Loquat flower bud detection in UAV remote sensing imagery based on PSMF-YOLO

Loquat flower buds are typically developed in dense and clustered distributions during their natural growth cycle. Timely and precise thinning is often required for optimal nutrient allocation at the early bud stage. Fruit size disparities can also be prevented for the final commercial quality of the fruits. However, it is still challenging in the accurate identification and spatially localize these small targets under practical agricultural scenarios. The difficulty can be attributed to the small volumetric size of the buds, the clustered growth patterns, and the severe visual similarity in color to the complex surrounding background foliage. In this study, a target detection, termed PSMF-YOLO, was proposed to integrate unmanned aerial vehicle (UAV) remote sensing imagery with advanced deep learning. A loquat flower bud dataset was constructed using high-resolution UAV images in orchard environments. An improved detection was developed using the YOLOv11n architecture as the baseline to effectively balance computational efficiency and detection accuracy. Structural modifications were sequentially implemented to systematically optimize the network for the microscopic targets. (1) A P2 small-object detection layer was introduced into the architecture. Shallower network layers were used to preserve high-resolution spatial features, thereby detecting minute and densely packed targets. (2) Space-to-Depth Convolution (SPD-Conv) module was incorporated into the feature extraction network. Space-to-depth transformation retained information, thus preventing the unnecessary loss in the conventional stride convolutions. Detection accuracy was significantly improved for low-resolution images and small objects. (3) The C2PSA module was combined with a multi-scale dilated attention (MSDA) mechanism. Cross-level perception was achieved in multi-scale dilated convolutions and dynamic feature fusion. The network was established for more effective global context modeling. (4) A Focaler-DIoU bounding box regression loss function was employed to effectively alleviate the sample imbalance in small-object detection tasks. The precise localization of bounding box regression was improved for the overall robustness of the model when deployed under complex orchard conditions. Experimental results demonstrate that the superior performance of the PSMF-YOLO model was achieved, with 4.9, 2.3, and 4.2 percentage points in mean average precision (mAP_0.5), precision (P), and recall (R) metrics, respectively, compared with the baseline YOLOv11n model. In conclusion, the PSMF-YOLO can provide a highly effective and accurate algorithmic solution for UAV small-target detection tasks in agricultural remote sensing. Consequently, this finding can offer robust technical support in an intelligent and precise orchard.