Grassland degradation has posed a serious threat to ecological stability, forage yield, and the sustainable animal husbandry in recent years. Agricultural robots (such as unmanned aerial vehicle (UAV) platforms) can be expected to monitor and reseed the degraded grassland patches for ecological restoration. However, existing UAV operations have been limited to the low accuracy of patch identification and patch positioning, due to easy target loss or unstable tracking. In this study, a computer vision was proposed to integrate the advanced object detection and tracking algorithms for the efficient detection and positioning of grassland patches. The YOLO series model was used to detect the patch. An adaptive feature pyramid network was introduced to enhance multi-scale feature fusion, thereby improving detection accuracy in complex grassland backgrounds. Lightweight convolution was used for channel compression and nonlinear activation, while the expression of deep features was optimized for the computational efficiency. The Swin Transformer sliding window self-attention mechanism and residual convolution structure were adopted to establish long-range dependencies between patches and the grassland background within the local receptive field. Input RGB patch images were converted to the YUV color space. The single-channel luminance map was extracted as the input to enhance the luminance contrast of the grassland and patch texture structures. In terms of temporal consistency and target tracking, the DeepOCSORT algorithm was combined to perform multi-target tracking of patches in consecutive image sequences. The results showed that the improved SCTD-YOLO detector significantly outperformed the baseline YOLOv8. Ablation experiments indicated that the better performance was achieved to combine the different modules. The accuracy and recall rate increased to 94.7%, and 86.2%, respectively, using the improved SCTD-YOLO model, indicating the best performance; The average precision and mAP increased by 2.4 percentage points and 2.2 percentage points, respectively. The high robustness was also obtained to significantly reduce the missed detection and false rates under different lighting conditions, complex backgrounds of grassland texture, partial patch occlusion by plants, and significant changes in patch size. In multi-target tracking, the algorithm with the DeepOCSORT tracker was effectively maintained the consistency of patches in multi-frame sequences, with an IDF1 of 85.15%, an MOTA of 88.2%, and an MOTP of 86.71%, significantly reducing false detections and ID switches. DeepOCSORT model was still stably maintained the target trajectories in the scenarios of rapid camera panning, grass leaf shaking caused by wind, and dynamic background, significantly improving the reliability of long-term monitoring. Finally, the latitude and longitude coordinates were predicted after image modeling and external parameter calculation, compared with high-precision RTK-GPS measurement coordinates. The spatial positioning error was calculated using the Vincenty formula. The improved SCTD-YOLO and stable tracking algorithm were combined to significantly reduce the overall positioning error, with an average error of 0.316 1 m and an error range of 0.278 m to 0.423 m. The higher positioning accuracy and error stability were found in all test samples, effectively improving the overall consistency of target detection, tracking, and positioning. The overall positioning accuracy was fully met the requirement of less than 0.5m in intelligent reseeding of degraded grassland patches. The SCTD-YOLO and DeepOCSORT framework were integrated for patch detection, tracking, and positioning in complex multi-scenario and multi-interference environments for agricultural robots. The high efficiency and reliability of grassland patch reseeding can be expected for the decision-making generalization of agricultural robot vision in different agronomic models and environmental conditions, sustainable grassland, resource conservation, and ecological restoration. Further exploration can be conducted on seasonal monitoring and adaptive reseeding using multi-source remote sensing data.

Multi-unmanned aerial vehicle (UAV) cooperative operations in mountainous orchards face four critical challenges: prohibitively high computational complexity due to large-scale 3D path planning, slow algorithmic convergence in rugged terrain, geometrically unstable flight path generation, and uneven energy consumption leading to premature battery depletion among heterogeneous drone fleets. To address these interconnected issues, this study proposes an integrated computational framework synergistically combining an Improved Artificial Hummingbird Algorithm (IAHA) with a recursive energy-balancing model. Empirical validation was conducted in a lychee orchard (23°9' N, 113°22' E), where centimeter-resolution digital elevation models (DEMs) were constructed via UAV photogrammetry at a spatial resolution of 5.41 cm/pixel. These DEMs enabled precise georeferencing of 57 mission-critical waypoints, comprising one depot at 25 m elevation and 56 operational points dynamically maintained at 10 m above heterogeneous tree canopies to accommodate variable crown heights. The core innovation lies in the IAHA, which significantly enhances the foundational artificial hummingbird algorithm through three algorithmic advancements: (1) Hybrid population initialization strategically injects one high-quality solution generated via nearest-neighbor greedy heuristic into a Monte Carlo-sampled population at a 1:99 ratio, simultaneously seeding solution quality while preserving population diversity. (2) Adaptive parameter control employs cosine annealing for dynamic step size modulation across iterations and exponential decay for directional perturbation coefficients, enabling nuanced exploration-exploitation tradeoffs. (3) Periodic local optimization automatically triggers a 2-opt refinement every 10 generations to eliminate topological path crossings, ensuring physically flyable trajectories. To counteract search stagnation, a visit-table-driven guidance system implements inverse probability selection to prioritize visitation of underexplored spatial regions, complemented by a tournament-based migration strategy that replaces 10% of the poorest-performing individuals per generation to maintain evolutionary pressure. For multi-UAV energy equilibrium, physics-based rotor power consumption models were extended through hierarchical recursive segmentation with bidirectional boundary