Deep learning model optimization for crop prediction and recommendation using Harris Hawks optimization

Abstract These days, farmers often expect crop recommendations and guidance, frequently seeking to agriculture officers for assistance. Fortunately, many websites now provide essential crop datasets that can help recommend suitable crops. The primary aim of this research is to deliver timely and accurate crop recommendations by leveraging deep learning methods and crop data. The novelty lies in combining Harris Hawks Optimizer (HHO) with deep learning models, introducing a practical crop suitability score that simplifies crop prediction and recommendation for farmers without relying on traditional hawks’ behavior calculations. The proposed efficient models like Long Short-Term Memory (LSTM), Bidirectional LSTM, (Bi-LSTM) and Recurrent Neural Network (RNN) were used to analyze environmental factors such as nutrients (N, P, K), pH, rainfall, temperature and humidity. The HHO, a leading swarm-based technique, was employed to enhance soil health analysis and to optimize crop prediction parameters. The Adam Optimizer refines hyperparameters such as learning rate, batch size, iterations, dropout rates, and hidden layers, improving training accuracy. This system recommends crops like rice, wheat, and pomegranate based on a dataset of 22 crop labels with environmental and nutrient conditions, helping farmers optimize resources, reduce waste, and promote sustainable agriculture. HHO was employed to optimize and achieving 100% accuracy for crop prediction. the deep learning models, such as Bi-LSTM achieved the best performance with an accuracy of 90.23%, precision of 90.66%, recall of 90.22%, and f1-score of 90.01%, excelling in capturing bidirectional dependencies in temporal data. The RNN demonstrated solid results with 84.01% accuracy, 87.82% precision, 84.01% recall, and an f1-score of 81.91%, proving its strength in sequential data processing. The LSTM attained 81.36% accuracy, 87.52% precision, 81.36% recall, and an f1-score of 80.37%, effectively handling long-term dependencies. The framework delivers superior results compared to traditional approaches, enabling sustainable farming practices and bridging the gap between deep learning and real-world agriculture.