This study aims to address two critical challenges faced by agricultural robots in protected tomato cultivation environments: significant light variations and unstable GPS signals that lead to difficulties in rail finding and precise alignment. To enhance autonomy and operational efficiency in such environments, we propose an integrated method combining two-dimensional (2D) LiDAR and an improved YOLOv8-seg network for rail-region segmentation and alignment control. The proposed method consists of three main steps. First, during inter-row navigation, a 2D LiDAR scans the crop rows to obtain point cloud data. An improved DBSCAN clustering algorithm, which adapts to local density variations, is employed to process the point cloud data. This algorithm dynamically adjusts the ε and MinPts parameters based on the local and global density distributions of the point cloud. Combined with Theil-Sen estimation, it fits row lines and enables real-time extraction of the distance and yaw angle of the robot relative to the crop rows. This information is used to control the robot’s navigation along the inter-row path with centimeter-level accuracy. Second, as the robot approaches the target rail region, a side-mounted RGB camera captures images of the rail area. An enhanced YOLOv8-seg network is utilized to identify and segment the rail region. The network incorporates several improvements over the original YOLOv8-seg architecture. A multi-scale fusion module (MSC2f) is introduced to replace the original C2f module. This module uses parallel 3×3 and 5×5 convolutions to capture fine details in bright regions and larger structures in shadowed areas. Additionally, efficient channel attention (ECA) is embedded to suppress glare effects by dynamically adjusting channel weights. To further reduce computational complexity and improve inference speed, depthwise separable convolution and GhostConv modules are implemented. These modifications enable the network to achieve an mAP₅₀ of 96.43%, representing a 1.3 percentage point improvement over the original YOLOv8-seg, while maintaining a lightweight model with only 1.41 million parameters and a frame rate of 96 frume/s on an RTX 4060 GPU. Finally, based on the segmentation results, the least squares method is used to fit the rail centerline. The lateral deviation between the robot and the rail centerline is calculated, and the robot’s pose is adjusted accordingly. The robot performs a 90° in-place turn and enters the rail region once the deviation is within the acceptable range. Comprehensive experiments were conducted in a multi-span greenhouse at the Lvgang Modern Agricultural Research Institute in Suqian City, Jiangsu Province, where tomato plants were cultivated in coconut coir bags with rail laid between ridges (rail width: 550 mm, adjacent rail spacing: 1760 mm, total rail length: 50 m). A total of 1216 rail images were collected under four typical light conditions (normal light: 8000-25000 lx, strong light: >25000 lx, low light: 2000-8000 lx, local strong light: >25000 lx in local areas), These raw images were first divided into training, validation, and test sets, and then enhanced using techniques such as dynamic blur, mirror flipping, and brightness adjustment. The final dataset consisted of 2 653 training images, 663 validation images, and 332 test images. Experimental results demonstrate that the proposed method achieves an average lateral deviation of 2.16 cm and a maximum deviation of 4.95 cm under different lighting conditions, with a 100% success rate in 130 trials. The average distance error from the robot center to the crop row, as determined by LiDAR clustering and fitting, is 2.67 cm, and the average heading angle error is 0.16°, validated against data from an Xsens MTi-670 inertial measurement unit (IMU). The improved DBSCAN clustering algorithm achieves a clustering accuracy of 98.07%, which is 1.05 percentage points higher than the traditional DBSCAN algorithm, with a single-frame processing time reduced from 4.6 ms to 1.7 ms (a 63% speedup). In conclusion, the proposed method provides an effective solution for rail-region segmentation and alignment control in protected tomato cultivation environments. By integrating 2D LiDAR and an improved YOLOv8-seg network, the method achieves high precision and robustness in navigation and alignment tasks under varying lighting conditions, offering a promising technical pathway for the development of autonomous agricultural robots capable of operating efficiently in protected environments and potentially extendable to other protected crops such as strawberries and cucumbers, contributing to the advancement of intelligent protected agriculture.

