The use of unmanned aerial vehicle (UAV) imagery for specific plant detection has emerged as an important approach. In sporadic poppy detection scenarios, poppies are often intercropped with other crops in highly dispersed and concealed environments, such as courtyard corners and balconies, thereby increasing confusion with background vegetation. Moreover, substantial morphological variations across different growth stages further complicate accurate detection. Existing detection methods often fail to capture the precise contours and texture characteristics of poppies, leading to false positives and false negatives. Accordingly, a poppy detection network based on Position Aware Extraction and Feature Joint Enhancement (PAFJ-Net) is proposed. First, the Position-Aware Multi-Scale Feature Extraction (PAM) module employs global average pooling to extract horizontal and vertical positional information for feature-channel encoding, thereby enhancing the capability for local feature detection. This enables the network to focus on critical poppy features and effectively distinguish them from complex vegetation backgrounds. Second, a Multi-Level Joint Feature Enhancement (MJFE) module is introduced. The module constructs multi-branch heterogeneous convolutional pathways to enhance the network's perception of poppy features across different scales and morphological characteristics while simultaneously strengthening spatial correlations among features. Finally, the Global Context Feature Fusion (GCFF) structure utilises cross-level connected channels to capture richer and more detailed poppy features, thereby alleviating information loss caused by insufficient shallow–deep feature fusion during feature propagation. Experiments were conducted using a UAV-based remote-sensing dataset of containing poppy images. PAFJ-Net achieved an mAP 50 of 86.6%, demonstrating excellent detection performance. To further meet the demands of real-time applications, a lightweight version, termed L-PAFJ-Net, was developed. The proposed network maintains a high mAP 50 of 85.5% while reducing the number of parameters by 40.1% and the computational cost by 51.4%, providing an effective solution for detecting sporadically cultivated poppies in complex environments.
Zhang et al. (Fri,) studied this question.