{"id":28608,"date":"2026-11-08T23:24:20","date_gmt":"2026-11-08T22:24:20","guid":{"rendered":"https:\/\/www.proefschriftmaken.nl\/portfolio\/chao-zhang\/"},"modified":"2026-11-15T20:16:04","modified_gmt":"2026-11-15T19:16:04","slug":"chao-zhang","status":"publish","type":"us_portfolio","link":"https:\/\/www.proefschriftmaken.nl\/en\/portfolio\/chao-zhang\/","title":{"rendered":"Chao Zhang"},"content":{"rendered":"","protected":false},"featured_media":28609,"template":"","meta":{"_acf_changed":false},"university_us_portfolio":[],"class_list":["post-28608","us_portfolio","type-us_portfolio","status-publish","has-post-thumbnail","hentry"],"acf":{"main_text":"","naam_van_het_proefschift":"Artificial Intelligence for Salmon Farming","samenvatting":"Aquacultuur speelt een belangrijke rol in de mondiale voedselzekerheid en menselijke voeding, maar de verdere ontwikkeling ervan vereist een meer verantwoord en duurzaam beheer. Dit is bijzonder uitdagend bij de kweek van Atlantische zalm in open kooien op zee, waar vissen worden gekweekt in grote en dynamische mariene omgevingen en informatie over omgevingsomstandigheden, gedrag, gezondheid, welzijn en regelgevingsrisico's vaak onvolledig, vertraagd of onzeker is. Effectief beheer hangt daarom af van het verbeteren van de manier waarop boerderijgegevens, wetenschappelijke kennis en regelgevende informatie worden verkregen, ge\u00efnterpreteerd en gebruikt.\n\nHet overkoepelende doel van dit proefschrift is om te onderzoeken hoe op kunstmatige intelligentie (AI) gebaseerde methoden de waarneming van de toestand van de boerderij, welzijnsbeoordeling en besluitvormingsondersteuning in de zalmkwekerij kunnen ondersteunen, en zo kunnen bijdragen aan een transparanter en wetenschappelijk onderbouwd beheer van de aquacultuur. De kweek van Atlantische zalm is gekozen als casussysteem vanwege de combinatie van productie in open omgevingen, welzijns- en bioveiligheidsrisico's en wettelijke vereisten.\n\nHet proefschrift behandelt drie onderling verbonden managementlagen die elk een cruciaal knelpunt vormen tussen gegevensverwerving, kennisgebruik en managementbeslissingen. Zalmluizen dienen als representatieve casus op de waarnemingslaag, waar AI-gebaseerde detectie van vroege planktonische stadia in zeewater wordt onderzocht. Op de welzijnslaag wordt AI-ondersteunde extractie en classificatie van wetenschappelijke uitspraken onderzocht ter ondersteuning van welzijnsmodellering. Op de besluitvormingslaag wordt een modulair AI-agent-raamwerk verkend voor het integreren van informatie in op bewijs gebaseerde besluitvorming. Het proefschrift is als volgt opgebouwd:\n\nHoofdstuk 2 onderzoekt de detectie van planktonische stadia van zalmluizlarven in zeewater met een nieuw online beeldverwervingssysteem. Een dataset van 8316 beelden werd gebruikt om zestien objectdetectiemodellen te evalueren, waarbij RT-DETR de beste resultaten behaalde (97,2% precisie). Dit is de eerste studie naar snelle detectie van larven in verschillende stadia in zeewatermonsters.\n\nHoofdstuk 3 breidt dit uit naar complexe real-world omgevingen waar larven schaars zijn te midden van onzuiverheden. Een geautomatiseerde synthetische datageneratiemethode op basis van het Segment Anything Model (SAM) verbeterde de modelprestaties aanzienlijk, met een stijging van de recall met 11,0% tot 87,0% in praktijktests.\n\nHoofdstuk 4 introduceert 'Online Synthetic Training' (SDOST), waarbij tijdens de training dynamisch nieuwe beeldcomposities worden gegenereerd. Dit mechanisme vermindert overfitten en verbetert de F1-scores aanzienlijk, vooral onder omstandigheden met weinig beschikbare data.\n\nHoofdstuk 5 richt zich op AI-ondersteunde kennisextractie voor dierenwelzijn. Er werd een NLP-raamwerk ontwikkeld om automatisch welzijnsgerelateerde uitspraken uit wetenschappelijke literatuur te identificeren met een precisie van 96,0%. Dit ondersteunt de systematische constructie van welzijnsmodellen en vermindert handmatig screenwerk.\n\nHoofdstuk 6 introduceert het 'Salmon Farm Multi-Agent Decision Support System' (S-MADSS). Dit systeem combineert traceerbare bewijsvoering met gespecialiseerde agenten die adviseren vanuit regelgevende, welzijns- en technische perspectieven. Casestudies tonen aan dat S-MADSS effectief verschillende factoren kan co\u00f6rdineren binnen een uniforme workflow.\n\nHoofdstuk 7 vat de bevindingen samen en bespreekt de bredere implicaties voor duurzame aquacultuur en de rol van menselijk oordeel bij AI-ondersteund beheer.\n\nDit proefschrift legt de methodologische fundamenten voor een meer op bewijs gebaseerde en transparante zalmkwekerij. De ontwikkelde methoden zijn bedoeld om menselijke besluitvorming te ondersteunen door complexe data en regelgeving expliciet en inzichtelijk te maken.","summary":"Aquaculture plays an important role in global food security and human nutrition, but its continued development requires more responsible and sustainable management. This is particularly challenging in open sea-cage salmon farming, where fish are raised in large and dynamic marine environments, and information about environmental conditions, behaviour, health, welfare, and regulatory risks is often incomplete, delayed, or uncertain. Effective management therefore depends on improving how farm data, scientific knowledge, and regulatory information are acquired, interpreted, and used.