Real-time detecting and counting of citrus fruits in hilly and mountainous areas using UAV imagery

Citrus industry can greatly contribute to the agricultural production and rural development in China. However, the conventional fruit counting is confined to the numerous practical challenges in the orchard terrains of the hilly and mountainous. The densely distributed citrus fruits are often partially occluded by branches and leaves, thus difficult to accurately identify and count. Manual counting is also labor-intensive and time-consuming, highly susceptible to the human error, due to the operator fatigue, inconsistent lighting, and the variability in fruit visibility. Such inefficient management cannot fully meet the increasing demand for the modern, efficient, and intelligent agriculture, particularly with the continuous expansion of the orchard scale. In this study, an efficient and light-weight citrus fruit detection and counting model was proposed to balance the high-resolution images that captured by unmanned aerial vehicles (UAVs) and deep learning. The YOLOv8n object detection framework was improved the performance with a series of the targeted architecture under resource-constrained scenarios. Firstly, the GhostConv module was introduced as a lightweight alternative to the standard convolution. Simultaneously, the C3Ghost module was used to reconstruct the C2f feature extraction structure. The computational complexity was significantly reduced to lower the number of the parameters, floating-point operations, and inference time. The detection accuracy was maintained suitable for the deployment on the mobile or edge-computing devices, such as UAV platforms. Secondly, the citrus fruits were distinguished from the complex backgrounds. The Convolutional Block Attention Module (CBAM) was embedded within the neck of the network. This attention mechanism was improved on the key object features, in order to suppress the interference from the irrelevant background. In addition, the Wise Intersection over Union (WIoU) loss function was also replaced to accelerate the network convergence. The accuracy of the bounding box regression was enhanced the overall model performance. Furthermore, the robust and continuous tracking of the individual fruits was achieved across video frames. The Deep Simple Online and Real-time Tracking (DeepSORT) algorithm was integrated into the counting pipeline. Experimental results demonstrate that the GhostConv and C3Ghost modules were reduced the parameters, floating-point operations, and inference time by 42.8%, 37.0%, and 43.2%, respectively. The CBAM module was improved the precision and mean average precision (mAP) by 1.6 and 2.7 percentage points, respectively. The WIoU loss function was led to an additional increase in the precision, recall, and mAP by 1.2, 0.2, and 0.6 percentage points, respectively. Overall, the optimal GCW-YOLOv8n model was achieved in the high accuracy of the detection, with the increases of 4.2, 3.4, and 3.7 percentage points in the precision, recall, and mAP, respectively. The computational demands were shortened the inference time by 75.7%. In fruit counting tests, the better performance was achieved in a peak accuracy of 98.51% and an average accuracy of 95.57% in the offline video analysis. In online real-time counting experiments under varying weather conditions, the stable performance was obtained with the minimal fluctuation in the CA values, indicating its robustness and adaptability. This finding can also provide a reliable technical solution to the real-time estimation of the citrus yield in the intelligent orchard under complex hilly environments.