Xisha watermelon, one variety of selenium-rich economic crops, has been widely cultivated in the semi-arid regions of northwest China. It is critical to estimate its yield for the decision making on the harvesting and market. This paper presents a yield estimation model for Xisha watermelon using drone remote sensing in digital agriculture. Firstly, the watermelon weight was estimated in three steps: object detection with YOLOv8n, contour fitting with the functions from the OpenCV library, and weight estimation of individual watermelon through a volume-to-weight model. Then, two yield estimation strategies were developed. 1) Sampling: the total yield of watermelons over the entire plot was calculated using the average yield within sampled units and the plot area. 2) Global scanning: an overall yield distribution of watermelons was obtained to scan the entire plot using orthoimagery. Finally, a series of field tests was carried out to verify the estimation in plantations. The results reveal that the average accuracy of the detection model was 0.986 using YOLOv8n. Once the number of watermelons exceeded 45, the relative error between the total estimated and the measured weight was less than 1.00%. The speed of sampling was 27.37 m²/s for a 9000 m² field size of Xisha watermelon, approximately 50 times higher than that of global scanning. Compared with global scanning, the sampling-based estimation underestimated the count by 1.77% and the total weight by 5.10%, both of which fall within an acceptable range. Each estimation can be suitable for the specific scenarios of application. The sampling can be expected to provide the higher efficiency for the total field yield. While the global scanning can effectively represent the overall yield distribution of Xisha watermelons in the field. This study provides a new research approach and direction for fruit and vegetable yield estimation in precision agriculture based on UAV remote sensing technology.

To address the limitation of low harvesting efficiency in intelligent mechanized harvesting in standardized orchards, a negative pressure adsorption apple harvesting robot has been designed and developed. The robot is based on a Cartesian coordinate system and incorporates direct negative pressure adsorption picking combined with multi-stage buffering collection methods. Additionally, the proposed model pipeline integrates YOLOv8 and Segment Anything Model for precise apple picking point localization. Finally, field trials of the apple harvesting robot were conducted in a V-shaped layout apple orchard at Experiment and Demonstration Orchard at Tianping Lake. The experimental results showed an apple recognition rate of 90.54%, an overall harvesting success rate of 83.65%, an average picking efficiency of 4.83 s per fruit, and a damage rate of 13.61%. It demonstrates the potential of the robot in improving the efficiency and reliability of automated apple harvesting. At the same time, the results highlight the need to focus on enhancing the robustness of apple recognition algorithms under varying lighting conditions, and reducing apple damage rates by shortening the transport pipeline and optimizing the structure of the collection device. This study provides a promising solution for addressing global challenges in agricultural automation, offering insights into the future optimization of intelligent harvesting technologies.