RAPID: A Reproducible Multi-Agent Pipeline for Interpretable Disaster Damage Assessment from Satellite and Street-View Imagery

---

Abstract

With the increasing frequency and intensity of extreme climate events, there is a growing demand for intelligent, scalable, and autonomous approaches to disaster damage assessment. Existing methods, largely based on supervised learning and task-specific fine-tuning, struggle to generalize under domain shifts, long-tailed data distributions, and heterogeneous geospatial data sources. This paper introduces RAPID, an autonomous multi-agent pipeline for interpretable disaster damage assessment — encompassing the assessment of damage levels, interpretation of damage types and degrees, and generation of actionable suggestions for response, remediation, and recovery.

Unlike conventional approaches that rely on single-task supervised models, RAPID coordinates multiple specialized agents to perform cross-view understanding, image restoration, structured damage recognition, and geographical reasoning across heterogeneous data modalities. Without task-specific fine-tuning, RAPID supports zero-shot damage assessment by jointly leveraging complementary information from remote sensing and ground-level perspectives. Experiments show that RAPID achieves an overall accuracy of 0.92 on multi-disaster type classification and up to 0.627 on cross-view damage severity prediction.

Keywords: Disaster Assessment · Vision-Language Models · Cross-View Imagery · Zero-Shot Learning · Multi-Agent Pipeline

---

Overview

Figure 1. RAPID: An Autonomous Multi-Agent Framework for Disaster Damage Intelligence

RAPID addresses three core research questions:

• RQ1 — How can an autonomous multi-agent pipeline achieve multimodal disaster understanding across different geospatial data sources without task-specific fine-tuning, while maintaining robustness under domain shifts?
• RQ2 — How can multiple agents efficiently coordinate perception, data restoration, damage recognition, and reasoning to generate structured, interpretable assessment results?
• RQ3 — To what extent can such a system automatically produce location-specific and decision-relevant disaster intelligence, and what are its strengths and limitations in supporting real-world disaster assessment?

---

Pipeline

The framework comprises four collaborative specialized agents coordinated through task decomposition:

Agent	Abbreviation	Core Role	Key Output
Disaster Perception Agent	DPA	Zero-shot identification of disaster type, image modality, and structural context; plans the downstream workflow	Disaster label + confidence score + task plan
Image Restoration Agent	IRA	Diagnoses image quality issues (blur, haze, low-light) in SVI/RSI; applies constrained enhancement strategies to preserve disaster-relevant visual evidence	Restored imagery + quality score (Q)
Damage Recognition Agent	DRA	Structured damage diagnosis across cross-view and bi-temporal settings; severity classification without task-specific fine-tuning	Severity label · object-level indicators · confidence scores
Disaster Reasoning Agent	DReA	High-level cognitive synthesis; causal interpretation, recovery recommendations, and structured disaster report generation	Decision-relevant disaster report

Agent Details

Disaster Perception Agent (DPA) comprises three modules: ModePerceiver (zero-shot image mode + disaster type recognition), DisasterReasoner (natural-language scene explanation with visual evidence), and TaskPlanner (downstream agent orchestration).

Image Restoration Agent (IRA) evaluates three restoration branches — heuristic baseline, Gemini-guided planner, and image-only Gemini enhancement — and accepts a branch output only when the composite quality score Q (combining contrast, sharpness, and NIQE-proxy) improves beyond a preset margin over the original.

Damage Recognition Agent (DRA) operates through four complementary tasks: (1) cross-view hurricane damage prediction from paired RSI+SVI, (2) bi-temporal visual change analysis from pre/post SVI, (3) wildfire-specific five-level classification, and (4) object-level detection + instance segmentation for spatial damage characterization. Evaluated with both standard metrics and the severity-aware Normalized Cross-Severity Error (NCSE):

$$\text{NCSE} = \frac{1}{N}\sum_{i=1}^{N}\frac{\left|y_i - \hat{y}_i\right|}{K-1}$$

Disaster Reasoning Agent (DReA) ingests structured JSON outputs from DRA and applies templated chain-of-thought reasoning to generate causal explanations, secondary risk assessments, and FEMA-guideline-aligned recovery recommendations — evaluated by both LLM-based and human expert scoring across factual consistency, causal plausibility, information completeness, and actionability.