\n\nThe overall aim of this thesis is to investigate how artificial intelligence (AI)-based methods can support farm-state perception, welfare assessment, and decision support in salmon farming, thereby contributing to more transparent and evidence-based aquaculture management. Atlantic salmon farming is selected as the case system for its combination of open-environment production, welfare and biosecurity risks, and regulatory requirements.\n\nThe thesis addresses three interconnected management layers, each representing a key bottleneck between data acquisition, knowledge use, and management decisions. Salmon lice serve as the representative case at the perception layer, where this thesis investigates AI-based detection of early planktonic-stage lice in seawater. At the welfare layer, it examines AI-assisted extraction and classification of welfare-related scientific statements to support semantic welfare modelling. At the decision-support layer, it explores a modular AI-agent-based framework for integrating regulatory and management-related information into evidence-grounded decision support. Specifically, the thesis is structured as follows:\n\nChapter 2 explores the detection of planktonic-stage salmon lice larvae in seawater. In this chapter, an online image acquisition system is introduced to continuously collect images from seawater samples. Using this system, a dataset of 8316 images was constructed for salmon lice detection. Sixteen object detection models, including one-stage and two-stage CNN-based models and Transformer-based models, were evaluated for salmon lice detection. RT-DETR achieved the best overall performance with a precision of 97.2% and a recall of 97.8%, while most YOLO-family models maintained high precision and recall with fast inference speeds. To our knowledge, this is the first study to apply machine learning to the rapid detection of different developmental stages of salmon louse larvae in seawater samples. The results demonstrate the feasibility of machine learning-based planktonic lice detection and establish a baseline for subsequent studies on rapid, easy-to-use, and cost-effective salmon lice monitoring in seawater.\n\nChapter 3 extends salmon lice detection from relatively clean and controlled seawater samples to more complex real-world seawater backgrounds. In such environments, both data acquisition and detection become more difficult because salmon louse larvae are small and sparse, while seawater samples contain diverse non-target organisms, particles, and impurities. To address the lack of annotated data under these conditions, this chapter presents an automated synthetic data generation method based on the Segment Anything Model (SAM), designed to improve the detection of multi-stage salmon louse larvae in complex seawater environments. By extracting annotated lice instances and combining them with real target-free background images, the method makes effective use of large numbers of unlabeled background images and generates diverse, realistic synthetic training data. Orthogonal experiments show that the diversity of non-target regions in synthetic images was important for data quality and model performance. Across different detection models, the use of synthetic data consistently improved detection performance. In real-world seawater tests, synthetic data increased model recall by 11.0% to 87.0%, further demonstrating its value for improving detection under complex seawater conditions.\n\nChapter 4 further explores online synthetic data generation for salmon lice detection and proposes Salmon Lice Detection based on Online Synthetic Training (SDOST). Unlike offline synthesis, where synthetic images are fixed before training, SDOST generates new image compositions during training, allowing the model input to remain dynamic and reducing overfitting. A feature-proximal replay mechanism was used to reuse synthetic samples with features close to the real-data distribution, while a fast gradient sign method (FGSM)-based adversarial perturbation was introduced to improve robustness. Compared with models trained only on fixed offline synthetic data or the original limited real data, SDOST improved the F1-score by 4.3% to 17.7% and 13.8% to 166.4%, respectively. These improvements were especially evident under few-shot conditions, demonstrating that online synthetic training can support the detection of small, sparse, and visually challenging salmon lice larvae and may also be useful for other aquaculture perception tasks where annotated target data are scarce.