AI-Assisted Detection of Major Pests in Red Chili Plants: Accuracy–Efficiency Trade-offs of YOLO11 Variants

Abstract

Reliable and timely pest detection is essential for protecting red chili plants and supporting effective pest monitoring. However, automated detection remains challenging because agricultural pests are often small, densely distributed, visually similar, and observed against complex natural backgrounds. This study evaluates the accuracy–efficiency trade-offs of three YOLO11 model variants—YOLO11n, YOLO11s, and YOLO11m—for detecting major pests in red chili plants. Experiments were conducted using 4,957 annotated images containing 10,561 bounding boxes representing four pest classes: Bemisia tabaci, caterpillar, Myzus persicae, and thrips. All models were initialized with pretrained weights and trained under a controlled experimental protocol using identical data splits, an input resolution of 640 × 640 pixels, a maximum of 100 epochs, and early stopping. Detection performance was evaluated using precision, recall, mAP50, and mAP50–95. YOLO11n achieved the highest test mAP50 of 0.700 and mAP50–95 of 0.310 with only 2.591 million parameters. YOLO11s obtained an mAP50–95 of 0.298 at 81.31 frames per second, whereas YOLO11m achieved 0.283 at 34.23 frames per second despite producing the highest recall of 0.754. The results demonstrate that increasing model capacity does not necessarily improve pest localization performance. Among the evaluated variants, YOLO11n provides the most favorable balance between detection accuracy and computational efficiency, indicating its potential as a lightweight model for real-time AI-assisted pest monitoring in red chili cultivation.