Scheduling-driven attention CNN–BiLSTM framework for aquaculture disease prediction using water quality parameters
TL;DR
An IoT based deep learning system for the prediction of important water quality parameters based on multi-sensor time-series data, such as pH, Dissolved Oxygen and Oxidation–Reduction Potential, and the framework is expanded to include disease risk categorization according to forecasted water quality which will enable early warning and proactive management.
Abstract
In intelligent and sustainable shrimp aquaculture, rapidly changing variation environmental conditions can affect the accurate prediction of water quality parameters, therefore this is essential to the productivity and survival rates. The paper introduces an IoT based deep learning system for the prediction of important water quality parameters based on multi-sensor time-series data, such as pH, Dissolved Oxygen (DO) and Oxidation–Reduction Potential (ORP). To effectively model the spatial feature correlation and bidirectional temporal dependency, a hybrid attention-based Convolutional Neural Network with Bidirectional Long Short-Term Memory (attention CNN–BiLSTM) model is used. To enhance training stability and convergence efficiency two learning rate scheduling strategies ReduceLROnPlateau and CosineAnnealingLR are integrated and comparatively analysed. The experimental results show that the scheduler-enhanced attention CNN–BiLSTM model achieves the best performance over traditional machine learning and deep learning models. The proposed model prediction performance for pH is an RMSE of 0.0957 and R2 of 0.9813 and for DO the prediction performance is R2 > 0.998. ReduceLROnPlateau demonstrates a strong robustness for nonlinear parameter estimation as the integration of ReduceLROnPlateau, with MAE of 4.7582 and 1.8201, does not drastically change the prediction of ORP. Moreover, CosineAnnealingLR brings down the training time by about 40% and retains the prediction accuracy comparable with the state of the art. The results show that the proposed framework can achieve significant improvement of the prediction accuracy, stability and computational efficiency, which is well suitable for real-time IoT-based aquaculture monitoring system. Furthermore, the framework is expanded to include disease risk categorization according to forecasted water quality which will enable early warning and proactive management.