Abstract
Aquaculture is increasingly dependent on continuous monitoring and timely management of environmental and production conditions. Variables such as dissolved oxygen, temperature, pH, turbidity, biomass, growth, feeding efficiency, and fish behavior are strongly interconnected, while conventional monitoring approaches often consider these factors separately. Recent advances in artificial intelligence (AI), machine learning (ML), Internet of Things (IoT), and computer vision provide an opportunity to transform aquaculture management from predominantly reactive practices toward predictive and data-driven decision support. Recent reviews have highlighted applications of AI and deep learning in water-quality forecasting, biomass estimation, fish detection and counting, growth prediction, health monitoring, feeding optimization, and behavioral analysis, while also identifying data availability, model generalization, and real-world validation as major challenges [1–3].
This work presents an AI-driven smart aquaculture architecture designed to integrate heterogeneous information sources into a unified monitoring, prediction, and farm-level decision-support system. The proposed architecture combines three complementary data streams: IoT-based environmental measurements, farm production and intervention records, and fish images/video. Instead of treating individual ML models as independent prediction tools, the architecture organizes their outputs within a common risk-fusion and decision-support layer. The objective is to convert model predictions into interpretable operational information that can support actions such as increasing aeration, adjusting feeding strategies, investigating water chemistry, inspecting fish health, and scheduling management interventions. The overall architecture is therefore intended not simply to improve prediction accuracy, but to connect prediction with actionable farm management.
System Architecture and Methodology
The ML layer is organized into five specialized analytical models, each addressing a distinct aspect of aquaculture management. The first model uses a Long Short-Term Memory (LSTM) network to forecast dissolved oxygen (DO), an important indicator of water quality and fish stress. The second is a proposed LSTM model for forecasting pH and identifying potential water-chemistry risk. The third uses XGBoost regression to estimate total fish biomass. The fourth is a dedicated gradient-boosting regression model intended to predict growth and feed-efficiency indicators. The fifth is a planned YOLO/CNN-based computer-vision model for fish detection, counting, size estimation, and behavioral or health assessment. A Vision Transformer (ViT) is also identified as a candidate architecture for the computer-vision component.
The study uses the Data_Model_IoTMLCQ_2024.xlsx dataset [4], containing approximately 6,216 observations collected from January to June. The available variables include average fish weight, survival rate, disease occurrence, temperature, dissolved oxygen, pH, turbidity, temporal information, and intervention indicators. Data preparation includes chronological ordering, validation of temporal fields, construction of Date, Time variables, lagged features, rolling statistics, cyclic encoding of hour, feature scaling for LSTM models, and construction of sequential input windows. Production variables are additionally engineered to support biomass, growth, and feed-efficiency analysis. For the computer-vision component, the proposed system requires appropriately annotated underwater images or video frames containing labels for detection, counting, size, behavior, and health.
Implemented Models and Experimental Results
The currently implemented components provide encouraging initial results. The DO-LSTM prototype uses environmental and temporal information together with intervention records and employs a 70% training, 15% validation, and 15% testing division. Its reported performance is MAE = 0.18 mg/L, RMSE = 0.27 mg/L, and R2 = 0.95, indicating strong predictive performance for the available development data. The predicted DO is subsequently transformed into a risk score, allowing anticipated low-oxygen conditions to trigger an aeration-related recommendation.
The second implemented component is the biomass-XGBoost model, which uses fish age, fish count, feed quantity, environmental variables, survival, disease occurrence, and growth-related indicators. The reported performance is MAE = 12.4 kg, RMSE = 18.7 kg, and R2 = 0.97. The resulting biomass estimate can support feed planning, stock management, harvest forecasting, and production optimization, and is also converted into a biomass-risk indicator for integration with the overall decision-support layer.
Acknowledgement: This work is sponsored by a grant from Carnegie Mellon University Africa to American University in Cairo under Afretec Initiative
References
Aung, T., Abdul Razak, R., & Md Nor, A. R. (2025). Artificial intelligence methods used in various aquaculture applications: A systematic literature review. Journal of the World Aquaculture Society, 56, e13107. DOI: 10.1111/jwas.13107.
Wu, A.-Q., Li, K.-L., Song, Z.-Y., Lou, X., Hu, P., Yang, W., & Wang, R.-F. (2025). Deep Learning for Sustainable Aquaculture: Opportunities and Challenges. Sustainability, 17(11), 5084. DOI: 10.3390/su17115084.
Application of Artificial Intelligence in Aquaculture – Recent Developments and Prospects. (2025). Aquacultural Engineering, 111, 102570. DOI: 10.1016/j.aquaeng.2025.102570.
Baena-Navarro, R., Carriazo-Regino, Y., & Torres-Hoyos, F. (2024). IoT Monitoring Dataset of Water Quality and Tilapia (Oreochromis niloticus) Health in Aquaculture Ponds in Monter��a, Colombia (2024) (Version 1). Mendeley Data. \t "_new"https://doi.org/10.17632/3g2b4sh65m.1