BO-CNN: A deep learning framework with Bayesian hyperparameter tuning for nutrient stress classification in Piper nigrum

Nutrient deficiency (ND) detection in agricultural systems remains a critical challenge requiring accurate and efficient diagnostic tools. This study presents a Bayesian-optimized Convolutional Neural Network (BO-CNN) framework for identifying nutrient deficiencies in black pepper plants using leaf imagery. The approach combines deep feature extraction with Bayesian hyperparameter optimization, enabling efficient parameter selection without exhaustive grid searches. To support this work, we introduce BPNutriDef05, a curated dataset of 10,325 leaf images across five balanced nutrient deficiency classes. Experimental results show that the BO-CNN achieves 97.09% accuracy, outperforming baseline models. This high level of accuracy translates to reliable early detection of nutrient deficiencies, supporting timely corrective measures, optimized fertilizer use, and reduced yield losses, thereby advancing sustainable crop management. Comparative ablation studies further confirm the contribution of Bayesian optimization to model performance. The proposed framework offers both a practical solution for precision agriculture and a methodological advancement in integrating optimization with deep learning for agricultural applications.