Source-linked AI summary
RAPIDMap: Rapid Multi-Agent Pipeline for Interpretable Disaster Mapping from Satellite and Street-view Imagery
Yifan Yang, Lei Zou
TL;DR
Traditional disaster maps primarily visualize spatial damage, while RAPIDMap uses a zero-shot multi-agent pipeline combining remote sensing and street-view imagery for structured disaster assessment and mapping. The framework generates map-ready disaster intelligence, and the results demonstrate its potential to support timely emergency response and recovery decisions.
Problem
Traditional disaster maps primarily visualize the spatial distribution of damage rather than richer intelligence for each mapped location.
Source: Disaster Mapping Agent (DMA) · p. 8Method
RAPIDMap combines remote sensing and street-view imagery in four agents for disaster perception, image restoration, damage recognition, and disaster mapping.
Source: Multi-Agent Pipeline · p. 2, Experimental Datasets · p. 3Results
The results demonstrate RAPIDMap's potential for automated disaster perception, damage recognition, reasoning, and structured disaster-intelligence generation.
Source: Conclusion · p. 9Takeaways & Limitations
For each mapped location, RAPIDMap can provide disaster type, damage severity, object recognition, confidence scores, and recovery recommendations.
Source: Disaster Mapping Agent (DMA) · p. 8
Abstract
from arXiv · showhide
Rapid and reliable disaster mapping of impacted areas, damaged infrastructure, and affected populations is essential for emergency response and recovery. However, existing AI-based approaches often require extensive manual annotation, lack cross-hazard generalization, and rely on single-modal observations. To address these challenges, this paper proposes RAPIDMap, a rapid multi-agent pipeline for zero-shot interpretable disaster mapping from satellite and street-view imagery. The framework integrates four intelligent agents: Disaster Perception Agent (DPA), Image Restoration Agent (IRA), Damage Recognition Agent (DRA), and Disaster Mapping Agent (DMA). By combining remote sensing and street-view data, RAPIDMap eliminates the need for manual fine-tuning, generalizes across multiple disaster categories, and generates structured, map-ready disaster intelligence with recovery recommendations.
Multi-Agent Pipeline
RAPIDMap organizes disaster assessment and mapping into four coordinated agents that interpret imagery, restore degraded inputs, recognize damage, and produce map-ready outputs. The framework evaluates multiple model backbones and adds map-level reasoning for structured disaster intelligence.
Multi-Agent Pipeline
- RAPIDMap comprises four agents: DPA for perception and planning, IRA for image quality control, DRA for structured damage diagnosis, and DMA for mapping.DPA identifies modality and disaster type; IRA applies constrained enhancement; DRA produces severity classifications; DMA performs georeferencing, alignment, and GIS-ready output generation.Source: Multi-Agent Pipeline · p. 2
- DPA generates structured task-planning signals, while IRA preserves disaster-relevant evidence during constrained enhancement of street-view and remote-sensing imagery.Source: Multi-Agent Pipeline · p. 2
- The pipeline evaluates GPT-5.1, GPT-5.1-mini, Gemini-2.5-flash, Gemini-2.5-Pro, and Gemini-3-Pro for DPA, IRA, and DRA, while DMA uses GPT-5.2 for map-level reasoning.Source: Multi-Agent Pipeline · p. 3
Experimental Datasets
The study assembles disaster imagery to test cross-view, bi-temporal, and multi-hazard capabilities across hurricane and wildfire events in multiple U.S. regions. Table 1 summarizes how these datasets support evaluation of the RAPIDMap agents.
Experimental Datasets
- The experimental system contains cross-view remote-sensing and street-view pairs, bi-temporal pre- and post-disaster street-view pairs, and multi-hazard post-disaster street-view data.Source: Experimental Datasets · p. 3
- The datasets include representative hurricanes and wildfires across multiple U.S. disaster areas, including California and Florida.Source: Experimental Datasets · p. 3
- Figure 3 presents the geographic distribution of the included disasters together with example images.Source: Experimental Datasets · p. 3, Figure 3 · p. 3
- Table 1 summarizes dataset types, composition, characteristics, and their roles in evaluating different RAPIDMap agents.Source: Experimental Datasets · p. 3, Table 1 · p. 4
Disaster Perception Agent (DPA)
The Disaster Perception Agent classifies input images into seven disaster categories using zero-shot perception. Across three evaluated language-model backbones, overall accuracy is strong, although earthquake recognition is comparatively difficult.
