A Dual CNN Framework For Binary and Multiclass Classification of Napier Grass (Penissetum Purpureum): Leaf-Based Diagnosis of Pests and Diseases

Abstract — Napier grass (Pennisetum purpureum) a key livestock forage faces yield losses of up to 50% from pests and diseases threatening tropical food security. This study introduces dual Convolutional Neural Network system based on InceptionV3 to support precision agriculture. The system consists of two components: (1) a binary classification model to distinguish Napier grass from non-Napier entities and (2) a multiclass classification model to identify twelve pest and disease classes along with a healthy class. Trained on 6,400 images which includes 3,200 Napier and 3,200 non-Napier, the binary model achieved 83.90% validation accuracy and 83.90% F1 score. The multiclass model fine-tuned on 4,160 images reached 94.00% accuracy and 94.00% F1 score outperforming VGG16 has 89.21% accuracy, 89.20% F1 score and matching DenseNet which achieved 94.48% accuracy, 94.52% F1 score. Evaluation metrics including accuracy, precision, recall, and F1 score confirm the reliability of both models. Integrated with smartphone imaging this system supports automated monitoring potentially reducing yield losses by 20–30% and aligning with Sustainable Development Goals the SDG 2: Zero Hunger, SDG 12: Responsible Consumption, SDG 15: Life on Land. Larger datasets and advanced architectures could further enhance scalability for smallholder farmers.