adjustment. This technique recursively partitions waypoint clusters while co-optimizing segment boundaries to balance cumulative energy demands, subsequently employing weight-aware path aggregation that explicitly accounts for UAV payload differentials (e.g., spraying modules vs. sensors). The model enforces a strict upper bound of 5% total energy variance across the fleet—critical for operations in mountainous terrain where elevation changes exponentially impact power draw. Validation experiments demonstrated IAHA's superiority in single-UAV path optimization. Compared to five benchmark algorithms (AHA, ACO, GA, SA, PSO), IAHA achieved path length reductions of 25.7%, 2.0%, 20.6%, 12.74%, and 19.19%, respectively, in real-world orchard flights. Significantly, it maintained a path length standard deviation ratio of merely 1.24%, highlighting exceptional solution stability across trials. In multi-UAV deployment scenarios using a heterogeneous fleet (2×DJI Mavic 3 + 1×DJI Phantom 4 varying in thrust-to-weight ratios), the integrated IAHA-energy framework achieved an unprecedented energy consumption Coefficient of Variation (CV) of 1.16%. Concurrently, task completion time decreased by 56.52% compared to sequential single-UAV operations. The framework's efficacy stems from synergistic algorithmic innovations: dynamic parameter adaptation prevents premature convergence while maintaining exploration capability; visit-table guidance sustains population diversity; and recursive energy segmentation enables near-equitable workload distribution—the latter proving particularly critical for operations in mountainous topography where altitude variances induce nonlinear power demands. Statistical analysis confirms these components interact multiplicatively rather than additively. Theoretical contributions include establishing convergence guarantees for the modified hummingbird dynamics and proving recursive segmentation's computational tractability under constrained energy variance bounds. This study validated the feasibility of the algorithm framework through simulation. In future field flight tests, we will focus on verifying the engineering applicability of path planning and energy consumption optimization, including: three-dimensional trajectory tracking accuracy in complex mountainous environments, consistency between actual battery consumption and model predictions, and balance in flight duration among heterogeneous fleets.

Abstract: Grassland restoration has been a critical initiative of ecological engineering in China. Drones have effectively facilitated the reseeding in large-scale and diversified terrains at present. However, the accuracy of gap localization has been confined to the high workload during drone operations. This study aims to propose a novel approach for gap localization of degraded grassland, termed YOLOFG (YOLO for Gap), using super-resolution reconstruction and Transformer technology. Firstly, the challenges were addressed on the limited quantity and widespread distribution of gap samples in grassland scenes, in terms of sample acquisition. Drones were also employed to combine the fieldwork and simulation. A total of 132 and 282 aerial images were collected to constitute a mixed gap dataset. Secondly, three data augmentations were selected: random angular rotation, exposure adjustment, and brightness modulation, in order to enrich the mixed gap datasets. The objective of these techniques was to enhance the generalization of the model. Furthermore, 1 242 enhanced gap images were annotated to utilize the open-source labeling tool LabelImg. The annotated data was stored, according to the format specifications of the Pascal VOC dataset. Subsequently, the blurred texture was targeted in the unmanned aerial vehicle (UAV) aerial photography dataset. The modifications were made to the YOLOv5s model, thereby completing the training of the data. Finally, the spot anchor frame information was utilized to obtain from the recognition model. The imaging principles of the gimbal camera were used to obtain the conversion relationships among multiple coordinate systems. The real-time attitude information was selected from the UAV. A geodetic coordinate resolution model was established for the aerial photography targets, thus enabling precise localization of the target gaps. According to the theoretical framework of YOLOv5s target detection, a cascaded module was introduced to extract the feature texture in the neck network. The high-resolution feature maps were generated with enhanced textures, thereby enabling the model to focus on intricate feature textures. There was a great decrease in the significant scale variations and blurred textures in the drone-captured gap images. Furthermore, ShuffleNetV2 was selected as the backbone network for YOLOv5s. The number of model parameters was reduced to integrate a highly interactive Transformer self-attention structure into the backbone network. More comprehensive mid-to-high-level visual semantic information was obtained from the drone imagery. The information interaction was also enhanced among pixel blocks. More differentiated features were ultimately extracted among pixels to enhance the capture precision for the gap edge pixels. The experimental results demonstrated that the YOLOFG model achieved a mean Average Precision (mAP) of 96.57%, indicating an improvement of 11.86, 9.65, and 6.82 percentage points, respectively, compared with the YOLOv4, YOLOv7, and YOLOv8 models. The parameter count was approximately 6.24 M, with a model size of 13.9 MB. Compared with the YOLOv5s, the model size remained comparable, yet the precision was enhanced by 3.84 percentage points, and the parameter count was reduced by approximately 11.2%, thus facilitating the deployment. With an inference time of 39.6 FPS(frames per second), the improved model outperformed the baseline networks, including YOLOv5s, YOLOv8, and YOLOv4, thus demonstrating the high efficiency and robustness. Finally, 54 target points were measured with an average positioning error of 0.440 4 m between the center points of the targets and the anchor boxes. In summary, the finding can offer a better trade-off between speed and accuracy, in order to realize the rapid and efficient identification of gap distributions across different communities. Additionally, the precise localization requirements were fully met for the drone-based reseeding in grassland gaps. Thus, robust technical support was provided for the subsequent vegetation restoration and reconstruction in degraded grasslands.