论文信息
论文解读
AI 辅助整理- 方法
- 多智能体框架+LLM对话理解需求,跨模态检索文本与遥感影像线索。
- 发现
- 检察办案人员评测显示检索性能与可用性良好,简单易学。
- 机会点
- 可将该智能体范式迁移至农业面源污染、耕地破坏等线索的跨模态自动发现。
论文摘要
摘要 在生态保护红线区域内,不符合规定的生产建设活动所造成的生态环境破坏问题日益严重。现有方法通常利用遥感影像和深度学习技术自动识别违规建设,或从公众举报、官方通报等文本信息中检索生态环境破坏线索。前者往往依赖专业经验知识,后者则停留在被动检索层面;二者均依赖预定义规则和专业经验,且存在时间滞后。近年来兴起的大语言模型(Large Language Model,LLM)智能体兼具自然语言理解与多步任务执行能力,使得通过人机对话快速、灵活地获取并满足用户的检索需求成为可能。受此启发,本文基于多智能体框架构建了一种面向生态环境破坏的多模态线索检索智能体。该智能体通过大语言模型与缺乏计算机专业技术背景的用户进行自然语言交互,明确其对指定区域、时间段和破坏类型的检索需求,随后调度不同子智能体以跨模态方式自动检索相应时空范围内的文本线索和遥感影像线索,并将文本与图像证据汇总呈现给用户,以支持问题线索的审查与分析。经检察办案领域研究人员测试与评估,所提方法取得了令人满意的检索性能和可用性,且简单易学,可为生态环境破坏问题的筛查与评估提供技术支撑。
全文
Scientific ReportsARTICLE | Article in Press
https://doi.org/10.1038/s41598-026-80668-x
This Article in Press is shared early to give you faster access to new research. It is citable and carries a permanent DOI.
The final edited version will replace it automatically.
Scientific Reports
Scientific Reports
LLM-agent-based cross-modal retrieval method for
detecting clues of ecological environment damage
Man Zhao · Lu Bai · Shiyao Lu · Xiaoqing He · Liangfu Chen · Fucai Ben · Zhibao Wang
Received: 27 July 2026 / Accepted: 10 November 2026
© The Author(s) 2026
Abstract
Within Ecological Conservation Redline zones, ecological environmental damage caused by non-compliant
production and construction activities has become increasingly severe. Existing methods typically leverage
remote sensing imagery and deep learning techniques to automatically identify unauthorized construction, or
retrieve clues of ecological environmental damage from textual information such as public complaints and
official circulars. The former often relies on professional experiential knowledge, while the latter remains at
the level of passive retrieval; both rely on predefined rules and professional experience and suffer from a time
lag. Recently emerging Large Language Model (LLM)-based agents, which possess both natural language
understanding and multi-step task execution capabilities, make it possible to acquire and fulfill users’ retrieval
demands in a fast and flexible manner through human–machine dialogue. Inspired by this, this paper constructs
a multimodal clue retrieval agent for ecological environmental damage based on a multi-agent framework. The
agent interacts with users who lack computer professional and technical backgrounds through natural language
via LLMs to clarify their retrieval demands regarding specified regions, time periods, and damage types. It then
orchestrates different sub-agents to automatically retrieve textual clues and remote sensing image clues within
the corresponding spatiotemporal scope in a cross-modal manner, and aggregates and presents textual and
image evidence to users to support the review and analysis of problem clues. Testing and evaluation conducted
by researchers in the field of procuratorial case handling show that the proposed method achieves satisfactory
retrieval performance and usability, and is simple and easy to learn. It can provide technical support for the
screening and assessment of ecological environmental damage issues.
Keywords Environmental degradation · Multi-agent · Large language model
Introduction
The ecological environment constitutes the material foundation for human survival and development. However,
issues including environmental pollution, ecosystem degradation and unsustainable resource exploitation have
become increasingly pronounced1,2, rendering the protection of the ecological environment and natural resources
an imminent necessity. Many countries worldwide have implemented zoned ecological and environmental
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
management for ecosystem conservation 3,4, among which the Ecological Conservation Redlines (ECRs) repre -
sent a pivotal institutional innovation in China’s environmental protection regime. ECRs refer to spatial boundar-
ies and management thresholds where stringent protection measures are enforced to safeguard natural ecosystem
services, environmental quality security, and the rational exploitation of natural resources5.
In 2023, China’s Ministry of Natural Resources completed the nationwide delineation of ECRs, which cover
a total area of over 3.15 million square kilometers, and any non-compliant production and construction activities
are prohibited within these redline zones 6. Nevertheless, the supervision and regulation of such illegal activi -
ties face formidable challenges, including the vast coverage of redline areas, heavy workloads, a wide diversity
of violation types, and delayed detection of violations. For example, a wind farm project in Shoushan, Anhui
Province, illegally occupied 18.709 ha of the Bagong Mountain Scenic Area in 2015, and the violation was
rectified in 2019 through public interest litigation7. In another case, a company in Wenzhou, Zhejiang Province,
conducted unapproved aquaculture activities in the marine waters of Longwan District in 2021, and relevant
law enforcement authorities launched a case investigation into the violation in June 2024 8. Illegal logging and
artisanal mining prevail in the Colombian Amazon rainforest. Restricted by inaccessible terrain and insufficient
law-enforcement patrol personnel, most illegal land-use violations are verified one to three years after ecological
damage actually occurs9. Over the past decade, forests in Cameroon have been cleared for cocoa cultivation. Such
scattered and concealed deforestation activities escape routine field inspections and can only be verified with a
lag of two to four years10.
Today, remote sensing technology has emerged as a pivotal technical means for monitoring ecological environ-
mental damage, by virtue of its advantages in large-scale coverage and long-term continuous observation11,12. Tra-
ditional remote sensing-based methods primarily conduct environmental monitoring through image mapping13,14,
ecological index calculation15,16 and machine learning-based identification 17,18. These approaches underpin the
quantitative analysis of environmental damage, yet they often rely on manually designed features and empirical
rules, which renders the intelligent retrieval of environmental damage issues challenging to achieve. With the
advancement of artificial intelligence, deep learning based approaches have been widely applied to the analysis
of high-resolution remote sensing imagery 19–25, which has remarkably enhanced the automation level and thus
enabled the automatic identification of non-compliant construction and production facilities within ECRs. How-
ever, most of these methods remain in passive monitoring mode, merely capable of detecting “where changes
have occurred”; they lack the integration of cross-domain knowledge, as well as the ability to actively compre -
hend and discover the complex semantic implications of environmental damage. Furthermore, sole reliance on
image clues is often insufficient to fully reflect the nature of ecological environmental damage issues. In practical
applications, it is necessary to integrate textual information such as public opinions, news coverage and official
circulars to construct a complete and reliable clue chain for ecological environmental damage. Consequently,
how to extract textual and image clues from massive internet and remote sensing data has become a major chal -
lenge for the user-demand-driven cross-modal retrieval of ecological environmental damage clues.
The rise of large language models (LLMs) has brought new opportunities to address this challenge 26. LLMs
exhibit robust capabilities in semantic understanding and reasoning, enabling cross-modal information integra -
tion, complex problem analysis, and natural language explanation generation. These advantages not only make
up for the deficiencies of deep learning methods in semantic understanding and cross-domain reasoning 27 but
also drive the rapid development of frameworks combining LLMs with the Agent architecture 28. By integrating
mechanisms of environmental perception, task planning and tool invocation, the agent framework endows LLMs
with the capabilities of interactive operation and multi-step task execution, transforming them from pure natu -
ral language models into intelligent agent systems that can translate user demands into concrete retrieval tasks.
In this workflow, the steps performed autonomously by the agent include textual clue retrieval, remote sensing
image identification, and the automatic association of textual and image clues, whereas demand specification,
intermediate result screening, and final verification of the associated evidence are performed by the user.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Furthermore, the multi-agent system 29,30 refers to a system design that enables complex task completion
through the collaboration of multiple intelligent agents. Individual sub-agents can grasp user requirements more
accurately and provide personalized services via continuous user interaction 31. The intelligent applications in
the ecological environmental protection field built on the multi-agent framework allow non-professional users
to directly invoke corresponding tools to solve complex tasks based on prompt instructions. Mainstream multi-
agent frameworks at present include AutoGen31, LangGraph32, AgentScope33, and Coze Multi-Agent34.
This LLM-based agent framework pioneers new research directions for monitoring tasks in the ecological
environmental domain 35. However, the current application of LLM-based agent systems to the cross-modal
retrieval of ecological environmental damage remains in the initial exploratory stage. Most existing studies focus
on tasks such as disaster response and decision-making36,37.
In summary, the main contributions of this paper are as follows. First, relying on AI Agent technology, this
study proposes an LLM-agent-based cross-modal retrieval scheme for detecting clues of ecological environmen-
tal damage. Second, a multimodal retrieval agent for ecological damage is constructed. The agent leverages the
natural language comprehension and reasoning capabilities of large language models to acquire user retrieval
demands through dialogue and to retrieve textual clues associated with ecological damage, invokes Remote -
CLIP38 to automatically identify image evidence of damage, and automatically associates textual and image clues
to provide auxiliary evidence supporting user verification, thereby realizing the user-demand-driven cross-modal
retrieval of ecological environmental damage clues. We note that, from a software-architecture perspective, the
Dispatch, Search and Image agents communicate through a shared JSON object in Shared Memory and are routed
by static if-else transitions; this is a Coze-based static workflow orchestration, chosen for maintainability and for
ease of adding new damage types. The LLM-agent architecture serves as the enabling framework through which
this retrieval flow is implemented; the novel contribution of this work lies in the user-demand-driven, human-
in-the-loop cross-modal retrieval process itself, rather than in a newly proposed agent-organization mechanism.
Related works
The emergence of LLMs such as GPT-4 and DeepSeek 39 has advanced natural language processing. Building
on the established concept that intelligent agents are computer systems capable of performing tasks and mak -
ing decisions with minimal or no human intervention 40, this progress has accelerated the research, development
and application of agents, leading to the rapid development of Autonomous AI agents41,42. These LLMs possess
human brain-like reasoning capabilities and can handle natural language tasks in different application scenarios.
