JellyNovaNet-JSO: A Hybrid TabNet–BiLSTM Model for IoT-Based Crop Yield Prediction

Abstract

Precise prediction of crop yield is essential for sustainable agriculture, resource maximization, and food security. As the use of IoT and Wireless Sensor Networks (WSNs) gains momentum, huge amounts of heterogeneous and time-series environmental data have become readily available from intelligent greenhouses. Despite this, it is still difficult to obtain meaningful insights from these data due to their high dimensionality, noise, and nonlinear temporal behavior. Traditional machine learning and statistical approaches usually fail to effectively capture static as well as sequential relationships, and most current models are difficult to tune hyperparameters and have problems with dealing with data heterogeneity and do not generalize across dynamic environments. To overcome these shortcomings, this paper introduces JellyNovaNet-JSO, a new hybrid deep learning architecture that integrates TabNet and BiLSTM architectures, designed using the Jellyfish Search Optimization (JSO) algorithm. The model exploits TabNet sparse attention for static feature modeling and the temporal memory of BiLSTM for time-series sensor data. The innovation is in utilizing attention-guided tabular learning with bidirectional temporal modeling, with a metaheuristic optimization layer to perform automatic hyperparameter tuning. Experimental outcomes based on realworld IoT greenhouse data demonstrate that JellyNovaNet-JSO attains MAE of 0.012, RMSE of 0.017, R2 of 0.991, and MAPE of 1.89%, outperforming state-of-the-art CNN-LSTM, Random Forest, and SVM models substantially. In comparison with the prior approaches, JellyNovaNet-JSO enhances prediction accuracy by as much as 25% while ensuring scalability and robustness. This innovation provides a viable, interpretable, and deployable solution for precision agriculture, enabling smarter irrigation, climate control, and yield management.