Enhancing Aquaculture Health : ML-Based Early Warning System for Fish Disease Detection

Aquaculture faces significant challenges in early disease detection, particularly for small-scale farmers with limited resources. This study proposes a novel hybrid CNN-LSTM model with federated learning to enhance fish health monitoring while preserving data privacy. Unlike existing approaches that rely on centralized data or lack interpretability, our framework integrates IoT-based water quality sensing with an attention-based deep learning architecture, enabling accurate and explainable disease prediction. The model processes real-time sensor data (pH, dissolved oxygen, ammonia, temperature, and salinity) to classify fish health into Normal, Stressed, or Diseased states, achieving 93.7% accuracy surpassing traditional machine learning and standalone deep learning benchmarks. Federated learning ensures scalability and privacy by aggregating model updates from multiple farms without sharing raw data, making the solution viable for decentralized aquaculture operations. Key innovations include: (1) a self-attention mechanism that improves detection of critical disease-linked patterns, (2) edge-compatible model quantization for low-cost deployment, and (3) interpretable predictions to aid farmer decision-making. Experimental results demonstrate robust performance across diverse farm conditions, with an F1-score of 0.92, highlighting its practical applicability. This research advances smart aquaculture by addressing critical gaps in accessibility, privacy, and real-time monitoring, offering a sustainable pathway for early disease intervention in resource-constrained settings.