💻 computer science

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

This paper introduces RAPID, a reproducible multi-agent pipeline that leverages zero-shot reasoning across satellite and street-view imagery to provide interpretable, fine-grained disaster damage assessments and actionable recovery reports without requiring task-specific fine-tuning.

Original authors: Yifan Yang, Wenjing Gong, Kaili Zhang, Lei Zou, Zhengzhong Tu, Hao Li, Zongrong Li, Xinyue Ye

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yifan Yang, Wenjing Gong, Kaili Zhang, Lei Zou, Zhengzhong Tu, Hao Li, Zongrong Li, Xinyue Ye

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a team of specialized detectives working together to figure out exactly what happened after a massive storm, earthquake, or fire. Instead of relying on one super-smart detective who tries to do everything at once, this paper introduces RAPID, a "team of agents" that works together to assess disaster damage using photos taken from space (satellites) and from the ground (street views).

Here is how RAPID works, broken down into simple steps using everyday analogies:

The Problem: The "One-Size-Fits-All" Trap

Usually, computer programs used to check for disaster damage are like students who have memorized answers for a specific test. If the test changes (a different type of storm, a different city, or a new camera angle), they get confused. They also often need a teacher to manually show them thousands of examples before they can learn anything. This is too slow when a disaster just happened and we need answers now.

The Solution: The RAPID Detective Squad

RAPID is different. It doesn't need to be retrained for every new disaster. Instead, it uses four specialized "agents" (AI assistants) that pass the work down a line, like a relay race.

1. The Scout (Disaster Perception Agent)

What it does: This agent looks at the incoming photos and asks, "What kind of disaster is this? Is it a hurricane, a flood, or an earthquake? And do we need to clean up these blurry photos first?"
The Analogy: Think of this as the team leader who walks into a messy crime scene, identifies the type of crime, and decides which tools the rest of the team needs. It can tell the difference between a flood and a fire just by looking, even if it has never seen that specific flood before.

2. The Restorer (Image Restoration Agent)

What it does: Disaster photos are often dark, foggy, or blurry. This agent tries to sharpen them up. But here's the catch: it's careful. It tries three different ways to fix the picture and only keeps the version that actually looks clearer without inventing fake details.
The Analogy: Imagine a photo editor who tries three different filters to fix a blurry photo. They pick the one that makes the details pop but stops if a filter starts making the sky look purple or the trees look like they are melting. They only pass on the "cleanest" version of the truth.

3. The Inspector (Damage Recognition Agent)

What it does: Now that the photos are clear, this agent looks for specific damage. It counts broken windows, fallen trees, and downed power lines. It also gives a "severity score" (e.g., "mild damage" vs. "destroyed").
The Analogy: This is the insurance adjuster walking through the house with a clipboard. They don't just say "it's broken"; they say, "The roof has a hole, the tree is on the car, and the power is out." They can look at a satellite view and a street view at the same time to get the full picture.

4. The Reporter (Disaster Reasoning Agent)

What it does: This agent takes all the facts from the previous steps and writes a report. It explains why the damage happened and suggests what to do next (like "clear the debris" or "check the building's safety").
The Analogy: This is the press secretary or emergency coordinator. They take the inspector's notes and turn them into a clear story for the mayor or the public: "Here is what happened, here is the damage, and here are the first steps we should take."

How They Tested It

The researchers tested this team on real disasters like hurricanes, wildfires, and floods. They used a mix of "before and after" street photos and satellite images.

The Results: The team was very good at identifying what kind of disaster happened (about 92% accuracy).
The Limits: They were good at spotting damage, but sometimes they struggled to say exactly how bad the damage was on a precise scale (getting about 60% accuracy on severity). Also, while the "Reporter" agent was great at describing the facts, it sometimes struggled to come up with specific, practical solutions for how to fix things, because that requires deep local knowledge that a computer might not have.

The Bottom Line

RAPID is like a fast, reliable, and transparent early-warning system. It doesn't replace human experts or official inspectors. Instead, it acts as a powerful assistant that can quickly scan a huge area, clean up the photos, spot the obvious damage, and write a draft report for humans to review.

The paper emphasizes that this system is reproducible (you can run it again and get the same steps) and interpretable (you can see exactly which agent made which decision, so you know where the system might have made a mistake). It's designed to get emergency responders a head start when time is running out.

Technical Summary: RAPID – A Reproducible Multi-Agent Pipeline for Interpretable Disaster Damage Assessment

1. Problem Statement

The increasing frequency and intensity of extreme climate events necessitate rapid, accurate, and spatially specific disaster damage assessment for emergency response and resource allocation. Current deep learning approaches, primarily relying on supervised learning and task-specific fine-tuning of convolutional or transformer-based models, face three critical limitations:

Generalization Failure: Models struggle to generalize across domain shifts, long-tailed data distributions, and heterogeneous geospatial sources, particularly when labeled data for specific disaster events is scarce.
Modality Silos: Traditional research often relies on single-modal data (e.g., only satellite or only street-view imagery), neglecting the complementary nature of multi-source information required to capture diverse disaster impacts.
Lack of Interpretability and Reasoning: Existing systems often function as "black boxes" for pattern recognition, lacking the ability to integrate multimodal data to generate interpretable, actionable, and reasoning-based reports without extensive manual annotation.

