrieval and evaluation.

B Wildfire Events and Statistics

C Case Study: Daily Forecast Trajectory

The CZU trajectory illustrates GAL-conditioned forecasts that adjust resources to changing exposure, operational context, and signal availability while maintaining temporal stability.

C Case Study: Daily Forecast Trajectory

  • During the early phase, staffing rises as activity moves toward populated areas, even while FRP declines, prioritizing exposure and accessibility.The trajectory and its numbered callouts show personnel and cost forecasts with auditable rationales.Source: C Case Study: Daily Forecast Trajectory · p. 14
  • As intensity eases, personnel and cost decline gradually without oscillations, with rationales citing crew fatigue and logistical readiness.The authors attribute smooth reductions to analog-conditioned soft bounds.Source: C Case Study: Daily Forecast Trajectory · p. 14
  • On September 6, the model makes a small resource increase when the footprint is concentrated nearer populated zones, citing higher intensity and population risk.The response draws on buffer-based proximity and exposure density.Source: C Case Study: Daily Forecast Trajectory · p. 14
  • With minimal or missing hotspot detections, the model retains conservative patrol and mop-up baselines rather than reducing allocations to zero.The passage links this stability to analog-based priors and contingency readiness.Source: C Case Study: Daily Forecast Trajectory · p. 14
  • When the footprint expands again, personnel and cost rise alongside renewed increases in FRP and spatial extent.The rationales cite augmented aerial support and logistical coordination.Source: C Case Study: Daily Forecast Trajectory · p. 14
  • GAL supports explicit trade-offs: staffing increases when exposure or access risks outweigh intensity, then scales back when trends and logistics allow.The representation uses 3/7-day anchors and incremental updates for stability while analog-conditioned priors calibrate personnel and cost magnitudes.Source: C Case Study: Daily Forecast Trajectory · p. 15

D Extended Benchmark: Time-Series Baselines and Additional LLMs

Across four held-out California fires and two prediction tasks, the extended benchmark compares GAL-grounded agents with time-series and additional LLM baselines.

D Extended Benchmark: Time-Series Baselines and Additional LLMs

  • The stricter subset contains CZU August Lightning, El Dorado, LNU Lightning Complex, and SCU Lightning Complex, evaluated on daily total personnel and estimated daily cost-to-date.Competitors include four time-series models and Text-RAG, which uses unstructured FRP and area summaries with historical-analog retrieval.Source: D Extended Benchmark: Time-Series… · p. 15
  • Across both tasks and all four fires, GAL-grounded LLM agents consistently match or outperform all baselines.The comparison includes modern sequence models and retrieval-based alternatives.Source: D Extended Benchmark: Time-Series… · p. 15
  • DLinear, TimesNet, and PatchTST perform comparably to or worse than the Physical baseline, while TimeLLM is more competitive on cost but trails top GAL agents on most fires.Text-RAG consistently underperforms full GAL agents, indicating that generic retrieval alone does not match structured grounding.Source: D Extended Benchmark: Time-Series… · p. 15
  • Among added LLMs, GPT-5-web and Gemini-3-Pro are competitive in select settings, while DeepSeek-V3.2 and GPT-5.1 show moderate performance.GPT-5-web approximates a web-search retrieval baseline without structured geospatial context.Source: D Extended Benchmark: Time-Series… · p. 15

E Token Usage and API Cost per Fire

The paper reports per-fire token use and estimated API cost for models evaluated on the four-fire subset, noting that strong GAL configurations pair competitive prediction with modest costs.

E Token Usage and API Cost per Fire

Loading 2510.12061v2…