Plant protection is moving toward mechanization and intelligence in the process of modern agriculture in recent years. Autonomous operation of unmanned aerial vehicles (UAV) can reduce the quality errors from the visual images or manual experience, particularly for labor cost saving. Among them, trajectory optimization is one of the most important steps during the autonomous operation of plant protection UAVs. A reasonable algorithm of trajectory optimization can be expected for the UAVs to more efficiently and safely realize the tracking operations. The current algorithm of mainstream optimization focuses mainly on trajectory accuracy and algorithm efficiency in the tracing scene. But there is an overshoot of state quantity and serious trajectory deviation in the special scene of agricultural plant protection, due to complex situations, such as line feed and turn. Therefore, it is very necessary to optimize the trajectory suitable for the operation scenarios of plant protection. Simple time allocation cannot fully meet the requirements of trajectory optimization in the traditional minimum-SNAP algorithm, such as the large trajectory deviation error. In this study, an improved minimum-SNAP algorithm was proposed for trajectory optimization in the operation scenarios of plant protection. Firstly, the real-time kinematic (RTK) integrated navigation system carried by the plant protection UAV was used to collect the topographic data of the plant protection plots. Then the data was imported into the robot operating system (ROS) in the central controller. The surface fitting was also utilized to obtain the three-dimensional spatial map. The cattle ploughing method was selected to plan the initial operation path. Secondly, the task path was optimized using an improved minimum-SNAP algorithm. The cost function of time allocation was also constructed using the state parameter variable of the initial operation path. In addition, the change rate of the initial path baseline angle was introduced as a reference factor, in order to obtain the optimal time allocation after optimization. Thirdly, the coefficient optimization cost function of the minimum-SNAP algorithm was reconstructed using the time allocation information. The gradient penalty factor of position offset was then added to the optimization function. The trajectory expression of the piecewise polynomial was then obtained to solve the coefficient of the trajectory polynomial after optimization. Finally, the UAV position control period was substituted into the piecewise trajectory polynomial for the track point, which was taken as the location expectation to control the UAV motion. A simulation test was carried out to preliminarily verify using the Matlab platform. The initial path was taken as the "U" type round-trip operation track in two-dimensional space. The average acceleration of the improved algorithm was 0.390 m/s², which was 19.75% lower than the traditional. Furthermore, the average trajectory offset was 0.051 m/s², which was 66.88% lower than the traditional. The physical test site was selected as the mulberry forest on the slope with the complex terrain. The test results show that the improved algorithm reduced the average acceleration and deviation by 7.82%, and 45.56%, respectively, compared with the traditional under the same running time. An outstanding inhibition was achieved in the trajectory deviation with the reduced overshoot during line feed turning. The finding can also provide a technical and theoretical basis for the UAV trajectory optimization in the scenario of plant protection.

Accurate trajectory tracking is often required for autonomous wheeled robots in orchard and forestry environments. However, the control performance is also constrained by the external disturbances and parameter uncertainties caused by uneven terrain, soft soil, and frequent steering maneuvers, particularly when the wheel slip and steering-induced slip occur in practical agricultural scenarios. This study aims to propose a prediction-accuracy-enhanced explicit nonlinear model predictive control (ENMPC) algorithm integrated with a nonlinear disturbance observer (NDOB) for autonomous orchard wheeled robots. An extended kinematic model was firstly established to incorporate the typical disturbances in the field environments (including the wheel slip and steering slip) into an ideal kinematic framework. The trajectory tracking error dynamics were approximated over a finite receding horizon using a Taylor series expansion. An explicit analytical control law of the ENMPC was derived offline under the assumption that external disturbances were measurable. The requirement was fully met for the nonlinear optimization rather than the online one. As such, the controller significantly reduced the computational complexity for the real-time performance, which was critical for the embedded agricultural robotic systems with the limited computing resources. A nonlinear disturbance observer was designed to compensate for the external disturbances in real time in order to enhance the robustness against time-varying disturbances and modeling uncertainties. Furthermore, the disturbance estimates were treated as the additional corrective inputs using conventional feedforward compensation. By contrast, the approach directly integrates the disturbance estimation into the output prediction of the ENMPC. Therefore, the predicted system outputs were explicitly accounted for the disturbances in the prediction horizon, thereby achieving the zero steady-state offset in trajectory tracking. The closed-loop stability of the NDOB-ENMPC controller was analyzed under bounded disturbances. Simulation studies were conducted to evaluate the performance under multiple disturbance scenarios, including the constant disturbances, time-varying disturbances, and combined slip. The results demonstrate that