S-MADSS: An Agentic Framework for Evidence-Grounded, Compliance-Oriented Decision Support in Atlantic Salmon Farming
Smart Agricultural Technology · 2026-11 · 102734
DOI: 10.1016/j.atech.2026.102734 ↗Kurz gefasst
Fachartikel in der Zeitschrift Smart Agricultural Technology, 2026-11. Verfasst von Chao Zhang und 3 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Dutch Research Council, China Scholarship Council.
Abstract (Originalsprache)
Atlantic salmon farming requires operational decisions that must satisfy regulatory requirements, protect fish welfare, and remain feasible under farm conditions. Existing digital tools mainly support isolated functions, such as lice forecasting or feeding optimization, and rarely provide traceable decision support across a full operational workflow. To address this gap, we propose S-MADSS, an evidence-grounded decision-support system for Atlantic salmon farming. The system supports two main functions: (i) evidence-grounded question answering (QA) for informational queries and (ii) compliance- and risk-oriented decision support for Atlantic salmon sea-cage operations. The system combines a traceable retrieval-augmented generation (RAG) pipeline with a multi-agent auditing workflow. The evidence pipeline integrates BM25 retrieval, dense vector retrieval, cross-encoder reranking, and local context expansion to construct supporting evidence from a domain knowledge base of regulations, manuals, and scientific literature. On top of this evidence layer, specialized agents perform answer generation, review, action proposal, and multi-perspective risk auditing from regulatory, welfare, and feasibility perspectives. We evaluate S-MADSS through four complementary settings: retrieval benchmarking, evidence-grounded QA, multi-hop evidence integration, and scenario-based risk identification, followed by a real-farm workflow demonstration. Hybrid retrieval with reranking achieved the best retrieval performance, with a mean Hit@k of 0.910 and Hit@1 improving from 0.478 to 0.752 over the BM25-only baseline. Local context expansion improved evidence coverage and answer correctness in multi-hop QA. In risk identification, the best retrieval-based configuration achieved a precision of 0.779, recall of 0.819, and F1 of 0.799, compared with a maximum F1 of 0.604 for the retrieval-free baselines. A real-farm case further showed that monitoring data can be translated into compliance-oriented recommendations with explicit assumptions, risks, and supporting evidence. These results indicate that S-MADSS extends retrieval-augmented generation from information access to structured, traceable, and compliance-oriented decision support in salmon farming, and highlight the potential of evidence-grounded agentic workflows for regulation-sensitive agricultural domains beyond salmon farming.
Förderung
- Dutch Research Council
- China Scholarship Council
Zitieren
APA
Zhang, C., Gansel, L. C., Bracke, M., & Torres, R. D. S. (2026). S-MADSS: An Agentic Framework for Evidence-Grounded, Compliance-Oriented Decision Support in Atlantic Salmon Farming. Smart Agricultural Technology, 102734. https://doi.org/10.1016/j.atech.2026.102734
Vancouver
Zhang C, Gansel LC, Bracke M, Torres RDS. S-MADSS: An Agentic Framework for Evidence-Grounded, Compliance-Oriented Decision Support in Atlantic Salmon Farming. Smart Agricultural Technology. 2026:102734. doi:10.1016/j.atech.2026.102734
BibTeX
@article{Zhang2026,
author = {Zhang, Chao and Gansel, Lars Christian and Bracke, Marc and Torres, Ricardo da Silva},
title = {S-MADSS: An Agentic Framework for Evidence-Grounded, Compliance-Oriented Decision Support in Atlantic Salmon Farming},
journal = {Smart Agricultural Technology},
year = {2026},
pages = {102734},
doi = {10.1016/j.atech.2026.102734}
}
EUVIMED