CangLing-KnowFlow

A Unified Knowledge-and-Flow-fused Agent for Comprehensive Remote Sensing Applications

---

Important

Project status. The paper is currently under peer review. To protect the work during the review process, the full code and datasets are not yet publicly released. We plan to fully open-source this project once the paper is accepted.

In the meantime, the code and data are available on request for academic and non-commercial use — please see Data & Code Availability.

---

Overview

Automated and intelligent processing of massive remote sensing (RS) datasets is critical for Earth observation. Existing automated systems are typically task-specific and lack a unified framework to manage diverse, end-to-end workflows — from data preprocessing to advanced interpretation — across a wide range of RS applications.

CangLing-KnowFlow is a unified intelligent agent framework that fuses expert Knowledge into adaptive and verifiable Flows. It integrates three tightly-coupled modules:

• 📚 Procedural Knowledge Base (PKB) — 1,008 expert-validated workflow cases spanning 162 practical RS tasks. It guides planning and substantially reduces the planning hallucinations common in general-purpose agents.
• 🔄 Dynamic Workflow Adjustment — during runtime failures, the agent autonomously diagnoses the issue and replans a recovery strategy, remaining robust under the variability of real-world RS data and unexpected errors.
• 🧠 Evolutionary Memory — a long-term memory that continuously learns from execution: it solidifies successful adjustments into new workflows and attributes failures to heuristic knowledge, enabling continuous, experience-driven improvement.

Together, these modules let CangLing-KnowFlow adapt, learn, and operate reliably across diverse and complex Earth observation tasks.

Key Contributions

1. Codified expert knowledge for scientifically valid planning. A procedural knowledge base that constrains LLM-induced fallacies and ensures the scientific validity of generated plans.
2. Autonomous real-time adaptation. Real-time monitoring and dynamic re-planning that deliver robust performance in dynamic, unseen scenarios.
3. Persistent, experience-driven learning. A long-term memory mechanism that addresses the critical learning gap in current agentic workflows.
4. A comprehensive benchmark. KnowFlow-Bench, comprising 162 distinct RS task types and 324 real-world workflows, for rigorously evaluating an agent's workflow-planning ability under given tool constraints.

KnowFlow-Bench

We evaluate CangLing-KnowFlow on KnowFlow-Bench, a new benchmark of 324 workflows inspired by real-world applications, across 13 top LLM backbones (from open-source to commercial). The benchmark covers four main categories:

• Semantic Segmentation
• Object Detection
• Change Detection
• Remote Sensing Inversion

Across all complex tasks, CangLing-KnowFlow surpasses the Reflexion baseline by at least 4% in task success rate.

📦 Data & Code Availability

Component	Status
Paper (preprint)	✅ Available on arXiv
Source code	🔒 Available on request  ·  full release upon paper acceptance
Procedural Knowledge Base	🔒 Available on request  ·  full release upon paper acceptance
KnowFlow-Bench	🔒 Available on request  ·  full release upon paper acceptance

The code and data are available for academic and non-commercial research on request. If you would like early access, please send an email to the contact below with a short note describing your intended use and affiliation:

📧 Haoran Wang — wanghaoran23@mails.ucas.ac.cn

We will be happy to share the relevant materials.

Citation

If you find our work useful in your research, please consider citing:

@article{chen2025cangling,
  title   = {CangLing-KnowFlow: A Unified Knowledge-and-Flow-fused Agent for Comprehensive Remote Sensing Applications},
  author  = {Chen, Zhengchao and Wang, Haoran and Yao, Jing and Zhang, Jianshe and Ghamisi, Pedram and Zhou, Jun and Atkinson, Peter M. and Zhang, Bing},
  journal = {arXiv preprint arXiv:2512.15231},
  year    = {2025}
}

Contact

For questions, collaborations, or data/code access, feel free to reach out:

• Haoran Wang — wanghaoran23@mails.ucas.ac.cn

State Key Laboratory of Remote Sensing and Digital Earth, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China.

---

_{⭐ If you are interested in this project, please consider starring the repo — the full open-source release will land here after acceptance.}