the better performance was achieved under complex and uncertain working conditions. The NDOB-ENMPC algorithm effectively suppressed the disturbance influences. The superior accuracy and robustness were achieved in the trajectory tracking, compared with the conventional nonlinear model predictive control. Field experiments were further performed using an autonomous orchard wheeled robot in a forest grassland environment. Experimental results show that the NDOB-ENMPC reduces the maximum absolute lateral and longitudinal tracking errors by 39.42% and 49.01%, respectively, compared with a standard NMPC algorithm. While the mean absolute errors are also reduced by 29.45% and 44.01%, respectively. In addition, the average computation time is reduced by 97.47%. The maximum absolute lateral and longitudinal errors are reduced by 17.86% and 37.64%, respectively, and the mean absolute errors by 16.41% and 20.59%, respectively, compared with a Feedforward Compensated NMPC algorithm. While a 97.57% reduction is obtained in the average computation time. The NDOB-ENMPC algorithm fully meets the requirements of the real-time implementation, high tracking accuracy, and robustness for the autonomous navigation of the wheeled robots in complex orchard environments. The finding can provide the computationally efficient solution to deploy the optimal control strategies on the low-cost agricultural robotic platforms. There is also strong potential for practical applications in intelligent orchard and forestry environments.

Autonomous navigation has been widely used in complex orchard environments. Global Navigation Satellite System signals can be obtained after navigation. Specifically, Real-Time Kinematic solutions are frequently interrupted by dense canopy occlusion, leading to signal loss. However, the conventional laser Simultaneous Localization and Mapping can often suffer from substantial drift, due to the discrete, unstable geometric features and high local environmental similarity. Particularly, the feature extraction can rely heavily on the ground or planar structures. Such features are often absent, ambiguous, or highly repetitive in agricultural settings. The reliable localization and mapping are often required for the intelligent spraying agricultural robots in precise positioning tasks. In this study, a robust and high-precision framework was proposed for LiDAR-inertial positioning and mapping. The stable features of the tree trunk were also extracted within the arboreal environment. Thereby, the more robust observational constraints were added to the Kalman filter. Specifically, the Euclidean range was employed to effectively filter out the distant background noise. Distance smoothness criteria were used for the spatial continuity of the point segments. The concavity-convexity consistency was evaluated to verify the external shape of the trunks. The curvature regularity was then assessed for the properties of the cylindrical geometry. A series of tests was carried out to calibrate the experimental parameter. The valid features were effectively distinguished among inclined trunks in the quasi-dynamic orchard. Unstable features were removed to thereby reinforce the geometric constraints in the frame-to-frame registration. A tree loop descriptor was then introduced to optimize the global consistency on the back end. The spatial distribution of the trees was obtained for their topological relationships. Geometric triangle descriptors were constructed using inter-tree distances and rotation-invariant radial binary descriptors. Effective loop closures were identified and then registered using a dual-distance screening mechanism. The reliable loop detection was realized to eliminate the cumulative errors. A series of experiments was also conducted on both the publicly available BotanicGarden dataset and a self-collected nursery dataset. The quantitative analysis was utilized to take the Root Mean Square Error of the Absolute Pose Error as the primary metric, where the high-precision differential signals served as the ground truth trajectory. The results demonstrate that the better performance significantly outperformed the current state-of-the-art algorithms in complex orchard environments. Specifically, the trajectories also exhibited a high degree of overlap with the ground-truth paths on the BotanicGarden dataset. Statistical analysis indicated that the diverse challenging environments were also optimized, including the dense tree scenarios, riverside areas with sparse vegetation, and mixed tree-lawn zones. Consequently, the localization accuracy was improved by approximately 34.9% and 47.1%, respectively, compared with the widely-used FAST-LIO2 and the FAST-LIO-SAM algorithm. The self-collected dataset was utilized to specifically design for the different scenarios. Five scenarios were represented by different vegetation types and spatial structures, including a mixed planting area of Red Maple and Crape Myrtle, a complex mixed zone of Cactus Balls and Chinese Rose trees, a densely vegetated area of Begonia trees, a sparse potted Sweet Osmanthus section, and a high-density Cactus Ball planting zone. Stable trunk features were extracted to provide the robust observational constraints. While the tree-based descriptor was screened for the loop closure detection, thereby significantly enhancing the localization precision. As a result, the high accuracy was then improved by approximately 42.4% over FAST-LIO2 and 48.3% over SC-LIO-SAM. The robot paths were estimated to be closely aligned with the high-precision ground-truth references, with no significant cumulative drift during long-term operation. In conclusion, the stable trunk geometric features were fused with a topology-aware loop closure. The robustness and accuracy of the localization were significantly enhanced under complex orchard environments in precision agriculture.

