Towards Intelligent Urban Park Development Monitoring: LLM Agents for Multi-Modal Information Fusion and Analysis

As an important part of urbanization, the development monitoring of newly constructed parks is of great significance for evaluating the effect of urban planning and optimizing resource allocation. However, traditional change detection methods based on remote sensing imagery have obvious limitations in high-level and intelligent analysis, and thus are difficult to meet the requirements of current urban planning and management. In face of the growing demand for complex multi-modal data analysis in urban park development monitoring, these methods often fail to provide flexible analysis capabilities for diverse application scenarios. This study proposes a multi-modal LLM agent framework, which aims to make full use of the semantic understanding and reasoning capabilities of LLM to meet the challenges in urban park development monitoring. In this framework, a general horizontal and vertical data alignment mechanism is designed to ensure the consistency and effective tracking of multi-modal data. At the same time, a specific toolkit is constructed to alleviate the hallucination issues of LLM due to the lack of domain-specific knowledge. Compared to vanilla GPT-4o and other agents, our approach enables robust multi-modal information fusion and analysis, offering reliable and scalable solutions tailored to the diverse and evolving demands of urban park development monitoring.

💡 Research Summary

The paper addresses the growing need for intelligent monitoring of newly built urban parks, a task that traditional remote‑sensing change‑detection methods struggle to fulfill because they are limited to pixel‑level, quantitative analysis and lack semantic interpretation. To overcome these limitations, the authors propose a multi‑modal Large Language Model (LLM) agent framework that leverages the advanced language understanding and reasoning capabilities of models such as GPT‑4o, while integrating heterogeneous data sources (high‑resolution satellite imagery, GIS vector layers, tabular CSV data, LiDAR point clouds, and socio‑economic statistics).

The framework consists of two core stages. In the first stage, the agent decomposes a user query into subtasks and performs horizontal and vertical data alignment. Horizontal alignment converts all spatial data to a common coordinate reference system and synchronizes temporal granularity, thereby reconciling differences in resolution and acquisition frequency among modalities. Vertical alignment assigns a global unique identifier to each data element, enabling traceability of every transformation step across the entire workflow. This dual‑alignment strategy guarantees both spatial‑temporal consistency and data lineage, which are essential for reliable multi‑step analysis.

In the second stage, the agent selects and invokes domain‑specific tools from a curated toolkit. The toolkit includes a CSV column selector, GIS coordinate transformer, LiDAR‑to‑image conversion pipeline, and land‑use/land‑cover (LULC) computation modules. By calling these external tools, the LLM moves beyond pure text generation to perform concrete data operations, thereby mitigating the well‑known hallucination problem that arises when LLMs lack domain knowledge. The structured outputs from the tools are fed back to the LLM, which then performs logical reasoning over the combined results and generates a comprehensive, human‑readable report.

For evaluation, the authors built a question set from New York City open data, covering three difficulty levels: basic retrieval, qualitative synthesis, and quantitative change analysis. They benchmarked the proposed agent against (1) vanilla GPT‑4o, (2) default LangChain agents (SQL, Pandas DataFrame, CSV), and (3) a single‑modality CSV‑only agent. While the baseline models handled simple retrieval tasks, they failed on higher‑level queries due to token limits, lack of data lineage, and inability to fuse non‑tabular modalities, often producing hallucinated or incomplete answers. In contrast, the proposed multi‑modal agent consistently answered all questions correctly, accurately aligning and integrating data across modalities, and delivering reliable quantitative assessments such as land‑use proportion changes over time.

The paper’s contributions are threefold: (1) a novel multi‑modal LLM agent architecture tailored for urban park development monitoring; (2) a generalized horizontal/vertical data alignment mechanism that ensures consistency and traceability of heterogeneous datasets; and (3) a domain‑specific toolkit that reduces LLM hallucination by grounding reasoning in concrete tool‑based computations. The experimental results demonstrate that this combination yields superior performance, robustness, and scalability compared with existing approaches.

Future directions suggested include extending the framework to additional modalities (e.g., real‑time IoT sensor streams, social‑media imagery, audio), enriching the toolkit with more specialized analytics, and exploring multi‑agent collaboration for city‑wide monitoring platforms. By bridging LLM reasoning with traditional GIS and remote‑sensing pipelines, the work paves the way for smarter, data‑driven urban planning and resource allocation in the era of smart cities.