\n\nChapter 5 explores AI-assisted knowledge extraction for modelling of animal welfare. Semantic models for welfare assessment depend on scientific statements that describe environment-based welfare attributes, animal-based welfare measures grouped into weighting categories, and their interrelationships. However, these statements are scattered across large amounts of literature and are normally identified manually by domain experts. In this chapter, a natural language processing (NLP)-based framework is developed to automatically screen scientific documents, split them into sentences, and identify welfare-related statements. Different classification strategies were evaluated, including large language model (LLM)-based few-shot learning, two-stage classification, and end-to-end fine-tuning of pretrained Transformer models. The results show that end-to-end fine-tuning achieved the strongest overall performance and could identify welfare-related statements with a precision of 96.0% and a recall of 92.4% on the main test set. An interactive system was also developed so that users can upload documents, inspect classified sentences, and use the extracted statements as traceable evidence for the construction and updating of the welfare model. This chapter demonstrated that AI-based text analysis can reduce manual literature screening and support the systematic use of scientific knowledge in animal welfare assessment.\n\nChapter 6 explores AI-agent-based decision support for Atlantic salmon farming and introduces the Salmon Farm Multi-Agent Decision Support System (S-MADSS), an evidence-grounded, multi-agent framework for supporting farm-management decisions. The framework combines a traceable evidence-construction pipeline with a structured multi-agent workflow for evidence-grounded question answering and compliance-oriented decision support. Retrieved evidence is provided to specialized agents that generate answers, recommend actions, and candidate decisions from regulatory, welfare, risk, and feasibility perspectives. The results show that hybrid retrieval with reranking improves evidence quality, while local context expansion improves knowledge recall by incorporating adjacent text around retrieved knowledge fragments. In risk identification, the evidence-grounded auditing workflow achieves a maximum recall of 95.0%, compared with 63.7% for retrieval-free LLM-based identification. Scenario-based decision tasks and a real-farm case study on salmon lice treatment further demonstrate that S-MADSS can coordinate treatment feasibility, welfare concerns, regulatory compliance, and farm-specific information within a unified workflow. This chapter demonstrated that modular AI-agent systems can provide a feasible foundation for structured, traceable, and evidence-based decision support in salmon farming. The modular design also makes S-MADSS flexible and extensible. In future work, perception and analysis modules, such as broader salmon lice detection, welfare assessment, environmental monitoring, and farm-operation analysis, could be integrated into the framework so that decision support is grounded in a wider range of farm-specific evidence.\n\nChapter 7 synthesizes the findings from the preceding chapters and discusses how the three AI-based components contribute to salmon farming management. It reflects on the addressed knowledge gaps, the broader implications for AI-supported salmon farming and sustainable aquaculture, and the main limitations of the current studies. It also outlines future research directions, including broader commercial validation and integration of perception, welfare, environmental, and decision-support modules, and discusses wider societal implications related to animal welfare, regulatory transparency, environmental responsibility, and human judgement in AI-supported farm management.\n\nThe thesis contributes methodological foundations for more evidence-based, welfare-aware, and transparent salmon farming management, thereby supporting a broader transition towards more responsible and sustainable aquaculture. Rather than replacing farmer judgement, these methods are intended to support human decision-making under uncertainty by making relevant data, knowledge, assumptions, risks, and regulatory considerations more explicit. Although Atlantic salmon farming is used as the case system, the workflows developed in this thesis may also be relevant to other aquaculture and animal production systems facing similar challenges in perception, welfare assessment, and decision-making.","auteur":"Chao Zhang","auteur_slug":"chao-zhang","publicatiedatum":"2 december 2026","taal":"EN","url_flipbook":"https:\/\/ebook.proefschriftmaken.nl\/ebook\/chaozhangproof?iframe=true","url_download_pdf":"https:\/\/ebook.proefschriftmaken.nl\/download\/e9bb36c4-5475-4996-b057-e5963b3a13db\/highres","url_epub":"","ordernummer":"19961","isbn":"","doi_nummer":"10.18174\/682161","naam_universiteit":"Wageningen University","afbeeldingen":28609,"video_url":"","podcast_url":"","naam_student:":"","binnenwerk":"","universiteit":"Wageningen University","cover":"","afwerking":"","cover_afwerking":"","design":""},"_links":{"self":[{"href":"https:\/\/www.proefschriftmaken.nl\/en\/wp-json\/wp\/v2\/us_portfolio\/28608","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.proefschriftmaken.nl\/en\/wp-json\/wp\/v2\/us_portfolio"}],"about":[{"href":"https:\/\/www.proefschriftmaken.nl\/en\/wp-json\/wp\/v2\/types\/us_portfolio"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.proefschriftmaken.nl\/en\/wp-json\/wp\/v2\/media\/28609"}],"wp:attachment":[{"href":"https:\/\/www.proefschriftmaken.nl\/en\/wp-json\/wp\/v2\/media?parent=28608"}],"wp:term":[{"taxonomy":"university_us_portfolio","embeddable":true,"href":"https:\/\/www.proefschriftmaken.nl\/en\/wp-json\/wp\/v2\/university_us_portfolio?post=28608"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}