YOLO-UP: A High-Throughput Pest Detection Model for Dense Cotton Crops Utilizing UAV-Captured Visible Light Imagery

Abstract

Accurate detection of pest species in cotton fields is vital for effective agricultural management and the development of pest-resistant crops. However, achieving high-throughput and precise pest detection in cotton fields remains a challenging task. Although unmanned aerial vehicle (UAV) enable the rapid acquisition of extensive crop images, detecting pests accurately from these images is difficult due to the small size of pests and background interference. This study introduces YOLO-PEST, a novel pest detection model based on YOLOv8n, utilizing a personal cotton field imagery dataset acquired by UAVs. YOLO-PEST incorporates a custom-designed SC3 module to enhance low-level feature extraction and employs the GeLU activation function to address the vanishing gradient issue. Additionally, the model optimizes the neck design to re-duce semantic discrepancies between feature layers, improving small-target detection, and integrates large-kernel separable convolutions to bolster high-level feature processing. Experimental results demonstrate that YOLO-PEST outperforms original YOLOv8n models, with a 3.46 percentage points increase in mAP50, a 5.16 percentage points increase in Precision, and a 7.81 percentage points increase in Recall. YOLO-PEST also shows superior Precision compared to DenseNet and FasterNet, with improvements of 1.14 and 6.13 percentage points, respectively. Overall, YOLO-PEST offers high accuracy and a compact parameter footprint, making it highly effective for pest detection in UAV-acquired crop images.