Disaster Perception Agent (DPA)
- DPA assigns each input image to one of seven categories: drought, earthquake, flood, hurricane, ice storm, wildfire, or others.Source: Disaster Perception Agent (DPA) · p. 4, Table 2 · p. 5
- 0.86–0.92 overall accuracy was achieved across the three evaluated language-model backbones.Source: Disaster Perception Agent (DPA) · p. 4
- Earthquake recognition was the most challenging category, with F1-scores ranging from 0.65 to 0.71.Source: Disaster Perception Agent (DPA) · p. 4
- The evaluation reports per-class precision, recall, and F1-score alongside overall accuracy for multi-disaster classification.Source: Disaster Perception Agent (DPA) · p. 4, Table 2 · p. 5
Image Restoration Agent (IRA)
The Image Restoration Agent compares enhancement strategies using a composite image-quality score and qualitative visual assessment across satellite and street-view imagery. Performance depends on modality: planner-based methods lead on satellite images, whereas Gemini performs best on street-view images.
Image Restoration Agent (IRA)
- Q=0.4C+0.4S−0.2N combines normalized contrast, sharpness, and an NIQE proxy, with higher values indicating better visual quality.Source: Image Restoration Agent (IRA) · p. 5
- 0.62→0.73/0.71 Q gains were achieved by the baseline and planner-guided tool chains on satellite images.Source: Image Restoration Agent (IRA) · p. 5
- 0.75→0.79 Q was achieved by Gemini on street-view images, outperforming both baseline and planner methods.Source: Image Restoration Agent (IRA) · p. 5
- Figure 4 qualitatively compares baseline, planner-guided, and Gemini image-only restoration across illumination, contrast, structural clarity, and disaster-relevant details.Source: Image Restoration Agent (IRA) · p. 5, Figure 4 · p. 6
Damage Recognition Agent (DRA)
The Damage Recognition Agent evaluates damage severity across hurricane and wildfire datasets using accuracy and severity-aware error metrics. Results are dataset-dependent, but the most accurate model is also the most reliable under severity-weighted evaluation.
Evaluation setup
- DRA performance is dataset-dependent across three hurricane or wildfire damage-severity datasets.The evaluation compares overall accuracy with Normalized Cross-Severity Error (NCSE).Source: Damage Recognition Agent (DRA) · p. 6
- Overall accuracy measures the proportion of predictions that exactly match the ground-truth damage-severity label.NCSE weights errors by their distance from the true severity level and normalizes the result to [0, 1].Source: Damage Recognition Agent (DRA) · p. 6
Results
- In all three cases, the model with the highest accuracy also attains the lowest NCSE.This indicates aligned gains in exact correctness and severity-aware reliability.Source: Damage Recognition Agent (DRA) · pp. 6–7
Experimental scope
- Table 4 reports DRA performance across multiple disaster severity levels, while Table 5 compares multimodal models on cross-view and bi-temporal street-view tasks.The evaluated settings include joint satellite–street-view inference and pre-disaster versus post-disaster street-view comparison.Source: Damage Recognition Agent (DRA) · p. 7, Table 5 · p. 8, Table 4
Disaster Mapping Agent (DMA)
The Disaster Mapping Agent evaluates multimodal disaster reasoning and produces spatially explicit map outputs. Beyond damage visualization, the pipeline attaches structured disaster intelligence and recovery recommendations to mapped locations.
Evaluation
- DMA evaluation compares automatic LLM-generated assessments with manual assessments across four disaster-reporting dimensions.The dimensions are factual consistency, plausibility, completeness of information, and actionability of recovery recommendations.Source: Disaster Mapping Agent (DMA) · p. 8
Mapping outputs
- Each mapped location can include disaster type, damage severity, object recognition, confidence scores, and recovery recommendations.These outputs extend traditional maps that primarily show the spatial distribution of damage.Source: Disaster Mapping Agent (DMA) · p. 8
- The pipeline generates spatially explicit disaster mapping results for mapped locations.Figure 7 presents examples of these generated mapping outputs.Source: Disaster Mapping Agent (DMA) · p. 8, Figure 7 · p. 9
Conclusion
RAPIDMap is a zero-shot, multimodal multi-agent pipeline for disaster damage assessment and mapping. By integrating remote sensing and street-view data, it generates structured disaster intelligence intended to support timely emergency response and recovery decisions.
Contribution
- RAPIDMap performs zero-shot disaster damage assessment and mapping using multimodal imagery.The framework integrates remote sensing and street-view data within a multi-agent pipeline.Source: Conclusion · p. 9
Implication
- The framework automates disaster perception, damage recognition, and reasoning to generate structured disaster intelligence.These capabilities are presented as supporting timely decision-making for emergency response and recovery.Source: Conclusion · p. 9