Convolutional Neural Network Based on Transfer Learning for Pest Classification in Chili Plant

Pest infestation is a major cause of productivity loss in chili cultivation, requiring accurate and automated pest identification systems to support precision agriculture. This study proposes a lightweight MobileNetV2-based transfer learning model for multi-species chili pest classification using a validated Indonesian digital image dataset. The dataset consists of 2,450 images representing seven major chili pest species, collected from publicly available web sources and complemented with field images acquired under natural conditions. All images underwent manual curation, quality control, taxonomic verification, and annotation validation by experts to ensure dataset reliability. The dataset was divided into training and testing in a ratio of 80:20, and data augmentation techniques were applied to improve model generalization. MobileNetV2 was compared with ResNet50 and EfficientNetB0. All models were trained for 50 epochs using the Adam optimizer and the categorical cross-entropy loss function. Model performance was evalu-ated using accuracy, precision, recall, and F1-score metrics. Experimental results show that MobileNetV2 achieved the highest accuracy (98.29%), outperforming ResNet50 (97.84%) and EfficientNetB0 (97.56%) while maintaining low computational complexity suitable for resource-constrained environments. This study contributes (1) a validated Indonesian multi-species chili pest dataset, (2) an optimized MobileNetV2 transfer learning model for accurate pest classification, and (3) a practical Streamlit-based pest monitoring system for precision agriculture.