Accurate detection of small-volume targets, such as tree crowns and pedestrians, remains a significant challenge for the orchard robots in complex nursery environments. The high performance of detection is often required for agricultural settings. In this study, an enhanced version of the VoteNet model was proposed for the object detection of 3D point clouds. Three major modifications were utilized, according to the original VoteNet architecture. Firstly, the original module of asymmetric feature extraction was replaced by a symmetric feature extraction module (SFEM). A U-Net-like encoder-decoder structure was also adopted to integrate the multi-scale features. The symmetric design was used to more effectively fuse the low-level spatial and high-level semantic information. Thereby, the finer geometric features were preserved to detect the small objects. Secondly, a reverse attention feature fusion module (RAFFM) was introduced at the skip connections. The local feature representation was then enhanced using self-attention mechanisms and trilinear interpolation. The fine-grained structures were effectively highlighted to improve the feature consistency at the different scales. The network was focused on more discriminative regions of the small targets. Thirdly, a center point discrepancy loss (CPD-Loss) was incorporated to minimize the spatial offset between predicted proposal centers and ground-truth bounding box centers. The high accuracy was obtained for the object localization. This additional loss term regularized the voting process, leading to more stable cluster formation around object centroids. A series of experiments was conducted to validate the approach. A nursery point cloud dataset was constructed with three types of trees and pedestrians under 1147 scenes, according to the KITTI format. The dataset was partitioned into 60% for training, 20% for validation, and 20% for testing, in order to ensure fair evaluation. The good performance of the model was achieved, with an average recall (AR) of 88.06% and a mean average precision (mAP) of 55.05% at an IoU threshold of 0.25, which significantly outperformed the baseline VoteNet by 9.05% and 22.47% in AR and mAP, respectively. Particularly, the small objects were detected dramatically: The average precision (AP) for Ilex cornuta var fortunei increased from 4.40% to 28.77%, respectively, indicating a 6.5-fold enhancement. Each component was individually contributed after the ablation test. The SFEM provided more discriminative feature learning for the objects with sparse point distributions; The RAFFM effectively aggregated the contextual features to preserve the geometric details, leading to the higher recall for the pedestrians and small tree crowns; And the CPD-Loss further increased the accuracy of bounding box regression, in order to stabilize the training convergence. It was also compared with the traditional target detection network.The good performance was achieved under the nursery environments. The 3D detection framework can offer a robust perception solution for the autonomous orchard robots. More accurate environmental detection was also obtained in complex nursery settings. The detection facilitated the small-volume targets at various precision levels of the agriculture applications, such as targeted spraying, growth monitoring, and safe navigation around pedestrians. This work can contribute to the advancement of intelligent agricultural systems. The finding can provide a reliable 3D perception technology for efficient and autonomous nursery operations. Future research can optimize the network architecture for real-time performance, and then extend into the rest agricultural scenarios.

