Abstract
INTRODUCTION: Licorice has significant medicinal and ecological importance. However, prolonged overharvesting has resulted in twofold damage to wild licorice resources and the ecological environment. Thus, precisely determining the distribution and growth condition of wild licorice is critical. Traditional licorice resource survey methods are unsuitable for complex terrain and do not meet the requirements of large-scale monitoring. METHODS: In order to solve this problem, this study constructs a new dataset of wild licorice that was gathered using Unmanned Aerial Vehicle (UAV) and proposes a novel detection network named ALPD-Net for identifying wild licorice. To improve the model's performance in complex backgrounds, an Adaptive Background Suppression Module (ABSM) was designed. Through adaptive channel space and positional encoding, background interference is effectively suppressed. Additionally, to enhance the model's attention to licorice at different scales, a Lightweight Multi-Scale Module (LMSM) using multi-scale dilated convolution is introduced, significantly reducing the probability of missed detections. At the same time, a Progressive Feature Fusion Module (PFFM) is developed, where a weighted self-attention fusion strategy is employed to effectively merge detailed and semantic information from adjacent layers, thereby preventing information loss or mismatches. RESULTS AND DISCUSSION: The experimental results show that ALPD-Net achieves good detection accuracy in wild licorice identification, with precision 73.3%, recall 76.1%, and mean Average Precision at IoU=0.50 (mAP50) of 79.5%. Further comparisons with mainstream object detection models show that ALPD-Net not only provides higher detection accuracy for wild licorice, but also dramatically reduces missed and false detections. These features make ALPD-Net a potential option for large-scale surveys and monitoring of wild licorice resources using UAV remote sensing.