However, LLMs tend to generate content without factual basis or inconsistent with reality, which leads to hallu-
cinations when processing complex or ambiguous problems. They may also suffer from information forgetting or
fail to accurately associate contextual information when dealing with tasks with complex contexts. Furthermore,
trained on large-scale general datasets, LLMs need further adjustment and optimization to meet the demands of
specific domains in particular application scenarios.
By integrating agents with technologies such as Retrieval-Augmented Generation (RAG) and prompt engi -
neering, domain-specific knowledge can be supplied for LLMs, which addresses the problems of model hal -
lucinations and knowledge limitations of LLMs. Thus, such integrated frameworks are increasingly adopted as
the decision-making core of autonomous agents 43. Recent studies demonstrate the effectiveness of LLM-based
agent systems across different domains. For example, GMAC 43 leverages GPT to extract semantic informa -
tion and assists agents in generating action predictions for the next step, thereby improving the communication
efficiency of multi-agent systems. MetaGPT 44 effectively decomposes complex tasks by deploying different AI
agents to handle distinct components of a problem. The KwaiAgents system 45 takes an LLM as its cognitive
core and integrates a memory bank and time-aware search tools to efficiently fulfill user queries. The RS-Agent
system46 is composed of three parts: a Task Agent, a Prompt Agent, and a LoRA Agent, which leverages the natu-
ral language processing capabilities of LLMs to generate high-quality remote sensing imagery. The LLM-Geo
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
prototype system47 employs an LLM as its core reasoning mechanism to explore the autonomous solution of
GIS problems. The RescueADI system integrates LLMs with remote sensing image interpretation to assist in
disaster damage identification. These studies show that agent systems integrated with LLMs have demonstrated
tremendous potential across various fields 48. The semantic understanding capabilities of LLMs, combined with
the planning and tool invocation capabilities of agents, enable the transformation from passive task response to
interactive task execution.
More recently, with the advancement of LLMs and intelligent agent technologies 49, researchers have begun
to explore their application potential in the ecological environmental domain. Current research efforts primarily
focus on scenarios such as climate management and natural disaster response. For example, DisasterRespon -
seGPT49 leverages LLMs to assist users in formulating disaster response action plans, thus improving the speed
of manual emergency decision-making. ChatClimate 50 enables conversational interaction with users and inte -
grates domain knowledge of climate science to enhance the accuracy of LLMs in the climate field. To further
reduce model hallucinations, ClimateGPT51 has optimized the retrieval function of the model and can respond to
user queries with answers from diverse research perspectives. In the field of flood disasters, Rafaela Martelo et
al.52 proposed an intelligent assistant system driven by GPT-4, which leverages the natural language processing
capabilities of GPT-4 to retrieve flood alerts and integrate real-time early warnings with other data by respond -
ing to inquiries, thus supporting the formulation of emergency decision-making strategies. In the field of wildfire
disasters, Xie et al.53 proposed a Retrieval-Augmented Generation-based multi-agent LLM system that can assist
humans in conducting comprehensive analysis and making decisions in wildfire disaster scenarios. To achieve
a comprehensive analysis of disaster scenarios, Liu et al. 54 proposed RescueADI, which adopts an LLM-driven
adaptive agent to interpret remote sensing images of disaster scenes, a method that can address various com -
plex interpretation requests. Furthermore, in the field of marine environmental damage, Yang et al.55 designed a
conversational artificial intelligence system named Chatecology, which uses a domain fine-tuned LLM to inves-
tigate the causes of large-scale marine green tides and thus formulate targeted strategies for marine ecological
management.
In summary, existing studies indicate that the integration of LLMs and agents holds broad application pros -
pects in the ecological environmental domain. However, these approaches still have limitations in aspects such
as personalization and data utilization. Their interaction with users mainly focuses on addressing one-off prob -
lems, where agents reason based on user queries and return results. In practical applications, non-professional
users often require multi-turn interactions to accurately convey their actual needs, as they can hardly articulate
their problems clearly in one go. This may lead to misunderstandings or oversimplifications when agents handle
complex tasks. Furthermore, research on methods directly targeting remote sensing images for the agent-assisted
cross-modal retrieval of ecological environmental damage remains quite limited.
Different from existing remote sensing agents such as RS-Agent, LLM-Geo, and RescueADI, which typically
execute tasks in a largely automated manner upon receiving a single user instruction, our framework supports
multi-round human–agent interaction throughout the entire retrieval workflow. The proposed system contains
two specialized sub-agents: a textual retrieval agent and an image retrieval agent. It enables bidirectional agent-
assisted retrieval: retrieving corresponding remote sensing image evidence based on user-specified textual case
clues, or retrieving relevant textual case records from web resources according to the uploaded or selected remote
sensing imagery. Moreover, the agent assists users in automatically associating textual clues with image clues,
allowing users to further verify the associated heterogeneous evidence. Instead of a single autonomous pipeline
without intermediate human screening, the multi-round interactive mechanism is embedded into the cross-modal
retrieval loop to support ecological damage investigation.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Methodology
All methods were performed in accordance with the relevant guidelines and regulations. This study includes a
questionnaire-based user evaluation that was reviewed and approved by the Science and Technology Ethics Com-
mittee of Qiqihar University, with written informed consent obtained from all participants.
To address limitations including inadequate agent-assisted retrieval capability and limited clue sources in
existing ecological damage investigation, this paper constructs an agent-driven cross-modal retrieval framework
to break the constraints of single-data-source detection. The overall architecture of the proposed framework
is designed to integrate multi-source information for the retrieval and assisted comparison of ecological envi -
ronmental damage clues. It combines textual data from online sources with remote sensing imagery to enable
cross-modal analysis. The framework operates in a closed-loop manner, allowing information extracted from one
modality to guide the retrieval and analysis of the other. As shown in Fig. 1, the proposed framework can retrieve
Fig. 1 Flowchart of the proposed cross-modal retrieval framework for ecological environment damage.Icons in this and
subsequent figures (Figs. 1, 2, 3, 4, 5, 6, 7 and 8) are from Font Awesome Free ( h t t p s : / / f o n t a w e s o m e . c o m / ) , licensed under
CC BY 4.0.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
textual and remote sensing image clues of ecological environmental damage within a user-specified geographical
area and time period, and support the assisted comparison and user verification of clues across the two modalities.
The specific implementation of the framework encompasses two core functions: “Case-to-image Retrieval” and
“Image-to-case Retrieval.”
The “Case-to-image Retrieval” function first retrieves textual information related to ecological environmental
damage, including public sentiment information, news reports, and official bulletins, based on user queries. After
analysis and collation, the results are fed back to the user. Subsequently, based on the textual information selected
by the user, the system further retrieves remote sensing imagery within the corresponding spatiotemporal scope.
Finally, the system automatically associates the retrieved image evidence with the corresponding textual clues
and presents them to the user, assisting the user in verifying the authenticity and accuracy of the textual content.
The “Image-to-case Retrieval” function starts by searching for production-related facilities in remote sensing
imagery of the target area in line with user needs. The retrieved imagery results associated with ecological envi-
ronmental damage are then provided to the user. Combined with the user’s points of interest, the system conducts
further searches for relevant textual information on the internet, collates it, and feeds it back to the user. This
provides detailed textual evidence for users to assess ecological environmental damage issues, thereby achieving
in-depth integration of image clues and textual documentation.
To clarify the autonomy boundary of the proposed framework, the steps performed autonomously by the agent
include: (1) cross-modal retrieval of textual clues from online sources; (2) automatic identification of candidate
image evidence from remote sensing imagery via RemoteCLIP; (3) automatic association of textual and image
clues, providing auxiliary evidence for user verification; and (4) aggregation and presentation of evidence. By
contrast, user involvement is required at the following decision points: specifying retrieval demands (geographi-
cal area, time period, and damage type), selecting relevant cases or images from intermediate retrieval results,
verifying the authenticity and accuracy of the associated textual and image evidence, and providing feedback
to trigger iterative refinement of the retrieval. In this regard, the proposed framework is not an autonomous dis -
covery system but an automated cross-modal retrieval tool with a multi-round human-in-the-loop mechanism.
Unlike conventional interactive retrieval systems, in which users must manually query textual and image sources
separately and integrate heterogeneous evidence on their own, the proposed framework automates the labor-
intensive steps of cross-source evidence retrieval and association, while the verification of associated evidence
is performed by the user with the assistance of the framework, and domain experts retain control over screening
and decision-making.
As shown in Fig. 2, the proposed ecological environmental damage retrieval process primarily consists of three
components: text-based ecological environmental damage retrieval, image-based ecological environmental dam-
age retrieval, and user requirement analysis. Consequently, the model adopts a multi-agent architecture, which is
composed of one Dispatch Agent and two Mini Agents. Its operational mechanism is detailed as follows.
The Dispatch Agent is primarily responsible for receiving user task requests. It leverages the natural language
processing capabilities of LLMs to understand user needs, converts them into standard JSON text, and allocates
and schedules tasks to the two Mini Agents based on the task requirements. The Search Agent, acting as the
text processing unit, retrieves text-based issues related to ecological environmental damage according to the
JSON text information output by the Dispatch Agent. Through analysis and reasoning, it generates correspond -
ing textual clues. The Image Agent serves as the image processing unit, invoking and executing modules for
remote sensing image acquisition, preprocessing, and retrieval based on information such as the spatiotemporal
scope and task objectives in the JSON text, thereby completing the retrieval of clues for image-based ecological
environmental damage. Communication among the three agents is realized through Shared Memory. This design
is positioned as an engineering implementation choice rather than a methodological novelty. The “agents” are
implemented as Coze workflow nodes: the Coze platform handles LLM invocation, plugin execution, conversa-
tional dialogue management and Shared Memory storage, whereas the authors’ custom logic lies in demand pars-
ing, bidirectional cross-modal association, the relevance filter, and the multi-round human-in-the-loop screening
loop.Furthermore, to enhance the human-computer interaction experience, all three agents are integrated with
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
conversational interaction functions, enabling them to receive and understand user questions or instructions in
real-time and generate corresponding responses to feed back to the user.