This study aims to enhance the accuracy, real-time performance, and practicality of real-time target detection in spray robots in forest and fruit orchards. A novel approach was introduced using a lightweight PointNet deep learning network. Efficient detection of target was realized during spray under complex natural environments. Several key steps were consisted as follows. Firstly, the target was extracted from the original point cloud. Preprocessing techniques included the region extraction downsampling, ground segmentation, and an improved density-based spatial clustering of application with noise. The target points were isolated to facilitate the subsequent recognition. After extraction, the target point cloud was transformed into the format suitable for the input into the point cloud recognition network. The moving least squares upsampling was carried out to collect the appropriately prepared data. The recognition network of point cloud was constructed to specifically designed for the forest and fruit tree environments. The residual module was introduced to enhance the cyclic pruning algorithm into the PointNet network. The representation power of the network was improved to capture more complex patterns and features of target objects. The cyclic pruning algorithm was used to refine the network architecture, and then remove unnecessary connections, in order to reduce the computational overhead with high performance. The experimental results demonstrate that an impressive accuracy of 89.7% was achieved in the widely-used ModelNet40 dataset. The accuracy of target recognition reached 92.49% in actual nursery environments. Moreover, the misidentification and rejection rates were reduced to 13.4% and 6.47%, respectively. The accuracy of the improved PointNet was outperformed by 4.38 percentage points, compared with the original. While the misidentification and rejection rates were reduced significantly. Comparative analysis revealed that the recognition accuracy of the improved model was only 1.14 percentage points lower than that of PointNet++. Additionally, the misidentification rate and false rejection rate were slightly higher by 0.9 and 1.12 percentage points, respectively. Nevertheless, the number of parameters and computational operations were reduced remarkably in the lightweight PointNet model, only 11.5% of PointNet and 27.02% of PointNet++. The operations were reduced by 13.3 and 76.79 percentage points, respectively. The efficiency and response performance were promoted to reduce the computational complexity suitable for real-time applications. In conclusion, the improved lightweight PointNet deep learning network can offer the high real-time performance, accuracy, and robustness, which was well-suited for the target recognition in spraying operations within forest and fruit orchard environments. These findings can provided the strong reference to advance the field of target detection and recognition in spray operations.

Traditional particle swarm optimization(PSO) is prone to getting stuck in local optima and low search ability for energy consumption optimal solutions when planning plant protection operation paths. The study focuses on improving the PSO algorithm to improve the efficiency of crop protection drone operations and reduce drone losses. An adjacent competition (AC) algorithm for crop protection operation path planning based on a crop protection UAV model was proposed to solve the problem of poor effectiveness of traditional particle swarm optimization algorithms applied to UAV plant protection operations. The AC algorithm ensures that the total distance of crop protection UAV operations is constant, provides prior knowledge of the search direction, and ensures that special points are not missed, preventing the algorithm from getting stuck in local optimal solutions: Firstly, all particles initialize the homework distance range to prevent extreme situations where the single homework distance is too long or too short. Secondly, each drone is randomly assigned the operating distance for each particle within the operating range. Finally, adjacent particles compete with each other for the operating distance, indirectly changing the operating distance of each flight and searching for the optimal path. Matlab was used to conduct algorithm simulation verification on a simulated plant protection site of 420 m×200 m. Traditional PSO algorithm often gets stuck in local optima and the final planning results fluctuate greatly, while the AC algorithm obtains the optimal energy consumption solution 9 times out of 10 planning iterations, and the planning energy consumption of non-optimal solutions is similar to the optimal energy consumption, still better than all planning results of the traditional PSO algorithm, proving that the neighboring contention algorithm has stronger searchability. The results show that traditional PSO algorithms often got stuck in local optima and increased energy consumption by 16.16% to 38.14% compared to neighboring contention algorithms in 10 rounds of path planning. This indicated that the energy consumption of the AC algorithm proposed in this study for planning results is much lower than that of the traditional PSO algorithm, and the algorithm has stronger search capabilities. This study used the RflySim simulation platform to build a crop protection drone model and a 420 m×200 m work site in order to compare the simulated tracking results of the traditional PSO algorithm and AC algorithm in a virtual environment. The results show that the energy consumption of AC algorithm planning results is reduced by 25.15% compared to the traditional PSO algorithm. This study conducted actual flight tests using the drone M300 at a 420 m×200 m work site to validate the practicality of the algorithm. Actual flight experiments have shown that the energy consumption of AC algorithm planning results is reduced by 34.48% compared to traditional particle swarm optimization algorithms, making it more suitable for multiple crop protection operations. In actual experiments, the return points of multiple UAVs in the adjacent contention algorithm are closer to the supply point, the distance of departure and return is shorter, and the ineffective energy consumption is lower compared with the traditional PSO algorithm, indicating that its planning is more reasonable. Adjacent particles indirectly change the distance of each operation by competing for each other's operation distance, effectively improving the convergence speed of the algorithm, preventing the algorithm from getting stuck in local optima, and improving the efficiency of multiple crop protection operations with a single drone.