2. Methodology: The RAPID Framework

The authors introduce RAPID (Reproducible multi-Agent Pipeline for cross-view Intelligence in disaster Damage assessment), an autonomous multi-agent system designed for zero-shot damage assessment. Instead of a single predictive model, RAPID decomposes the assessment workflow into four collaborative agents that process heterogeneous data (remote sensing imagery and street-view imagery) through an auditable, end-to-end pipeline.

2.1 System Architecture

The pipeline operates as a sequence of auditable state updates rather than a black-box prediction. Let $X$ denote the visual evidence; the workflow is defined as:
$z_p = A_{DPA}(X) \rightarrow \tilde{X} = A_{IRA}(X, z_p) \rightarrow z_d = A_{DRA}(\tilde{X}, z_p) \rightarrow Y = A_{DReA}(z_p, z_d, \tilde{X})$

The four specialized agents are:

Disaster Perception Agent (DPA): Acts as the entry point, identifying the input modality (satellite vs. street-view), disaster category, and the need for image restoration. It outputs a perception state ( $z_p$ ) including modality, disaster type, and a planned agent route.
Image Restoration Agent (IRA): Functions as a guarded quality control module. It evaluates degraded imagery using no-reference image quality assessment (IQA) metrics (contrast, sharpness, naturalness). It selects the optimal restoration branch (baseline, planner-guided, or generative enhancement) only if the quality improvement exceeds a specific threshold ( $\delta$ ), ensuring the original image remains available for audit.
Damage Recognition Agent (DRA): Converts the audit-preserving image set into structured outputs. It supports cross-view (RSI+SVI), bi-temporal (pre/post SVI), and single-view inputs. It outputs ordinal severity levels and object-level damage indicators (e.g., damaged buildings, fallen trees) with confidence scores.
Disaster Reasoning Agent (DReA): Maps structured recognition outputs to causal explanations and decision-support reports. It synthesizes perception states, damage signals, and image evidence to generate reports detailing damage mechanisms, affected assets, secondary risks, and recovery suggestions.

2.2 Zero-Shot and Reproducibility Strategy

RAPID operates in a zero-shot setting, meaning no target-dataset fine-tuning or few-shot examples are used. The system relies on fixed label spaces and JSON-format constraints in prompts to ensure consistency. Reproducibility is achieved through an audit trail that stores visual evidence, intermediate agent states, restoration deltas, and evaluation scores, allowing for the tracing of errors to specific pipeline stages.

3. Experimental Setup

The framework was evaluated across multiple disaster scenarios (hurricanes, floods, wildfires, earthquakes) using three distinct dataset types:

Dataset A: Cross-view pairs of remote sensing and street-view imagery (Hurricanes).
Dataset B: Bi-temporal street-view pairs (pre- and post-disaster) (Hurricanes).
Dataset C: Multi-hazard post-disaster street-view imagery (including drought, earthquake, flood, ice storm, and wildfire).

Evaluation metrics included precision, recall, F1-score, and overall accuracy for classification; Normalized Cross-Severity Error (NCSE) for ordered severity prediction; and automated/human rubric scoring (factual consistency, causal plausibility, completeness, action feasibility) for report generation.

4. Key Results

Disaster Perception (DPA): The framework achieved an overall accuracy of 0.92 for multi-disaster-type classification (using GPT-5-mini). While precision was consistently high (>0.90) across most hazards, recall varied, with earthquake classification being the most challenging due to visual ambiguity.
Image Restoration (IRA): No single restoration strategy dominated all scenarios. Planner-guided and baseline methods provided the most stable improvements for remote sensing and most street-view subsets. Generative enhancement (Gemini) improved quality for specific hurricane street-view scenes but introduced artifacts or degraded quality in wildfire and drought scenarios.
Damage Recognition (DRA):
- Severity Prediction: The system achieved up to 0.627 accuracy for cross-view damage severity prediction (Dataset A, Gemini-3-Pro) and 0.591 for bi-temporal analysis (Dataset B, ChatGPT-5.1). Errors primarily occurred between adjacent severity levels rather than extreme categories.
- Object Detection: For cross-view hurricane analysis, Gemini-3-Pro achieved an F1-score of 0.77 and an overall accuracy of 0.92 in detecting specific damage elements (e.g., downed power lines, damaged buildings).
Disaster Reasoning (DReA): Generated reports scored highly on factual consistency and causal plausibility (often >4.0/5.0). However, scores for action feasibility were lower (2.0–2.8), indicating that while the system can describe damage and causes effectively, proposing specific, operationally feasible recovery measures remains a challenge requiring expert context.

5. Significance and Claims

The paper claims that RAPID represents a shift from traditional pattern recognition toward reasoning-based autonomous understanding in disaster intelligence. Its primary contributions are:

Zero-Shot Generalization: It demonstrates that a multi-agent pipeline can perform damage assessment across diverse disaster types and data modalities without task-specific fine-tuning, addressing the scarcity of labeled data in time-sensitive scenarios.
Interpretability and Auditability: By decomposing the task into distinct agents, RAPID exposes intermediate states (perception, restoration, recognition), making the assessment process transparent and traceable, unlike single-model pipelines.
Cross-View Integration: It successfully leverages complementary information from satellite and street-level perspectives to improve damage recognition and reasoning.

The authors maintain a modest stance, explicitly stating that RAPID is an auditable early-stage assessment pipeline intended to support, not replace, official field inspections or expert judgment. They acknowledge limitations, including reliance on proprietary models, the need for better uncertainty quantification, and the current gap in generating operationally feasible recovery recommendations. The framework is positioned as a foundational step toward autonomous disaster intelligence that prioritizes evidence-grounded reasoning over black-box prediction.