Dispatch agent
The invocation and workflow of the Dispatch Agent is illustrated in Fig. 3. The Dispatch Agent provides users
with two optional retrieval services: case-to-image retrieval and image-to-case retrieval. Users submit retrieval
requirements via natural language interaction with the Dispatch Agent. Leveraging large language models (LLMs),
the Dispatch Agent interprets user intentions and extracts core retrieval parameters, such as target region, time
span, types of ecological damage activities and available clue forms. These parsed parameters are stored in JSON
format within the system’s shared memory for data exchange with the Search Agent and Image Agent. Subse -
quently, the Dispatch Agent invokes the corresponding Mini Agent to fulfill designated retrieval tasks.
The Dispatch Agent leverages LLMs to interpret the user’s natural language input and outputs a standard-for-
mat JSON text to Shared Memory. To enable the Dispatch Agent to better understand and process user requests,
an Agent Prompt is configured to improve the Dispatch Agent’s performance. This Prompt defines the Dispatch
Agent’s role, skills, constraints, output text format, and task assignment rules, as detailed in Table 1 below.
Fig. 2 Framework of the proposed multi-agent retrieval model for ecological environment damage.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Search agent
As shown in Fig. 4, the Search Agent parses the JSON data from Shared Memory to obtain the user-specified
retrieval area, time period, requirement type, and retrieval demand. It then interacts with the user to confirm the
number of web clues to be returned. The search tool adopted is the Bocha Web Search tool 56, an answer engine
based on large language models and real-time search technology that provides real-time information and reliable
reference sources; the system invokes Bocha’s web search API to retrieve, from the internet, online information
related to the specified type of ecological environmental damage, such as public sentiment information, news
reports, and official circulars. The main input parameters of this tool include the number of returned results
(count), the search query (query), and the page number, while the output parameters include web page links and
web page snippets. After obtaining the retrieval results, the agent uses LLMs to analyze and organize the textual
Fig. 3 Invocation and task assignment workflow of the Dispatch Agent.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
content corresponding to the user-specified top-n most relevant search results, generates a brief report in a fixed
format, and presents it to the user. The agent then interacts with the user to ask whether the current retrieval
results are satisfactory. If the user is not satisfied, the agent re-executes the retrieval with adjusted parameters. If
the user is satisfied, the user specifies the relevant textual clues, and the agent updates the JSON data in Shared
Memory according to the specified textual clues. If a returned webpage becomes inaccessible during the user’s
Table 1 Prompt for the dispatch agent.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
review (e.g., a 404 error or removal), the agent marks the entry as invalid, supplements the search with adjusted
keywords, keeps the total number of entries unchanged, and avoids repeating previously provided links in sub -
sequent rounds.
If the user’s requirement in Shared Memory is “Case-to-image Retrieval,” the Search Agent updates the cor -
responding area and time information in Shared Memory based on the text content selected by the user and then
switches to the Image Agent node. If the user’s requirement in Shared Memory is “Image-to-case Retrieval,” the
task is completed and the system switches back to the Dispatch Agent node. Relevant parts of the Search Agent
Prompt are shown in Table 2 below.
The reverse geocoding tool {{geocode_to_poi}} encapsulates the AMap (Gaode) reverse geocoding service.
The input parameters of regeoAddress are the longitude and latitude in the WGS84 coordinate system. The
interface internally performs coordinate system conversion and then invokes AMap’s reverse geocoding inter -
face ( https://restapi.amap.com/v3/geocode/regeo)to obtain the corresponding administrative information. The
returned province, city, and other relevant information are then encapsulated and output as a JSON object. The
{{WebSearch}} tool encapsulates Bocha’s real-time web search service, invoking the Bocha search interface
(https://api.bocha.cn/v1/web-search) to obtain the corresponding retrieval results.
Image agent
The Image Agent takes Shared Memory as input and parses the JSON file through code to obtain the request,
retrieval area, retrieval time period, and retrieval type, as shown in Fig. 5. Subsequently, it guides the user
through interaction to submit the corresponding remote sensing images and redline data. The image processing
and retrieval model is invoked via API, and the retrieved remote sensing images are fed back and displayed to
the user.
If the user’s requirement in Shared Memory is “Image-to-case Retrieval,” the Image Agent modifies the JSON
file in Shared Memory based on the metadata corresponding to the image selected by the user, adds a coordinate
field, and then switches to the Search Agent node. If the user’s requirement in Shared Memory is “Case-to-image
Retrieval,” the task is completed and the system switches back to the Dispatch Agent.
As shown in Fig. 6, the image processing tool is developed based on Python and GDAL. The tool reads the red-
line boundary SHP file and remote sensing image uploaded by users, performs image projection transformation
and vector clipping, and splits the clipped image into JPG tiles with a uniform size of 256 × 256 pixels. For each
tile, the tool calculates the longitude and latitude of the top-left and bottom-right corners, determines whether
Fig. 4 Invocation workflow of the search agent.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Table 2 Part of the prompts used by the Search Agent.
Fig. 5 Workflow of the image agent.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
the tile falls within the retrieval range recorded in Shared Memory, and retains only tiles inside the user-specified
area. Meanwhile, tile name, image size, geographic coordinates and other information are saved into the metadata
file.
The retrieval area defined by the user is large, while the spatial coverage of the uploaded remote sensing image
is relatively small. For example, in the “image-to-case retrieval” task, the user initially sets Anhui Province as the
retrieval area and photovoltaic facilities as the retrieval target, but the uploaded image only covers Huainan City.
If text retrieval is executed directly, the retrieval results will contain a large number of irrelevant clues outside
Huainan City, reducing the matching precision between image clues and text clues. To solve this problem, after
image retrieval, the system updates Shared Memory according to the coordinates of the tile selected by the user
and writes coordinate information to narrow the retrieval scope of the Search Agent. Considering that regional
descriptions of online text clues are mostly based on city-level administrative divisions, only the top-left longi -
tude and latitude of image clues need to be recorded to meet the retrieval requirements.
To clarify the chain from image coordinates to textual retrieval, after the user selects a clue tile, the Image
Agent writes the tile coordinates into Shared Memory. When the Search Agent receives the task, it parses the
coordinate field and invokes the reverse geocoding tool ({{regeoAddress}}), which converts the WGS84 coor -
dinates into a city-level administrative region (e.g., “Huainan City, Anhui Province”). The search query is then
constructed by combining the place name, the damage type, and the time field specified in the retrieval request,
together with violation-related keywords (e.g., “Huainan City + photovoltaic facilities + 2025 + violation”). After
retrieval, a relevance filter is applied: only text clues whose reported location falls within the target city are
retained, while clues from other regions, which would otherwise reduce the matching precision between image
and text clues, are excluded. The uploaded remote sensing imagery is expected to fall within the time range speci-
fied by the user, so that the image and textual evidence refer to the same period.
Regarding the choice of the upper-left corner coordinate rather than the image center, the geometric centroid,
or the bounding box center of detected targets: because the administrative descriptions in online text clues are
Fig. 7 Workflow of the image retrieval.
Fig. 6 Workflow of the image processing.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
almost all at the city level, and a city-level region typically spans a large area, any reference point within a
selected tile (its upper-left corner, center, or centroid) is reverse-geocoded to the same city in practice. The upper-
left corner is adopted as the reference point because it is a uniformly recorded and reproducible anchor in the tile
metadata, consistent with the raster metadata convention. It should be noted that this choice was not validated
through a dedicated quantitative experiment; instead, the rationale rests on the spatial scale of the system. All
remote sensing images in our dataset have a spatial resolution finer than 10 m and are tiled at a uniform size of
256 × 256 pixels, so a single tile covers a ground extent of at most approximately 2.56 km per side, and the maxi-
mum offset among the candidate reference points (upper-left corner, center, or centroid) within a tile is therefore
on the order of a few kilometers at most. Because the reverse geocoding step converts coordinates to a city-level
administrative region, which typically spans tens of kilometers or more, the difference among the candidate refer-
ence points cannot alter the resulting city-level region in practice. Consequently, the reference-point choice does
not affect the retrieval scope at the city level, and a dedicated experiment was not considered necessary.
The feature extraction tool adopts a LoRA-fine-tuned RemoteCLIP-ViT-B-32 for batch feature extraction of
images, and stores the extracted features into the feature vector file. Afterwards, information including the spatial
region and time of image clues, the corresponding metadata files, and the feature vector files is loaded into the
database.
The image retrieval tool invokes the LoRA-fine-tuned RemoteCLIP-ViT-B-32 model and performs image
retrieval according to the ecological environment damage type specified by the user in Shared Memory. To
improve the accuracy of image retrieval, a prompt list is formulated for each case type. Similarity metrics are then
calculated between the corresponding prompts and all images within the region. All images are sorted by metric
distance, and the top n images with the highest similarity specified by the user are returned; meanwhile, the cor-
responding metadata is queried by image name and returned together with the images for the user’s final screen-
ing and interpretation. Considering that ground objects in remote sensing scenes are prone to confusion, domain
experts can identify false positive samples from the sorted candidate list. This approach takes vector similarity as
a quantitative reference while reserving the final decision-making authority for interpreters.
RemoteCLIP is a vision-language foundation model pre-trained specifically for remote sensing imagery. It
inherits the dual-encoder architecture of CLIP, in which a visual encoder and a text encoder map images and
texts into a shared embedding space where cosine similarity measures semantic relevance. Figure 8 illustrates
the overall framework. To improve retrieval accuracy for five types of environmentally harmful infrastructure in
remote sensing images, we constructed an infrastructure image dataset and designed prompt templates that build
image-text pairs. For each category, a positive prompt was defined to describe its typical visual content. Three
fine-tuning strategies were compared against the zero-shot baseline: full fine-tuning, which updates all param -
eters of both the visual and text encoders; vision-only fine-tuning, which freezes the text encoder and updates
only the visual encoder and the temperature parameter, thereby isolating the effect of visual adaptation; and low-
rank adaptation (LoRA), which freezes the pre-trained weight matrices and injects low-rank residuals into the
query/key/value projections and the MLP layers of both encoders. Based on the comparative experiments and the
analysis of results, the LoRA-tuned RemoteCLIP was finally adopted as the image retrieval model.