Nursery planting is ever expanding at present, particularly with the increasing demand for fruit and ornamental trees. Simple tools cannot fully meet the efficient work of large-scale production, due to the labor intensity and pesticide demand. Alternatively, spray robots can be expected to reduce the large number of tasks in modern agriculture. However, it is still lacking in the recognition accuracy and target identification of spray robots in nurseries. In this study, a nursery target detection was proposed using an improved YOLOv5s. Firstly, partial convolution (PConv) was introduced into the comprehensive convolution block (C3) to reduce the computational complexity, in order to improve the backbone network. A coordinate attention mechanism was added at the highest dimensional feature to enhance location awareness. Secondly, the neck structure of the network model was optimized to enhance the feature extraction of the model. At the same time, the bilinear interpolation was used for up-sampling operations during training. Finally, the original coupling detection head was replaced with the improved light decouple one, in order to further improve the detection accuracy. A nursery dataset was constructed with a total of 2 000 pictures, including three representative objects: trees, pedestrians, and cultivation pots. The images of the objects were collected from pedestrians with various postures at varying distances (far, medium, and near). The dataset included a total of 5 769 trees, 1688 pedestrians, and 6 178 pots. Only the trees and pots were marked in the first row during labeling, according to the operation requirements. Once the pedestrians are detected, they must be identified to consider the safety of pedestrians. Therefore, all pedestrians were labeled regardless of distance. The experimental results show that the detection of the model was improved differently via the C3 module and coordinate attention mechanism, as well as the optimized neck network structure and up-sampling mode using a lightweight decoupling head. Finally, the average mAP_0.5, mAP_0.5:0.95, accuracy, and recall reached 88.2%, 54.7%, 86.0%, and 82.4%, respectively. The size of the improved network model was 14.1 MB, the average detection speed of a single image was 19.5 ms, and the average frame rate was 51.3 frames. The mAP_0.5, mAP_0.5:0.95, accuracy, and recall increased by 4.6, 5.9, 1.8 and 3.4 percentage points, respectively, compared with the original YOLOv5s. The improved network had the highest accuracy, mAP_0.5 and mAP_0.5:0.95, compared with the current mainstream single-stage target detection YOLOv3-tiny, YOLOv7-tiny, and the latest YOLOv8s model. The experimental results show that the improved model performed better in accurately and rapidly identifying the pedestrians, pots, and trees in a complex environment. The research findings can also provide technical support to the operational activities of electric spraying robots in nurseries.

