Comparative Performance Analysis of Deep Learning Models for Planthopper Classification in Rice Plant Ecosystems

Abstrak

Planthoppers are among the most destructive pests in rice ecosystems, capable of causing severe yield losses if not detected early and accurately. Conventional identification methods are often time-consuming and dependent on expert knowledge, limiting their scalability in large agricultural areas. To address this challenge, this study investigates the application of deep learning for automatic planthopper classification using digital images. The objective is to evaluate and compare the performance of four widely used convolutional neural network (CNN) architectures EfficientNetB4, ResNet101, MobileNetV2, and VGG16 under a standardized experimental setup. All models were trained in PyTorch with CrossEntropyLoss, AdamW optimizer, and learning rate scheduling, using mixed precision to improve computational efficiency. The dataset consisted of three planthopper classes: Brown Planthopper (BPH), White-Backed Planthopper (WBPH), and Rice Leafhopper (RLM). Evaluation metrics included accuracy, precision, recall, F1-score, and confusion matrix analysis to assess per-class classification performance. The experimental results show that EfficientNetB4 achieved the best overall performance with 92.13% accuracy and 93.45% F1-score, demonstrating its strong generalization capability. ResNet101 followed closely with 92.80% accuracy and the highest recall of 94.46%, indicating effectiveness in minimizing false negatives. MobileNetV2 provided competitive results (91.85% accuracy) with high efficiency, making it suitable for lightweight, real-time deployment. In contrast, VGG16 achieved lower performance (87.51% accuracy) due to its shallower architecture and limited feature extraction capability. In conclusion, the findings highlight that modern architectures such as EfficientNetB4 and ResNet101 outperform older CNN designs in agricultural pest classification tasks. EfficientNetB4 is recommended for high-accuracy applications, ResNet101 for minimizing missed detections, and MobileNetV2 for resource-constrained environments. This study contributes to the development of intelligent pest monitoring systems, supporting precision agriculture and sustainable rice production.