## The Three Fundamental Barriers of Traditional Monitoring: Why Empirical Experience is No Longer Sufficient Before discussing the technical advantages of AIoT (Artificial Intelligence of Things), it is imperative to address the precarious operational realities of traditional aquaculture. **First, there is the widening gap in empirical knowledge.** Skills such as judging water quality by color or identifying algal phases are often regarded as subjective or quasi-mystical. In reality, these are non-standardized, implicit knowledge sets accumulated over decades. However, in modern high-density aquaculture, this knowledge is difficult to replicate. During my doctoral research at the National University of Singapore (NUS), I observed that empirical decision-making is frequently delayed and inconsistent. Should a veteran farm manager depart, the productivity of the entire facility often suffers an immediate decline. **Second, the existence of critical monitoring blind spots.** It is physically impossible for personnel to monitor water conditions 24 hours a day. Even for the most diligent workers, measuring water quality twice daily represents the upper limit of manual labor. Yet, water quality dynamics are continuous. Under intense convective weather conditions, dissolved oxygen ($DO$) levels can plummet from a safe threshold to lethal concentrations in as little as 30 minutes. The temporal gap between manual checks constitutes a "death zone" that can eradicate hundreds of thousands of dollars in production value. **Finally, the failure of response speed.** When an operator observes fish gasping at the surface, ammonia nitrogen ($NH_3$) or nitrite ($NO_2^-$) levels have likely exceeded safe limits for several hours. This lagging response keeps aquaculture categorized as a "high-risk" industry. In our field investigations, a single algal crash incident resulting from insufficient monitoring typically results in losses exceeding tens of thousands of dollars. Global aquaculture is currently at a historical inflection point, transitioning from experience-driven traditional farming to a precision industry powered by data and physical laws. In this transition, inductive machine learning—which relies solely on massive data sets—has revealed limitations in generalization and physical interpretability. Leading-edge academic research is now returning to "first principles," embedding physical laws such as computational biology, fluid mechanics, and thermodynamic energy balance models into the underlying architecture of AI and computer vision. Guided by these logical foundations, the complex systems of aquaculture are being deconstructed into mathematical models of ecological stoichiometry, allometric scaling, and fluid dynamics. By merging rigorous mathematical models with deep learning frameworks, researchers can construct hybrid intelligent systems capable of understanding causality rather than merely identifying patterns. Within this framework, IoT, underwater computer vision, acoustic sensors, and Autonomous Underwater Vehicles (AUVs) serve as the sensory and executive organs. This report provides a comprehensive analysis of how global academic teams are applying these technologies—from microscopic identification of fish behavior to macro-scale IoT architectures for water quality prediction and the digital twin engineering of high-end offshore cages in Europe. ## 1\. Decoding Fish Behavior—Quantitative Mechanisms for Feeding Intensity, Abnormal Responses, and Smart Feeding Decisions The core asset of aquaculture is the aquatic organism itself. For practical farming operations, metrics such as the feeding intensity of the fish school, hypoxia stress responses, abnormal behaviors, and school density are the foundational indicators that directly determine the Feed Conversion Ratio (FCR), survival rate, and farming profitability. Academic teams led by scholars such as Daoliang Li, Yingyi Chen, Huihui Yu, Chao Zhou, and Song Zhang are dedicated to converting the complex behavioral characteristics of fish schools into mathematical vectors that machines can precisely compute, thereby realizing a leap from "experience-based feeding" to "precision on-demand feeding". ### 1.1 Multimodal Data Fusion and Precise Quantification of Feeding Intensity Feed costs typically account for over 60% of the total operating expenses in commercial aquaculture. Therefore, accurately assessing the Feeding Intensity of a fish school to avoid feed waste and water quality degradation is a primary pain point in production management. When fish feed, significant changes occur in the school's shape, swimming speed, water surface ripples, and the hydrodynamic and acoustic signals generated. Teams led by Li and Chen have pointed out that traditional single-vision modalities easily fail in industrialized, high-density farming or turbid water environments. Consequently, Multimodal Fusion has become the cutting-edge trend.
