Abstract: To address the difficulty of achieving precise landing of unmanned aerial vehicles (UAVs) in complex farmland environments, this study designed a machine vision-based precision landing system for agricultural UAVs, with the objective of realizing accurate landing on a UAV airport platform. In this study, the landing marker was first improved to enhance recognition robustness under farmland conditions. A red circular pattern was combined with a double-layer nested ArUco (augmented reality University of Cordoba) marker to form the landing marker. The red circular feature improved the visual distinctiveness of the landing marker in field environments, while the nested ArUco structure enhanced the continuity and stability of marker recognition during landing. The complete marker measured 580 mm × 580 mm, with a first-level ArUco code of 310 mm × 310 mm for long-range recognition and a nested second-level code of 50 mm for precise close-range positioning. To further improve target detection performance, an improved YOLOv9s (you only look once, version 9, small) model was adopted for training and recognition of the landing marker. A total of 126 UAV remote-sensing images were collected over a corn field and divided into training and validation sets at a ratio of 8:2 before brightness augmentation expanded the dataset to 630 images. The use of the improved model helped balance detection accuracy and computational complexity, thereby providing technical support for practical deployment on UAV platforms. A star operation, which mapped low-dimensional features into a high-dimensional space through element-wise multiplication, was embedded in the backbone to strengthen feature representation. The resulting lightweight model comprised 7.28 million parameters and required 29.1 G floating-point operations. After target detection, Zhang’s camera calibration method was employed to establish the coordinate system and obtain the camera parameters. On this basis, the UAV position was calculated by combining the perspective-n-point (PnP) algorithm. By integrating marker detection, camera calibration, and pose estimation, a complete machine vision-based precision landing system was established. To achieve stable flight control during descent, a cascade control strategy combining model predictive control (MPC) and proportional-integral-derivative (PID) was adopted, where MPC generated optimal trajectory references and PID regulated attitude and position for precise trajectory tracking. To evaluate the effectiveness of the proposed system, simulation tests and UAV experiments were carried out. The experimental results showed that the improved YOLOv9s model effectively reduced model complexity while maintaining favorable detection performance. Compared with the original YOLOv9s model, the number of parameters of the improved model was reduced by 49%, and the mean average precision at an intersection-over-union threshold of 0.50 (mAP50) reached 91.96%. In the model evaluation, the improved model attained a precision of 98.6%, a recall of 78.0%, and an F1-score of 82%. These results indicated that the proposed model had good potential for real-time application on platforms with limited onboard computing resources. In addition, the improved landing marker significantly enhanced landing accuracy. The landing errors along the horizontal and vertical axes were reduced from 110.6 mm and 99.8 mm to 64.0 mm and 62.4 mm, respectively, showing that the redesigned landing marker could effectively improve the positional alignment accuracy of the UAV during landing. At heights of 5 to 20 m, its mean detection confidence ranged from 89.5% to 97.3%, whereas that of the conventional single marker ranged from 57.9% to 83.6%. The simulation results further verified the feasibility of the proposed system. When the UAV landed on the platform, the landing errors along the horizontal and vertical axes were 93 mm and 86 mm, respectively, corresponding to 18.6% and 17.2% of the respective side lengths of the 500 mm × 500 mm platform; both individual-axis errors were below 20%. In field experiments, six landing trials were conducted with the original marker, including one failed landing; only the five successful landings were included in the landing-error statistics. The UAV with the improved marker completed five precise landings, and descending from 20 m took about 75 s. This demonstrated that the developed system could meet the basic requirements of precision landing on a UAV airport platform. In summary, this study realized precise landing of UAVs on a UAV airport by developing a machine vision-based precision landing system. The proposed system integrated an improved landing marker, a lightweight target detection model, camera calibration, and visual position estimation, and provided a feasible technical solution for autonomous landing of agricultural UAVs in farmland environments. The findings of this study could provide a reference for the realization of precision landing of agricultural UAVs and support their further application in agricultural operations.