---

Datasets

RAPID is evaluated across three complementary multimodal disaster dataset categories, covering cross-view, bi-temporal, and multi-hazard scenarios:

Dataset	Data Type	Images	Disaster	Severity	Source	Agents
A	SVI + RSI pairs	300	Hurricane	3 levels	CVDisaster (Li et al., 2025)	DPA · IRA · DRA · DReA
B	Bi-temporal SVI	300	Hurricane	3 levels	BiTemporal (Yang et al., 2025)	DPA · IRA · DRA · DReA
C1	Post-disaster SVI	188	Drought / Earthquake / Flood / Ice Storm / Wildfire	N/A	Incidents Dataset (Weber et al., 2020)	DPA
C2	Post-disaster SVI	295	Wildfire	5 levels	LA DINS (2025)	DPA · DRA

Geographic coverage spans California and Florida, USA. Dataset A and B each contain 150 image pairs; Dataset C totals 483 images across diverse hazard types.

Geolocation Distribution	Dataset Statistics

---

Results

Disaster Perception Agent — Multi-Disaster Type Classification

Model	Overall Accuracy
GPT-5-mini	0.92
GPT-5.1	0.88
Gemini-2.5-flash	0.86

Image Restoration Agent — Visual Quality Enhancement

Restoration consistently improves quality scores across all disaster types and image modalities. Representative results:

Category	Image Type	Q_original	Q_baseline	Q_planner	Q_gemini
Dataset A — Hurricane	Satellite	0.62	0.73	0.71	0.69
Dataset A — Hurricane	SVI	0.75	0.78	0.76	0.79
Dataset B — Hurricane	SVI	0.76	0.78	0.78	0.79

Damage Recognition Agent — Severity Prediction

Model	Dataset A Accuracy	Dataset B Accuracy	Dataset C Accuracy
Gemini-3-Pro	0.627	0.493	0.442
GPT-5.1	0.573	0.591	0.570
GPT-5-mini	0.387	0.503	0.573

Errors concentrate at adjacent severity levels, validating the role of NCSE as a more sensitive evaluation metric.

Disaster Reasoning Agent — Report Quality

LLM-based and human evaluation of multimodal disaster reasoning across Gemini-3-Pro, Gemini-2.5-Pro, and GPT-5.1

Example Outputs

LLM-Based Object Detection	Final Structured Report

---

Repository Structure

RAPID/
├── Disaster Perception Agent/
│   ├── DisasterPerceptionAgent.py       # ModePerceiver + DisasterReasoner + TaskPlanner
│   └── Prompt--Disaster Perception Agent
├── Image Restoration Agent/
│   ├── test.py                          # IQA diagnosis + 3-branch restoration
│   └── Prompt--Image Restoration Agent
├── Damage Recognition Agent/
│   ├── SVI&RSI.py                       # Cross-view hurricane (Dataset A)
│   ├── SVI-pre&post.py                  # Bi-temporal street-view (Dataset B)
│   ├── SVI-wildfire.py                  # Wildfire severity (Dataset C2)
│   ├── zero_shot_object_detection_Agent3.ipynb
│   └── Prompt--Damage Recognition Agent
├── Disaster Reasoning Agent/
│   ├── Large Language Model-based evaluation.py
│   ├── test.py                          # Report generation + evaluation
│   └── Prompt--Disaster Reasoning Agent
└── figure/

---

Citation

---

Contact

Yifan Yang — Department of Geography, Texas A&M University
yyang295@tamu.edu · rayford295.github.io

Lei Zou (Corresponding) — Department of Geography, Texas A&M University
lzou@tamu.edu