Case study
The Coze platform 34 is adopted to construct an AI system for the assisted retrieval of ecological environmen -
tal damage clues, with the LLM invoked being doubao-seed-2-0-mini-260,428.The retrieval experiments were
executed in August 2026.Relevant case studies are presented as follows.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Case study 1: case-to-image retrieval – querying for clues on ecological damage caused by
photovoltaic panels in anhui province
This case study primarily involves querying cases of ecological damage caused by photovoltaic panels in Anhui
Province in recent years. Based on the case selected by the user, it retrieves the corresponding remote sensing
images to verify the existence of illegal photovoltaic constructions, thereby providing image evidence for the
case. The user’s main requirements are shown in Table 3 below.
According to the user’s requirements, the system first retrieves web pages related to the illegal construction of
photovoltaic projects in Anhui Province from 2020 to the present, returns the three most relevant web pages as
specified by the user, and presents them for the user’s review and selection of textual clues.
The translation of the retrieval results in Fig. 9 is presented as follows:
Table 3 User’s main requirements.
Fig. 8 Overview of the Proposed LoRA Fine-Tuning Framework for RemoteCLIP.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
[1] Time: Illegal construction began in 2015; exposed in 2022; demolition completed in November 2023;
reported by the Ministry of Water Resources in October 2024.
Location: Maoji District, Huainan City, and Yingshang County, Fuyang City, Anhui Province (Jiaogang Lake).
Summary: Three enterprises—Huainan Jinhui Photovoltaic, Huainan Zhongdian Jiaogang Lake Photovoltaic
Power Generation, and Yingshang Juan Photovoltaic—illegally occupied more than 3,400 mu of the waters of
Jiaogang Lake to build water-surface photovoltaic power stations since 2015, and part of the projects encroached
on the Jiaogang Lake Qian Shi (Euryale ferox) National Aquatic Germplasm Resources Conservation Zone. In
April 2023, Maoji District issued a demolition notice with a time limit, and Yingshang County formulated a
demolition plan; by November 2023, all photovoltaic panels and components had been completely removed.
The Ministry of Water Resources listed the case as a typical case of the “clearing four disorders” (Qing Si Luan)
campaign in rivers, lakes and reservoirs.
Link: https://www.kesolar.com/headline/269697.html.
[2] Time: Reported by the Ministry of Water Resources in October 2024 (involving A VIC and State Power
Investment Corporation backgrounds).
Location: Jiaogang Lake, Huainan City, Anhui Province.
Summary: The official WeChat account of the Ministry of Water Resources, “China Water Resources,” pub -
lished ten typical cases of the “Qing Si Luan” campaign in rivers, lakes and reservoirs, among which the case of
illegal photovoltaic projects encroaching on Jiaogang Lake in Anhui Province was included. In this case, three
photovoltaic power generation enterprises long-term illegally occupied lake water surfaces and a national aquatic
Fig. 9 Textual clues.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
germplasm resources conservation zone to construct photovoltaic projects, involving corporate backgrounds such
as A VIC and State Power Investment Corporation; they were required to demolish the illegal structures and
restore the aquatic ecology.
Link: https://www.nengapp.com/zixun/8523a9a056787324873d115236f18e7b.
[3] Time: October 30, 2024.
Location: Anhui Province (Ministry of Water Resources policy targeting the Jiaogang Lake case).
Summary: In light of typical cases including the illegal photovoltaic construction at Jiaogang Lake in Anhui,
the Ministry of Water Resources explicitly stated that photovoltaic power stations that impede flood discharge are
prohibited within the management scope of river courses. The notice pointed out the illegal nature of photovoltaic
power generation projects built on lake surfaces that occupy lake waters and affect flood discharge and aquatic
ecology, requiring all localities to comprehensively investigate and rectify illegal photovoltaic projects within the
management scope of rivers and lakes, and to strengthen routine supervision.
Link: https://m.nengapp.com/zixun/8b4292315ab3555a47a6d567e90df754.
The illegal photovoltaic construction case at Jiaogang Lake is the most typical case in Anhui Province and has
been widely reported on the web; the three returned web clues are all related to Jiaogang Lake but differ in report-
ing time. Based on the textual clues, the user uploaded the remote sensing image and redline data of this region.
The system processed the image and split it into multiple 256 × 256 tiles, and retrieved the 10 most likely illegal
photovoltaic images according to the user’s requirements. The retrieval results were presented on an interactive
result page. A Sentinel-2 L2A true-colour composite image (10 m resolution, composited from the B04/B03/B02
bands of the Copernicus Sentinel-2 archive) covering exactly the same spatial extent as the uploaded remote sens-
ing image was used as the default basemap (Contains modified Copernicus Sentinel data [2022]); on this base -
map, the spatial extent of each retrieved image clue was outlined with a red rectangle according to its geographic
coordinates. Below the map, the image link and the metadata (the corner coordinates of each tile) of every clue
were listed. A “User Imagery” button allowed the user to switch the basemap to the originally uploaded remote
sensing image, which covers the same spatial extent, with the red rectangles of the image clues overlaid on it,
so that the suspected targets could be examined against the actual imagery; clicking an image link opened the
corresponding original remote sensing image for further verification. This helped the user further localize the
suspected targets. The image retrieval results are shown in Fig. 10.
The basemap is a Sentinel-2 L2A true-colour composite image (10 m resolution, B04/B03/B02 bands, Con -
tains modified Copernicus Sentinel data [2022]) covering exactly the same spatial extent as the uploaded remote
sensing image. The result page was generated by a custom Python script (Python 3, https://www.python.org/)
with rasterio (https://rasterio.readthedocs.io/) and rendered with Leaflet 1.9.4 (https://leafletjs.com/); red r e c t a n g
l e s outline the spatial extents of the retrieved image clues according to their geographic coordinates.
Case study 2: image-to-case retrieval – querying for clues on ecological damage caused by coastal
aquaculture in lingao county, hainan province
This task focuses on investigating whether illegal coastal aquaculture activities have occurred within the Ecologi-
cal Conservation Redlines of Lingao County, Hainan Province, since 2024. By retrieving imagery of illegal aqua-
culture facilities from remote sensing images, and then retrieving corresponding textual clues from the internet,
it provides clue evidence for ecological damage caused by coastal aquaculture. The user’s main requirements are
shown in Table 4.
The system processes the remote sensing image and redline data provided by the user and splits them into
multiple 256 × 256 tiles. According to the user’s requirements, it retrieves the 10 most likely images of illegal sea-
enclosure aquaculture facilities. To facilitate the user’s further examination of the image clues, the system marks
the suspected facilities with red rectangles within the retrieval region and feeds back each candidate image clue
of illegal sea-enclosure aquaculture together with its metadata to the user, as shown in Fig. 11.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
The basemap is a Sentinel-2 L2A true-colour composite image (10 m resolution, B04/B03/B02 bands, Con -
tains modified Copernicus Sentinel data [2024]) covering exactly the same spatial extent as the uploaded remote
sensing image. The result page was generated by a custom Python script (Python 3, https://www.python.org/)
with rasterio (https://rasterio.readthedocs.io/) and rendered with Leaflet 1.9.4 (https://leafletjs.com/); red r e c t a n g
l e s outline the spatial extents of the retrieved image clues according to their geographic coordinates.
After the user reviews and analyzes the image clues and selects one image clue, the system determines the
retrieval region according to the spatial coordinates of the selected image and automatically retrieves relevant
textual information from the internet. The retrieved information is then summarized and organized before being
fed back to the user for comprehensive review and analysis, as shown in Fig. 12.
The translation of the retrieval results in Fig. 12 is presented as follows:
[1] Time: Reported on May 31, 2024 (penalty dated April 28, 2024).
Location: Daoliao Village, Xinying Village Committee, Xinying Town, Lingao County, Hainan Province.
Summary: Lingao Yongheng Deep-Sea Aquaculture Co., Ltd. constructed five deep-water cage aquaculture
projects without obtaining the Sea Area Use Right Certificate, starting from July 2018; from February 2021
onward, it additionally took over 15 cages owned by three companies and continued illegal aquaculture until it
Table 4 User’s main requirements.
Fig. 10 Image clues retrieved for Case 1.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Fig. 12 Summary feedback on
retrieved information.
Fig. 11 Aquaculture pontoon imagery.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
was investigated and penalized in February 2024, occupying the sea area illegally for a total of five years and
seven months. The Comprehensive Administrative Law Enforcement Bureau of Lingao County ordered the com-
pany to return the sea area, restore the original state, and imposed a fine of RMB 145,400.
Link:http://www.hndnews.com/p/676099.html.
[2] Time: Reported on April 19, 2024 (penalty dated April 12, 2024).
Location: Touzui Village and Touzui Village Wharf, Xinying Town, Lingao County, Hainan Province.
Summary: The Comprehensive Administrative Law Enforcement Bureau of Lingao County conducted a cen-
tralized investigation and penalty against four companies—Haifeng Fishery, Haisheng Aquatic Products, Haixun
Deep-Sea, and Shunfeng Deep-Sea—which either occupied sea areas without obtaining the right to use sea areas
for deep-water cage aquaculture, or continued aquaculture after the expiration of their Sea Area Use Right Cer -
tificates, with the illegal duration ranging from three to five years. The four companies were fined a total of RMB
1,464,500 and were ordered to return the illegally occupied sea areas and restore the original state of the sea areas.
Link:https://www.hndnews.com/p/671680.html.