Agricultural and rural modernization has been ever accelerating with the promotion of rural revitalization strategy. The transformation of agriculture to intelligence, precision, and environmental protection has been the main trend of development in recent years. Among them, autonomous navigation of unmanned agricultural machinery has drawn much attention to effectively improve accuracy and efficiency. Path tracking control is one of the key technologies in the autonomous navigation of agricultural machinery. A lot of research has been done under relatively ideal working conditions. However, agricultural machinery has to operate in complex and various soil conditions, as well as highly unstructured farmland environments, such as muddy paddy fields, rugged slopes, and slippery grasslands. Agricultural machinery can inevitably encounter various uncertain disturbances during operation in these scenarios, leading to the low accuracy of path tracking and the stability of the control system. Therefore, anti-disturbance and strong self-adaptability are required in the autonomous navigation control of agricultural machinery. This study aims to improve the path following control accuracy and anti-disturbance ability of four-wheel self-steering (4WSS) electrically driven unmanned agricultural machinery in complex farmland environments. A control system of non-singular fast adaptive terminal sliding mode was also proposed using a nonlinear disturbance observer. Firstly, a 4WSS chassis extended kinematic tracking error model of unmanned agricultural machinery was established to consider the influence of drive wheel slip on the chassis motion in the farmland environment, including unknown sideslip speed and steering slip angle. Secondly, a nonlinear disturbance observer was designed to estimate the external lumped disturbance with unknown slip parameters, according to the chassis extended kinematic tracking error model. The disturbance compensation was introduced into the controller. A nonlinear adaptive sliding mode was also constructed to slow down the jitter problem of the nonsingular fast terminal sliding mode, according to the exponential approximation. The controller outputs confirmed the acceptable smooth control amount of the steering actuator. The stability of the composite controller was verified at the same time. Once the 4WSS unmanned agricultural machine tracked the U-shaped target path at 0.4 and 0.6 m/s, the standard deviation in the lateral error of the target operation path was reduced to 8.1 and 9.2 cm at different speeds, respectively, compared with the traditional. Furthermore, the standard deviation of lateral error was reduced to 5.3 and 7.8 cm, respectively, in spray mode. The robustness and accuracy of the path tracking control of 4WSS unmanned agricultural aircraft were effectively improved in the complex farmland environment. Better anti-disturbance and strong adaptive ability were achieved to meet the needs of agricultural production.

A safer and more efficient inspection of water quality is highly required in an aquatic-air amphibious robot. However, the traditional inspection mode of water quality cannot fully arrive at multiple target points over a pond. It is very necessary to plan a reasonable path for the robot, in order to reduce the dwelling time with cost saving. Particularly, the traditional path zigzagging of A* algorithm cannot consider the different constraints of multi-amphibious robots. In this study, an improved A* algorithm of path planning was proposed for the high search efficiency in the water-air amphibious robots. Firstly, the distribution and height information of the obstacles was collected using the improved A* algorithm. A 2.5-dimensional raster map was then established to contain the multiple water areas; Secondly, the evaluation function of A* algorithm was used to add the different energy consumption, time, and safety costs of amphibious robots. The different weights were then adjusted to obtain the initial paths; The weights among energy consumption, time, and safety costs were also dynamically allocated to improve the path planning of A* algorithm. The heuristic function of A* algorithm was then improved to speed up the search efficiency. The target cost function was used to prioritize all detection targets of water quality, in order to realize the path planning of multiple target points in a partitioned fishpond; Finally, the initial paths were optimized to add the additional airborne mode-switching, thus deleting the redundant points. The B-spline algorithm was adopted to generate a three-dimensional smooth trajectory. As such, an optimal path was obtained to connect the multiple inspection points of water quality. Simulation results show that the number of iterations of the improved A* algorithm was reduced by 70.04% and 68.07%, respectively, compared with the traditional and land-air A* algorithms; while the length of the path was reduced by 35.44% and 7.6%, respectively; the total angle of turn was reduced by 83.63% and 8.65%, respectively; and the number of dangerous nodes was reduced by 80.67% and 33.33%, respectively. The real water test showed that the number of iterations of the improved A* algorithm was reduced by 84.89% and 83.78%, respectively; the length of the path was reduced by 12% and 0.6%, respectively; the total angle of turn was reduced by 73.21% and 22.1%, respectively; and the number of dangerous nodes was reduced by 84.62% and 80%, respectively. The improved A* algorithm was achieved in the safe and smooth path, according to the multiple target points in multiple pieces of water. The efficiency of water quality detection was effectively improved in the partitioned fishponds, compared with the traditional approach. The finding can also provide a strong reference for the autonomous navigation of multi-habitat robots.