A Metaheuristic-based Deep Learning Model for Accurate Crop Recommendation Using Agro-Environmental Data

Accurate crop recommendation is essential for sustainable agriculture, yet existing machine learning and deep learning models often struggle with high-dimensional, redundant soil and environmental features, leading to reduced generalization in real-world settings. To address this limitation, this study presents a hybrid Firefly-optimized Autoencoder model that integrates nonlinear representation learning with metaheuristic feature refinement. The Autoencoder compresses NPK nutrient values, temperature, humidity, rainfall, and pH into a compact latent space, while the Firefly Algorithm selects the most discriminative latent dimensions by maximizing feature variance. A deep neural classifier trained on the optimized features achieves highly precise multi-class prediction. The final results tested on a benchmark 22-crop dataset, conducted using Google Colab Pro, shows that the proposed model attains 99.32% accuracy, outperforming advanced methods such as TCN (99.09%), CNN–LSTM federated learning (98.77%), and tuned Random Forest (99.05%). The results highlight the significance of combining latent-space learning with swarm intelligence to deliver a robust, scalable, and high-accuracy crop recommendation system.