3 Time: Reported on November 15, 2025.
Summary: The case of illegal fishing of aquatic products prosecuted by the Lingao County Procuratorate
against defendant Li was heard in an open court session at Xinying Court. Driven by illegal profit, the defendant
used a trawl net failing to meet the mesh size requirements to illegally fish within the no-fishing line, constituting
the crime of illegal fishing of aquatic products, and pleaded guilty and accepted punishment in court. The case
was used, in an “on-site case-handling” manner, to conduct publicity on marine ecological protection laws among
fishermen; although it is not a sea-enclosure aquaculture case, it likewise constitutes an act that damages marine
fishery resources.
Location: Xinying Town, Lingao County, Hainan Province.
Link: https://hndnews.com/p/730170.html.
The temporal consistency between the image and textual evidence is enforced as described in Sect. 3.3: the
user specified the retrieval time range as 2024 to present through interaction with the system, and uploaded
remote sensing imagery of Lingao County acquired in 2024 (which falls within the user-specified time window);
the Search Agent accordingly constrains the text-search time window to 2024 to present, retaining only textual
clues published within this window so that the image and textual evidence refer to the same period. All three
returned textual clues (April 2024, May 2024, and November 2025) fall within this window.
Experiments
Dataset
We constructed a remote sensing image dataset covering five types of target objects associated with potential eco-
logical and environmental damage, and used it to fine-tune the RemoteCLIP visual-language model. All satellite
imagery was obtained from the Esri World Imagery Wayback service ( https://livingatlas.arcgis.com/wayback/),
downloaded in January 2025.The images were cut into non-overlapping tiles of 256 × 256 pixels, with spatial
resolutions ranging from 0.6 m to 6 m. Each tile was manually labeled into one of five categories: industrial stor-
age tanks, wind power facilities, photovoltaic panels, open-pit coal mines, and coastal aquaculture. The dataset
was randomly split into training, validation, and test sets at a ratio of 6:2:2, and the tile count per category is
summarized in Table 5.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Experimental environment
Training uses AdamW with weight decay 0.2, a warmup ratio of 0.1 followed by a cosine learning rate schedule,
a batch size of 16, and a learning rate of 1 × 10⁻⁶ for full and vision-only fine-tuning and 1 × 10⁻⁴ for low-rank
adaptation. The validation metric used for early stopping is the mean recall at rank 10 (mR@10) with a patience
of 5 epochs; the maximum number of epochs is 20. All experiments run on a single NVIDIA RTX 4060 Laptop
GPU with 8 GB memory, using PyTorch and the open_clip library. The experimental environment is summarized
in Table 6.
Results and analysis
Table 7 reports the comparison of the four methods. All three fine-tuning strategies substantially improve retrieval
over the zero-shot baseline: mean R@10 increases from 0.065 to 0.123 (+ 89%), mean R@20 from 0.123 to 0.247
(+ 101%), and mean R@30 from 0.168 to 0.370 (+ 120%), using the best-performing strategy. MRR rises from
0.803 to 1.000, indicating that every category’s top-ranked result is now a correct member after fine-tuning. The
zero-shot baseline fails completely on coastal aquaculture (R@30 = 0), a failure that all fine-tuning strategies
repair. The parameter efficiency of low-rank adaptation is notable: with only 1.30% of the model parameters
trainable and the shortest training time (155 s), low-rank adaptation achieves the highest R@20 (0.247) and
R@30 (0.370) among all methods, marginally surpassing full fine-tuning (0.243 and 0.367).
Table 8 reports the per-category recall of the zero-shot baseline and the low-rank adaptation model, the best-
performing fine-tuned model. The contrast is most pronounced for coastal aquaculture: the zero-shot model
retrieves no coastal aquaculture image in the top 30 for any rank threshold, whereas low-rank adaptation lifts
R@30 to 50.0%. Among the categories with a non-zero zero-shot baseline, Industrial storage tanks show the larg-
est absolute gain in R@30 (28.8% to 57.7%), followed by photovoltaic panels (25.0% to 37.5%) and coal mines
(16.9% to 24.2%). Wind power facilities improve more modestly (13.0% to 15.6%).
Table 7 Ablation results on the test split.
Method Trainable params Ratio R@10 R@20 R@30 MRR Time (s)
RemoteCLIP zero-shot 0 0.00% 0.065 0.123 0.168 0.803 —
Full fine-tuning (vision+text) 151,277,313 100.00% 0.122 0.243 0.367 1.000 165
Vision-only (text frozen) 87,849,217 58.07% 0.123 0.245 0.365 1.000 169
Low-rank adaptation (vision+text) 1,966,081 1.30% 0.123 0.247 0.370 1.000 155
Item Content
CPU 13th Gen Intel(R) Core(TM) i7-13620 H
GPU NVIDIA GeForce RTX 4060 (8 GB)
Memory 16 GB
CUDA 13.1
OS Windows 11
Table 6 Experimental
environment.
Type Total Train Validation Test
Coal mine 628 377 126 125
Coastal aquaculture 312 187 62 63
Industrial storage tanks 260 156 52 52
Photovoltaic panels 400 240 80 80
Wind power facilities 972 583 194 195
Total 2572 1543 514 515
Table 5 Dataset summary.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Qualitative evaluation via domain expert feedback
Evaluation methodology
User feedback serves as a critical dimension for evaluating the performance of the AI agent. To evaluate the
performance of the proposed intelligent assisted retrieval system for ecological environmental damage issues
(IASRE), we invited three researchers, trained by domain experts, to conduct in-depth analysis on 20 cases. The
three evaluators were graduate researchers with backgrounds in environmental law and public administration,
and they received a two-week training program delivered by domain experts with more than five years of procu-
ratorial case-handling experience. The training covered the interpretation of ecological damage case types, the
use of the system interface, and the criteria for judging clue relevance. None of the evaluators was affiliated with
the development team, and no conflict of interest existed between the evaluators and the authors of this study.
All experimental procedures in this study were performed in accordance with the relevant national research-
ethical guidelines and regulatory requirements. The research protocol of this study was reviewed and approved
by the Science and Technology Ethics Committee of Qiqihar University. Written informed consent was obtained
from all participants prior to their participation in this research.
The 20 cases covered both retrieval modes—case-to-image retrieval and image-to-case retrieval—across the
five predefined damage types (industrial storage tanks, wind power facilities, photovoltaic panels, coal mine,
coastal aquaculture) in various regions of China. Ten cases were evaluated under the case-to-image mode and
ten under the image-to-case mode; the full case list is provided in Appendix Table A1. The researchers interacted
with the system through a visual interface, raised questions based on their professional domain knowledge, and
provided feedback on the results returned by the system and on the system’s performance during the interaction
process.
To evaluate the system in a more intuitive manner, we conducted interviews with the researchers to obtain their
practical requirements in case handling, and then designed a Task Retrieval Quality Questionnaire (TRQQ). For
each case, the three researchers independently rated the six TRQQ items immediately after their interaction with
the system. For each researcher, the per-item scores were then averaged across the 20 cases, and the values in
Table 9 represent these case-averaged scores. The evaluation content is shown in Table 10.
In Table 10, the questionnaire covered three dimensions: response integrity, clue matching accuracy, and infor-
mation purity. Each item was scored on a three-level scale (0, 0.5, 1), with higher scores representing better
task output quality. For each case, the total TRQQ score was calculated as the sum of the six item scores, with a
maximum of 6.
After completing the task evaluation, the System Usability Scale (SUS) 57 was used to measure the overall
usability of the system, and each researcher completed the SUS questionnaire once after the full set of 20 cases.
The SUS score was computed following the standard procedure: for each item, the raw score is converted by
Table 8 Per-category Recall@K (%) for zero-shot and low-rank adaptation.
Category Method R@1 R@5 R@10 R@20 R@30
Industrial storage tanks Zero-shot 1.9 5.8 11.5 19.2 28.8
Industrial storage tanks LoRA 1.9 9.6 19.2 38.5 57.7
Wind power facilities Zero-shot 0.5 2.6 5.2 10.4 13.0
Wind power facilities LoRA 0.5 2.6 5.2 10.4 15.6
Photovoltaic panels Zero-shot 1.2 6.2 10.0 20.0 25.0
Photovoltaic panels LoRA 1.2 6.2 12.5 25.0 37.5
Coal mine Zero-shot 0.8 4.0 5.6 12.1 16.9
Coal mine LoRA 0.8 4.0 8.1 16.1 24.2
Coastal aquaculture Zero-shot 0.0 0.0 0.0 0.0 0.0
Coastal aquaculture LoRA 1.7 8.3 16.7 33.3 50.0
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
subtracting 1 from the odd-numbered items and subtracting the raw score from 5 for the even-numbered items;
the converted scores of all ten items are then summed and multiplied by 2.5, yielding a total score ranging from
0 to 100. The specific questions are shown in Table 11.
In Table 11, questions 4 and 10 primarily assess the system’s learnability, while the other eight questions
mainly evaluate the system’s usability.
Evaluation results
All 20 cases were independently evaluated by all three researchers.To mitigate potential learning effects, the case
order was randomized for each evaluator. Before the formal evaluation, each evaluator completed a warm-up ses-
sion on one of the 20 cases to become familiar with the system interface and the judging criteria; this case was
also included in the formal evaluation, and only its formal assessment was scored.
Table 11 Specific questions of the SUS.
ID Questions Scale
Strongly
disagree
Disagree Neutral Agree Strongly
agree
1 I think that I would like to use this system frequently. 1 2 3 4 5
2 I found the system unnecessarily complex. 1 2 3 4 5
3 I thought the system was easy to use. 1 2 3 4 5
4 I think that I would need the support of a technical person to be able to use this
system.
1 2 3 4 5
5 I found the various functions in this system were well integrated. 1 2 3 4 5
6 I thought there was too much inconsistency in this system. 1 2 3 4 5
7 I would imagine that most people would learn to use this system very quickly. 1 2 3 4 5
8 I found the system very cumbersome to use. 1 2 3 4 5
9 I felt very confident using the system. 1 2 3 4 5
10 I needed to learn a lot of things before I could get going with this system. 1 2 3 4 5
ID Questions Scale
Yes Could
be
better
NO
1 Can the system respond to your query normally? 1 0.5 0
2 Is any required task content missing in the returned results? 0 0.5 1
3 Do the type, time range and spatial scope of textual clues match your
requirements?
