Fundamentally, WaterDoctor addresses not a singular water quality issue, but a highly coupled, complex system. Variables such as fish and shrimp feeding behavior, excretion, organic loading, DO, pH, ammonia nitrogen, nitrite, salinity, temperature, microbial communities, aeration, hydrodynamics, AquaMats biofilm maturity, effluent discharge standards, and client operational objectives fluctuate simultaneously. Traditional aquaculture relies heavily on empirical judgment, and conventional water treatment depends on static design parameters. However, modern high-density Recirculating Aquaculture Systems (RAS) and aquaculture effluent management necessitate an intelligent system capable of continuous perception, prediction, interpretation, and optimization. The textbook _Deep Learning_ by Ian Goodfellow et al., recommended by our team members, structurally builds deep learning upon mathematical foundations, machine learning principles, deep network practices, regularization, optimization, CNNs, RNNs, and generative models. This structural framework perfectly aligns with WaterDoctor's technological foundation as we upgrade from a "water treatment service provider" to an "AI-driven microbial water environment operating system company." ## I. AquaOS: From Water Quality Dashboard to "Water Environment Predictive Brain" The most direct empowerment of AI for WaterDoctor is the upgrade of AquaOS from a mere "data visualization platform" to a comprehensive "prediction, early warning, diagnosis, and recommendation" system. Existing aquaculture AI reviews indicate that artificial intelligence, machine learning, and deep learning are actively being applied across various domains, including water quality monitoring, disease detection, fish biomass estimation, feeding optimization, and intelligent decision-making. However, these applications simultaneously face challenges such as data standardization, model interpretability, insufficient labeled data, and difficulties in cross-scenario generalization. For WaterDoctor, AquaOS should be structured across three modeling layers. The first layer consists of traditional machine learning models—such as Random Forest, XGBoost, LightGBM, and SVM—utilized for predicting DO, ammonia nitrogen, nitrite, and pH risks under small-sample conditions. These models are particularly suitable for early-stage projects due to their strong interpretability and low training costs. Existing research on aquaculture water quality demonstrates that machine learning can effectively predict water quality and model critical parameters like DO, pH, and temperature. The second layer encompasses deep learning time-series models, including LSTM, GRU, Temporal CNN, and Transformer architectures. LSTM was initially proposed to resolve long-term dependency issues, making it ideal for processing delayed-effect phenomena such as the "daytime feeding—nighttime oxygen depletion—subsequent day ammonia nitrogen fluctuation" cycle. For WaterDoctor, many risks do not manifest instantaneously but exhibit precursors: a steepening DO decline slope, a gradual decrease in ORP, delayed ammonia nitrogen peaks post-feeding, or AquaMats biofilm loading approaching its upper limit. LSTM, GRU, and Transformer models can effectively extract these precursors from continuous time-series data. The third layer is a multi-modal predictive model that integrates sensor data, imagery, feeding logs, weather forecasts, engineering operational parameters, and manual inspection records. A 2025 study on IoT and machine learning for continuous aquaculture water quality monitoring suggests that integrating real-time IoT data acquisition with machine learning can robustly support sustainable aquaculture management. WaterDoctor can operationalize this concept within AquaOS: future iterations of the platform will not merely display "DO = 4.2 mg/L," but rather output actionable insights such as, "Orange alert for hypoxia risk in the next 3 hours; primary factors include an abnormal nighttime DO decline slope, excessively high feeding volume over the past 6 hours, and decreased aeration efficiency. Recommendation: proactively increase aeration and reduce the next feeding cycle by 15%." ## II. Computer Vision: Empowering Fish and Shrimp Health, Feeding Behavior, and Biomass Estimation Convolutional Neural Networks (CNNs) represent one of the most successful deep learning architectures in the visual domain. Early convolutional network research by LeCun et al. demonstrated the value of end-to-end gradient learning and convolutional structures in image recognition; the subsequent triumph of AlexNet on ImageNet further catalyzed the advancement of large-scale visual deep learning. WaterDoctor can position Computer Vision as a critical perceptual module within AquaOS. The first application is fish and shrimp health identification.