1 0.5 0
4 Do the type, time range and spatial scope of image clues match your
requirements?
1 0.5 0
5 Does the returned content contain redundant information? 0 0.5 1
6 Does the returned content contain invalid or useless information? 0 0.5 1
Table 10 Specific questions
of the TRQQ scale.
Question ID Researcher 1 Researcher 2 Researcher 3 Mean ± SD
1 1 1 1 1.00 ± 0.00
2 1 1 1 1.00 ± 0.00
3 0.5 0.5 0.5 0.50 ± 0.00
4 0.5 0.5 0.5 0.50 ± 0.00
5 0.75 0.9 0.7 0.78 ± 0.10
6 1 1 1 1.00 ± 0.00
Total 4.75 4.9 4.7 4.78 ± 0.10
Table 9 TRQQ scores aver-
aged over the 20 cases for
each researcher (maximum
total score: 6).
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
After the researchers conducted separate tests on the 20 cases, they completed the TRQQ and SUS question -
naires. The TRQQ scores are shown in Table 9, and the SUS scores are shown in Table 12.
The three individual SUS scores (77.5, 85.0, and 82.5) were computed as the sum of the converted scores (31,
34, and 33, respectively) multiplied by 2.5, yielding an overall mean of 81.67 ± 3.82 out of 100, which indicates
satisfactory system usability. For the TRQQ, the small standard deviations across the three evaluators (≤ 0.10 for
all item means) indicate a high degree of consistency in their ratings, and the overall score of 4.78 ± 0.10 out of 6
suggests that the system achieved satisfactory retrieval quality. To further quantify inter-rater reliability, we com-
puted the intraclass correlation coefficient (ICC) on the per-case total TRQQ scores of the three evaluators using
a two-way random-effects model with single measures and absolute agreement (ICC(2,1)); the resulting ICC of
0.944 indicates excellent inter-rater agreement.
According to the TRQQ evaluation results, the system was able to respond to user retrieval requests with sat -
isfactory response completeness, with few redundant or invalid outputs, and could match textual and image clues
in terms of category, time range, and spatial scope in most cases, although further optimization is still required.
According to the SUS evaluation results, the system features simple operation and achieved satisfactory usability,
while there remains room for improvement in system stability and overall functionality.
Objective retrieval evaluation
To further evaluate the core retrieval capability of the proposed system beyond subjective questionnaires, we
constructed a labeled test set based on the 20 cases.
The ground-truth clues were established as follows. For case-to-image retrieval, the domain experts annotated
all image tiles covering the case region and the corresponding time period: a tile was labeled as a ground-truth
clue if a facility of the corresponding damage type was present in the tile, and as a negative clue otherwise. The
annotation was based on the case facts and the remote sensing imagery itself rather than on the system outputs.
The annotation criteria were: (1) the image clue belongs to one of the five predefined damage types; and (2)
the facility is clearly identifiable in the remote sensing imagery. The annotation results of the ten case-to-image
cases are summarized in Appendix Table A2. For image-to-case retrieval, because textual clues are retrieved
from the live web and cannot be exhaustively predefined, all ten image-to-case cases were selected from events
with public web reports, so every case had a corresponding ground-truth event. After the system test, the domain
experts annotated the returned textual clues using the following criteria: (1) the clue belongs to one of the five
predefined damage types; (2) the location described in the clue covers the case region; and (3) the time of the clue
is not earlier than the specified time. We acknowledge that this constitutes a retrospective evaluation: the image-
to-case cases were chosen precisely because public information about the corresponding events exists, and the
text-retrieval hit rate may therefore overestimate the recall achievable in a blind setting in which no such public
information is available. If a facility detected in the imagery has no corresponding public web information, it
Question ID Raw score (R1/R2/R3) Converted score (R1/R2/R3) Mean ± SD
1 5 / 5 / 5 4 / 4 / 4 4.00 ± 0.00
2 2 / 1 / 2 3 / 4 / 3 3.33 ± 0.58
3 4 / 5 / 5 3 / 4 / 4 3.67 ± 0.58
4 2 / 2 / 2 3 / 3 / 3 3.00 ± 0.00
5 4 / 4 / 4 3 / 3 / 3 3.00 ± 0.00
6 2 / 2 / 2 3 / 3 / 3 3.00 ± 0.00
7 4 / 4 / 4 3 / 3 / 3 3.00 ± 0.00
8 2 / 2 / 2 3 / 3 / 3 3.00 ± 0.00
9 4 / 5 / 5 3 / 4 / 4 3.67 ± 0.58
10 2 / 2 / 2 3 / 3 / 3 3.00 ± 0.00
SUS score 77.5 / 85.0 / 82.5 81.67 ± 3.82
Table 12 SUS scores (raw
and converted scores per
researcher; each item is
scored on a scale of 0–4
after conversion).
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
should be recorded separately as an unavailable textual clue rather than counted as a retrieval failure; this situa -
tion did not occur in the present ten image-to-case cases.
The number of returned candidates follows the actual system configuration: the Image Agent returns the top
ten image candidates for the case-to-image mode, whereas the Search Agent returns the top five textual clues by
default for the image-to-case mode. Accordingly, image clues were evaluated at k = 10 and textual clues at k = 5.
We report precision@k together with the hit rate. Precision@k was computed as the proportion of relevant clues
among the top-k returned clues, judged by domain experts on the basis of the case facts; the denominator is k,
and the numerator is the number of clues judged relevant within those ranks. The hit rate indicates whether the
ground-truth clue was contained in the top-k returned clues; it serves as a binary measure of top-k retrieval suc -
cess and reflects the system’s recall capability without requiring an exhaustively enumerated ground-truth set,
which is not available for live web-sourced textual clues. The results are shown in Table 13.
The overall hit rate reached 0.80 (16 of 20 cases): all ten case-to-image cases contained the ground-truth clues
within the top ten image candidates, and six of the ten image-to-case cases contained the corresponding known
violation events within the top five textual clues. The mean precision@k was 0.38 for case-to-image retrieval and
0.40 for image-to-case retrieval. We acknowledge that these results leave room for improvement, and we briefly
analyze the likely causes of the failures here. For image retrieval, the moderate precision mainly reflects three
factors. First, several damage categories are visually similar to natural or anthropogenic background structures
(e.g., coal mines can be confused with other open-pit sites, and coastal aquaculture enclosures with similar water-
surface patterns), which is consistent with the per-category recall reported in Table 8: with the fine-tuned model,
wind power facilities reach R@30 = 15.6% and coal mines R@30 = 24.2%, whereas industrial storage tanks reach
R@30 = 57.7%. Second, the fine-tuned model was trained on a relatively small set (2,572 tiles), so the within-
category variation of the facilities is only partially captured. Third, a single tile-level candidate may contain only
part of a large facility, which lowers the precision of the tile ranking at k = 10. For text retrieval, the hit rate of
60% mainly stems from the dependence on public web coverage: some violation events are reported only in local
outlets, are described with place names that do not match the administrative-level filtering criterion, or were not
yet publicly reported within the retrieval window; events without public web coverage cannot be retrieved by any
web-based search engine by definition. Given the open-domain nature of the task, in which candidate clues are
retrieved from a large pool of web pages or image tiles while the ground-truth clues are sparse, these precision
levels and hit rates are acceptable for the current prototype, and the results confirm that the ground-truth clues
were generally contained in the returned candidate lists. The corresponding improvement directions (enlarging
the training set, adding multi-temporal verification, broadening textual clue sources such as law-enforcement
announcements and government information disclosure, and improving place-name normalization) are discussed
in Sect. 7.
The per-case results are provided in Appendix Table A3.
Conclusion
This paper proposes an LLM-agent-based cross-modal retrieval method for detecting clues of ecological envi -
ronmental damage. Combined with AI agent technology, the proposed system acquires the retrieval demands of
investigators through a conversational interface and realizes user-demand-driven cross-modal retrieval of tex -
tual and remote sensing image clues by invoking specialized sub-agents. It provides technical support for the
Table 13 Objective retrieval metrics on the labeled test set.
Metric Case-to-image (10 cases, k = 10) Image-to-case (10 cases, k = 5) Overall (20 cases)
Precision@k 0.38 0.4 -
Hit rate 1 0.6 0.8
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
subsequent screening and assessment of ecological damage issues and meets the needs of users without technical
backgrounds. Evaluations by procuratorial practitioners, together with retrieval experiments on a LoRA-fine-
tuned RemoteCLIP model, indicate satisfactory retrieval performance and usability.
This study has several limitations. First, the system is highly dependent on the availability and quality of pub-
lic web information, and the timeliness and stability of online sources may affect the retrieval results. Second,
the image retrieval model currently supports only five predefined facility types, which were selected as the most
common illegal construction categories within Ecological Conservation Redline zones; extending the framework
to additional types requires constructing new datasets and prompt templates. Third, the user evaluation involved
only three researchers, and the labeled test set is limited to twenty cases (ten per retrieval direction); objective
evaluation on a larger and more diverse test set remains future work. Fourth, the current system implements static
workflow orchestration rather than autonomous agent-to-agent negotiation.
In future work, we plan to expand the problem-clue database, broaden the acquisition channels of textual clues,
extend the framework to more damage types, introduce dynamic agent collaboration in which agents autono -
mously decompose sub-tasks and negotiate intermediate results, and conduct systematic ablation studies and
objective retrieval evaluation with a larger labeled test set and a larger sample of evaluators.
Appendix
Appendix Table A1.Summary of the 20 evaluation cases.
ID Mode Time Location Type
1 Case-to-image Recent years Ningxia Province Industrial storage tanks
2 Case-to-image Since 2020 Guangdong Province Industrial storage tanks
3 Case-to-image Since 2020 Guizhou Province Wind power facilities
4 Case-to-image Since 2020 Henan Province Wind power facilities
5 Case-to-image Since 2020 Anhui
Province
Photovoltaic panels
6 Case-to-image Since 2020 Jiangsu Province Photovoltaic panels
7 Case-to-image Since 2023 Shanxi Province Coal mine
8 Case-to-image Since 2020 Anhui Province Coal mine
9 Case-to-image Since 2020 Hainan Province Coastal aquaculture
10 Case-to-image Since 2020 Jiangsu Province Coastal aquaculture
11 Image-to-case Since 2021 Zhongwei City,
Ningxia Province
Industrial storage tanks
12 Image-to-case Since 2015 Zhongshan City,
Guangdong Province
Industrial storage tanks
13 Image-to-case Since 2016 Qiannan Prefecture,
Guizhou Province
Wind power facilities
14 Image-to-case Since 2020 Anyang City,
Henan Province
Wind power facilities
15 Image-to-case Since 2020 Bozhou City,
Anhui Province
Photovoltaic panels
16 Image-to-case Since 2020 Tianganghu Town,
Jiangsu Province
Photovoltaic panels
17 Image-to-case Since 2016 Xinzhou City,
Shanxi Province
Coal mine
18 Image-to-case Since 2020 Chizhou City,
Anhui Province
Coal mine
19 Image-to-case Since 2020 Lingao County,
Hainan Province
Coastal aquaculture
20 Image-to-case Since 2020 Lianyungang City,
Jiangsu Province
Coastal aquaculture
Appendix Table A2.Annotation results of the case-to-image cases on the labeled test set.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Case ID Total tiles Ground-truth tiles
1 24 8
2 4 1
3 21 7
4 1 1
5 63 4
6 624 145
7 6 6
8 1 1
9 18 8
10 84 29
Note. The number of annotated tiles varies across cases (from 1 to 624), reflecting differences in the extents of the case regions and the imagery
coverage.
Appendix Table A3.Per-case retrieval metrics on the labeled test set.
Case ID Mode Precision@k Hit (✓/✗)
1 Case-to-image (k = 10) 0.8 [✓]
2 Case-to-image (k = 10) 0.1 [✓]
3 Case-to-image (k = 10) 0.3 [✓]
4 Case-to-image (k = 10) 0.1 [✓]
5 Case-to-image (k = 10) 0.1 [✓]
6 Case-to-image (k = 10) 0.9 [✓]
7 Case-to-image (k = 10) 0.6 [✓]
8 Case-to-image (k = 10) 0.1 [✓]
9 Case-to-image (k = 10) 0.7 [✓]
10 Case-to-image (k = 10) 0.1 [✓]
11 Image-to-case (k = 5) 0 [✗]
12 Image-to-case (k = 5) 0 [✗]
13 Image-to-case (k = 5) 1 [✓]
14 Image-to-case (k = 5) 0 [✗]
15 Image-to-case (k = 5) 0.6 [✓]
16 Image-to-case (k = 5) 1 [✓]
17 Image-to-case (k = 5) 0.8 [✓]
18 Image-to-case (k = 5) 0 [✗]
19 Image-to-case (k = 5) 0.2 [✓]
20 Image-to-case (k = 5) 0.4 [✓]
Author contributions All authors contributed to the conception and design of the study. M.Z. performed model construc -
tion, experimental design, data analysis and drafted the manuscript. L.C. supervised the research and revised the manuscript.
The other authors participated in data collation, result validation and figure preparation. All authors reviewed and approved
the final manuscript.
Funding This research was supported in part by the Fundamental Research Funds in Heilongjiang Provincial Universities un-
der Grant No. 145309328, the National Key Research and Development Program of China under Grant 2022YFC330160204,
the Heilongjiang Province Higher Education Teaching Reform Project under Grant SJGY20200125, and the Tianfu Yongx-
ing Laboratory Organized Research Project Funding under Grant No. 2024KJGG18.
Data availability Data availability statement: The remote sensing imagery used in this study was obtained from the Esri
World Imagery Wayback service ( https://livingatlas.arcgis.com/wayback/). Under the applicable third-party licence, this
imagery is for internal research use only, and any redistribution is prohibited; consequently, neither the imagery nor the
derived image tiles can be made publicly available. The trained model weights of the LoRA-fine-tuned RemoteCLIP model
are available from the corresponding author upon reasonable request.Code availability: The implementation code is avail -
able at Zenodo: https://doi.org/10.5281/zenodo.23944660.
Code availability The implementation code is available at Zenodo: https://doi.org/10.5281/zenodo.23944660.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
Declarations
Competing interests The authors declare no competing interests.
Ethics and consent to participate declarations This study involving questionnaire surveys was approved by the Science and
Technology Ethics Committee of Qiqihar University. All participants provided informed consent prior to participating in the
questionnaire survey. All procedures were performed in accordance with the relevant guidelines.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use,
sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The
images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your
intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
References
- Barbarossa, V . et al. The many pathways of mining impacts on biodiversity. Conserv. L e t t . h t t p s : / / d o i . o r g / 1 0 . 1 1 1 1 / c o n l . 7
0 0 0 0 (2026). - Ripple, W. J. et al. The 2025 state of the climate report: A planet on the brink. BioScience75 (12), 1016–1027. h t t p s : / / d o
i . o r g / 1 0 . 1 0 9 3 / b i o s c i / b i a f 1 4 9 (2025). - Wu, L., Shen, W. M., Ma, W. D. & Bi, J. P. Research on ecological environment zoning management control monitoring
technology based on satellite remote sensing. Environ. Ecol. 114–120. (2024). - Xu, K. P. et al. Environmental function zoning for spatially differentiated environmental policies in China. J. Environ.
Manage.255, 109485. https://doi.org/10.1016/j.jenvman.2019.109485 (2020). - Gao, J. X. How China will protect one-quarter of its land. Nature569, 457. https://doi.org/10.1038/d41586-019-01563-2
(2019). - Ministry of Ecology and Environment of the People’s Republic of China. Notice on issuing the measures for the
supervision of the ecological environment of the ecological conservation red lines (trial). h t t p s : / / w w w . g o v . c n / g o n g b a o / c
o n t e n t / 2 0 2 3 / c o n t e n t _ 5 7 4 1 2 5 9 . h t m (2022). - Friends of Nature. The Bagongshan case closed by mediation: Public interest litigation promotes ecological restoration.
(2020). https://www.fon.org.cn/action/way/content/38 - Wenzhou Municipal Bureau of Agriculture and Rural Affairs. Case interpretation: A company in Wenzhou conducted
reclamation activities without approval. (2024). h t t p s : / / n y n c j . w e n z h o u . g o v . c n / a r t / 2 0 2 4 / 1 0 / 2 9 / a r t _ 1 2 1 2 8 8 6 _ 5 8 9 5 2 3 8 8 . h
t m l - Ungar, M. & Corredor-García, J. Enforcing environmental law in the Amazon. People Nat.6 (3), 70342. h t t p s : / / d o i . o r g /
1 0 . 1 0 0 2 / p a n 3 . 7 0 3 4 2 (2026). - Ingram, V ., Janssen, V ., Neh, V . A. & Pratihast, A. K. Cocoa driven deforestation in Cameroon: Practices and policy. For.
Policy Econ.177, 103533. https://doi.org/10.1016/j.forpol.2025.103533 (2025). - Bakhronova, S. et al. Assessment of ecological damage using GIS and remote sensing: A comprehensive literature review.
E3S Web Conf.563, 03009. https://doi.org/10.1051/E3SCONF/202456303009 (2024). - Wang, R. et al. Remote sensing application in ecological restoration monitoring: A systematic review. Remote Sens.16,
- https://doi.org/10.3390/RS16122204 (2024).
- Chen, C., Wang, L., Yang, G., Sun, W. & Song, Y . Mapping of ecological environment based on Google Earth Engine
cloud computing platform and Landsat long-term data: A case study of the Zhoushan Archipelago. Remote Sens.15, 4072.
https://doi.org/10.3390/RS15164072 (2023). - Baghdady, S. M. & Abdelsalam, A. A. Ten years of oil pollution detection in the Eastern Mediterranean shipping lanes
opposite the Egyptian coast using remote sensing techniques. Sci. Rep.14, 18057. h t t p s : / / d o i . o r g / 1 0 . 1 0 3 8 / s 4 1 5 9 8 - 0 2 4 - 6
7 9 8 3 - x (2024). - Yang, W. L., Zhou, Y . & Li, C. Z. Assessment of ecological environment quality in rare earth mining areas based on
improved RSEI. Sustainability15, 2964. https://doi.org/10.3390/SU15042964 (2023). - Song, W. et al. Environmental assessments in dense mining areas using remote sensing information over Qian’an and
Qianxi regions China. Ecol. Indic.146, 109814. https://doi.org/10.1016/J.ECOLIND.2022.109814 (2023). - Hu, L. et al. Ecological environment changes of mining areas around Nansi lake with remote sensing monitoring. Environ.
Sci. Pollut Res.28, 1–13. https://doi.org/10.1007/S11356-021-13849-Y (2021).
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
18. Chen, Y . X., Yu, X. Y ., Xu, D. M. & Peng, J. A spatial machine learning approach to exploring the impacts of coal mining
and ecological restoration on regional ecosystem health. Environ. Res.264, 120379. h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / J . E N V R E S .
2 0 2 4 . 1 2 0 3 7 9 (2024).
19. Xu, Y . & Lu, Z. An attention-fused deep learning model for accurately monitoring cage and raft aquaculture at large-scale
using Sentinel-2 data. IEEE J. Sel. Top. Appl. Earth Obs Remote Sens.17, 9099–9109. h t t p s : / / d o i . o r g / 1 0 . 1 1 0 9 / J S T A R S . 2
0 2 4 . 3 3 9 0 7 6 2 (2024).
20. Jabłońska, K. et al. MineCam: Application of combined remote sensing and machine learning for segmentation and
change detection of mining areas enabling multi-purpose monitoring. Remote Sens.16, 955. h t t p s : / / d o i . o r g / 1 0 . 3 3 9 0 / R S 1
6 0 6 0 9 5 5 (2024).
21. Xie, Y . et al. Localization, balance, and affinity: A stronger multifaceted collaborative salient object detector in remote
sensing images. IEEE Trans. Geosci. Remote Sens.63, 1–14. https://doi.org/10.1109/TGRS.2024.3506630 (2025).
22. Xie, Y . et al. Landslide extraction from aerial imagery considering context association characteristics. Int. J. Appl. Earth
Obs. Geoinf.131, 1–13. https://doi.org/10.1016/j.jag.2024.103950 (2024).
23. Song, W., Cui, X., Xie, Y ., Wang, G. & Ma, J. Monocular depth estimation via a detail semantic collaborative network
for indoor scenes. Sci. Rep.15 (1), 10990. https://doi.org/10.1038/s41598-025-96024-4 (2025).
24. Song, W. et al. A visual attention-guided approach for concrete crack detection in complex environments. Eng. Appl.
Artif. Intell.173https://doi.org/10.1016/j.engappai.2026.114439 (2026).
25. Song, W. et al. SDANet: A structure-detail aware network for image restoration in enclosed building spaces under fire-
smoke conditions. Developments Built Environ.27, 1–18. https://doi.org/10.1016/j.dibe.2026.101010 (2026).
26. Minaee, S. et al. Large language models: A survey. arXiv preprint arXiv . https://doi.org/10.48550/arXiv.2402.06196
(2024). :2402.06196.
27. Ye, Q. et al. mPLUG-Owl2: Revolutionizing multimodal large language model with modality collaboration. arXiv
preprint. https://doi.org/10.48550/arXiv.2311.04257 (2023). arXiv:2311.04257.
28. Zhang, Z. et al. A survey on the memory mechanism of large language model based agents. arXiv preprint. h t t p s : / / d o i . o
r g / 1 0 . 4 8 5 5 0 / a r X i v . 2 4 0 4 . 1 3 5 0 1 (2024). arXiv:2404.13501.
29. Guo, T. et al. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint. h t t p s : / / d o i
. o r g / 1 0 . 4 8 5 5 0 / a r X i v . 2 4 0 2 . 0 1 6 8 0 (2024). arXiv:2402.01680.
30. Zhang, Y ., Li, M., Wang, H. & Liu, Z. Mixture-of-agents enhances large language model capabilities. IEEE Trans. Artif.
Intell.5, 1892–1903. https://doi.org/10.1109/TAI.2024.3387654 (2024).
31. Wu, Q. Y . et al. AutoGen: Enabling next-gen LLM applications via multi-agent conversation. arXiv preprint arXiv .
https://doi.org/10.48550/arXiv.2308.08155 (2023). :2308.08155.
32. LangChain. LangGraph documentation. (2024). https://github.langchain.ac.cn/langgraph/tutorials/
33. Gao, D. W. et al. AgentScope: A flexible yet robust multi-agent platform. arXiv preprint. h t t p s : / / d o i . o r g / 1 0 . 4 8 5 5 0 / a r X i v .
2 4 0 2 . 1 4 0 3 4 (2024). arXiv:2402.14034.
34. Coze Coze documentation. (2024). https://www.coze.cn/open/docs/guides
35. Wang, C., Chen, L. & Li, X. Towards LLM agents for Earth observation. Remote Sens. Environ.298, 113789. h t t p s : / / d o i
. o r g / 1 0 . 1 0 1 6 / j . r s e . 2 0 2 4 . 1 1 3 7 8 9 (2025).
36. Smith, A., Jones, B. & Brown, C. MARSHA: Multi-agent RAG system for hazard adaptation. Nat. Hazards. 127, 890–
912. https://doi.org/10.1007/s11069-024-05876-x (2025).
37. DeepSeek-AI. Leveraging the DeepSeek large model: A framework for AI-assisted disaster prevention mitigation and
emergency response systems. Int. J. Disaster Risk Reduct. 106, 103892. https://doi.org/10.1016/j.ijdrr.2024.103892
(2025).
38. Liu, F. et al. RemoteCLIP: A vision language foundation model for remote sensing. IEEE Trans. Geosci. Remote Sens.62,
1–16. https://doi.org/10.1109/TGRS.2024.3390838 (2024).
39. DeepSeek-AI. DeepSeek LLM: Scaling open-source language models with longtermism. arXiv preprint. h t t p s : / / d o i . o r g /
1 0 . 4 8 5 5 0 / a r X i v . 2 4 0 1 . 0 2 9 5 4 (2024). arXiv:2401.02954.
40. Russell, S. & Norvig, P. Artificial Intelligence: A Modern Approach 4th edn (Pearson, 2020).
41. Stone, P. et al. Artificial intelligence and life in 2030: The one hundred year study on artificial intelligence. arXiv preprint.
https://doi.org/10.48550/arXiv.2211.06318 (2022). arXiv:2211.06318.
42. Wang, L. et al. A survey on large language model based autonomous agents. Front. Comput. Sci.18, 186345. h t t p s : / / d o i .
o r g / 1 0 . 1 0 0 7 / s 1 1 7 0 4 - 0 2 4 - 2 7 6 2 - 5 (2024).
43. Deng, X., Zhou, L., Dong, D. & Wei, J. Enhancing multi-agent communication collaboration through GPT-based semantic
information extraction and prediction. In: Proceedings of ACM TURC 2024, pp 81–85. h t t p s : / / d o i . o r g / 1 0 . 1 1 4 5 / 3 6 7 4 3 9 9
. 3 6 7 4 4 3 2 (2024).
44. Hong, S. et al. MetaGPT: Meta programming for a multi-agent collaborative framework. arXiv preprint. h t t p s : / / d o i . o r g /
1 0 . 4 8 5 5 0 / a r X i v . 2 3 0 8 . 0 0 3 5 2 (2024). arXiv:2308.00352.
45. Pan, H. et al. KwaiAgents: Generalized information-seeking agent system with large language models. arXiv preprint.
https://doi.org/10.48550/arXiv.2312.04889 (2023). arXiv:2312.04889.
Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-80668-x
46. Zhu, L. et al. RS-AGENT: Large language models guided agent system for remote sensing image generation. IEEE Int.
Geosci. Remote Sens. Symp. (IGARSS). 10640–10643. https://doi.org/10.1109/IGARSS54723.2024.10640804 (2024).
47. Li, Z. & Ning, H. Autonomous GIS: The next-generation AI-powered GIS. Int. J. Digit. Earth. 16, 4668–4686. h t t p s : / / d
o i . o r g / 1 0 . 1 0 8 0 / 1 7 5 3 8 9 4 7 . 2 0 2 3 . 2 2 7 8 8 9 5 (2023).
48. Xi, Z. et al. The rise and potential of large language model based agents: A survey. arXiv preprint. h t t p s : / / d o i . o r g / 1 0 . 4 8 5
5 0 / a r X i v . 2 3 0 9 . 0 7 8 6 4 (2023). arXiv:2309.07864.
49. Goecks, V . G. & Waytowich, N. R. DisasterResponseGPT: Large language models for accelerated plan of action
development in disaster response scenarios. arXiv preprint . https://doi.org/10.48550/arXiv.2306.17271 (2023).
arXiv:2306.17271.
50. Vaghefi, S. A. et al. ChatClimate: Grounding conversational AI in climate science. Commun. Earth Environ. 4, 480.
https://doi.org/10.1038/s43247-023-01108-4 (2023).
51. Thulke, D. et al. ClimateGPT: Towards AI synthesizing interdisciplinary research on climate change. arXiv preprint.
https://doi.org/10.48550/arXiv.2401.09646 (2024). arXiv:2401.09646.
52. Martelo, R., Ahmadiyeh Yazdi, K. & Wang, R. Q. Towards democratized flood risk management: An advanced AI
assistant enabled by GPT-4 for enhanced interpretability and public engagement. arXiv preprint. h t t p s : / / d o i . o r g / 1 0 . 4 8 5 5
0 / a r X i v . 2 4 0 3 . 0 3 1 8 8 (2024). arXiv:2403.03188.
53. Xie, Y . et al. A RAG-based multi-agent LLM system for natural hazard resilience and adaptation. arXiv preprint. h t t p s : /
/ d o i . o r g / 1 0 . 4 8 5 5 0 / a r X i v . 2 5 0 4 . 1 7 2 0 0 (2025). arXiv:2504.17200.
54. Liu, Z., Zhao, D., Yuan, B. & Jiang, Z. RescueADI: Adaptive disaster interpretation in remote sensing images with
autonomous agents. IEEE Trans. Geosci. Remote Sens. 63, 5611814. https://doi.org/10.1109/TGRS.2025.3532594
(2025).
55. Yang, J. et al. Chatecology: Revealing hidden drivers of green tide with the domain-tuned LLM in ecological engineering.
SSRN preprint. https://doi.org/10.2139/ssrn.5317264 (2025).
56. Bocha, A. I. Bocha AI Open Platform: Search API and Reranker API. https://open.bochaai.com/
57. Brooke, J. SUS-A quick and dirty usability scale. In (eds Jordan, P. W., Thomas, B. A., Weerdmeester, B. & McClelland,
I. L.) Usability evaluation in industry(189–194). Taylor & Francis. (1996).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional
affiliations.
Authors and Affiliations
Man Zhao1 · Lu Bai2 · Shiyao Lu3 · Xiaoqing He2 · Liangfu Chen4,5 · Fucai Ben6 ·
Zhibao Wang3,7
Man Zhao
03018@qqhru.edu.cn
Liangfu Chen
chenlf@aircas.ac.cn
1 School of Communication and Electronic Engineering, Qiqihar University, Qiqihar 161003, China
2 School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast, Belfast
BT9 6SB, UK
3 School of Computer and Information Technology, Northeast Petroleum University, Daqing 163318, China
4 State Key Laboratory of Remote Sensing Science, Aerospace Information Research Institute, Chinese Academy of
Sciences, Beijing 100101, China
5 University of Chinese Academy of Sciences, Beijing 100049, China
6 Beijing Zhongke Ruijing Technology Co., Ltd, Beijing 100101, China
7 Bohai-Rim Energy Research Institute, Northeast Petroleum University, Qinhuangdao 066004, China
论文原文与出版信息
Scientific Reports