Existing agricultural monitoring platforms in open-field environments generally suffer from limited endurance of unmanned aerial vehicles, insufficient viewpoint consistency during low-altitude repeated observations, and high susceptibility of ground robots to surface conditions and crop canopy occlusion, making it difficult to concurrently meet the demands of rapid deployment and highly repeatable observation. To achieve continuous acquisition and comparative analysis of agricultural information at the same location and from the same viewpoint over different periods, a wheel-wing hybrid cable-travelling inspection robot for agricultural monitoring was designed. Using overhead cables above farmland as fixed inspection paths, a dual-mode inspection scheme combining aerial cable docking and cable-travelling locomotion was proposed to meet the demand for repeated observation at fixed locations and viewpoints. On this basis, an integrated configuration composed of a passive trigger gripper, deformable arms, and wheel-wing hybrid end-effectors was developed to realize structural reuse of the aerial execution unit and the cable-travelling execution unit. In addition, key constraint models, including flight hovering, anti-slip on the cable, cable-travelling driving torque, and arm driving with over-center self-locking, were established, and SolidWorks and ADAMS were used to simulate mode switching and cable-travelling operation. The simulation results showed that the two arms of the robot could fold synchronously and the mode-switching process was continuous and stable. The robot could complete the transition from the flight-deployed mode to the cable-travelling working mode within approximately 5 seconds. In the cable-travelling mode, the platform moved continuously with stable wheel-cable contact, and the pitch angle variation of the front and rear wheels ranged from about −0.2 to 0.2 rad, indicating favorable path constraint and observation posture stability. The proposed robot can satisfy both rapid access to field monitoring areas and stable fixed-path inspection requirements, providing a reference for temporal agricultural monitoring, repeated observation data acquisition, and the development of low-cost agricultural inspection equipment.

Oil-powered plant protection unmanned helicopters play an indispensable role in modern precision agriculture due to their large payload and long endurance. However, their complex operational mechanisms, including broadband strong vibration, high-temperature engine radiation, and coupling among multiple subsystems, impose extremely high demands on condition monitoring. Existing monitoring methods mostly rely on closed flight control systems, which commonly suffer from issues such as missing on-board equipment status data, low telemetry data update rates, high transmission latency, and non-open interfaces, leading to severe deficiencies in real-time flight status monitoring and fault diagnosis capabilities. To address these problems, a multi-signal acquisition and condition monitoring system is designed that is completely independent of the original flight control system. At the hardware circuit level, by integrating an MPU6050 attitude sensor, laser rangefinder, thermocouples, and optoelectronic isolation detection circuits, the system achieves high-frequency acquisition of nine types of key parameters, including flight attitude angles, engine speed, exhaust gas temperature, generator three-phase voltage, and servo command/feedback signals (attitude at 5 Hz, status at 1 Hz). The extended MavLink protocol is used for data encapsulation, and data are transmitted to a ground station via 4G/5G networks in a time-sharing manner, combined with on-board cyclic data storage, thereby establishing a dual-redundant data link. On this basis, an engine overheating early-warning method based on a temperature-speed coupling model and a fault detection algorithm for servo actuators based on an Extended Kalman Filter (EKF) are proposed. Test results show that the system operates independently and stably, reducing the uplink latency of critical attitude data to less than 200 ms, and successfully achieving rapid identification of faults such as generator phase loss (response <1 s) and servo jamming. This system can serve as an independent on-board “black box”, providing a reliable data foundation and diagnostic tool for real-time condition monitoring, predictive maintenance, and post-flight incident analysis of oil-powered unmanned helicopters.

In the transition from traditional agriculture to smart agriculture, the demand for large-scale, high-precision information perception in agricultural scenarios is increasing. Traditional monocular and binocular vision, limited by their field of view, struggle to meet the requirements of all-round monitoring in complex environments. Panoramic vision technology, characterized by 360° unobstructed field of view, fast and complete information acquisition, and strong real-time performance, has become an irreplaceable key technical means in the intelligent upgrading of agriculture. This paper aims to systematically review the application status and development progress of panoramic vision systems in the agricultural field, providing theoretical reference for the in-depth integration of panoramic vision technology and agricultural equipment. Based on imaging principles and hardware structures, the systems are classified into four architectural types: pan-tilt rotating type, multi-camera array type, fisheye lens type, and catadioptric type. Combined with typical cases, the application evolution in fields such as crop monitoring, phenotypic investigation, and agricultural machinery navigation is analyzed. The research shows that panoramic vision has formed a multi-level application pattern: pan-tilt rotating systems have small distortion but poor real-time performance, suitable for static growth monitoring; multi-camera systems have high frame rates and small pose errors, capable of supporting dynamic high-dimensional phenotypic reconstruction and positioning; fisheye systems have a simple structure, and edge distortion can be suppressed through calibration to achieve wide-area monitoring; catadioptric systems are compact with no delay and small deviations in navigation and obstacle avoidance. Currently, panoramic vision in agricultural applications still faces four major issues: bottlenecks in adaptability to complex environments, shortcomings in real-time processing of dynamic scenes, insufficient efficiency and accuracy of 3D information reconstruction, and lack of technical standardization and system compatibility. Therefore, in the future, panoramic vision technology should focus on multi-modal sensor fusion perception, use deep learning-driven adaptive model optimization to improve generalization ability, and deeply collaborate with the agricultural Internet of Things to build an integrated perception-decision-regulation system. Meanwhile, strengthening technology open-sourcing and industry standardization will accelerate the large-scale application of panoramic vision technology in precision agriculture.

To address the limitations of current agricultural image recognition tasks, such as isolated models, weak multi-task synergy, insufficient real-time performance and generalization capability, this study develops an intelligent agricultural image recognition system based on deep learning and multi-task fusion. This system adopts a "device-edge-cloud" three-level collaborative architecture. By constructing a multi-task joint learning framework, the system enables concurrent training and collaborative inference for both crop height measurement and disease recognition. The system fuses the feature extraction capabilities of mobile convolutional neural networks (MobileNet) and visual transformers (ViT) with a coordinate attention (CA) mechanism; this design enhances key feature extraction, suppresses complex farmland background interference while significantly alleviating computational burden. Experimental results show that the system achieves remarkable performance across both core tasks. For crop height measurement, the determination coefficient R2 between the algorithm′s measurement results and the manual measured values reaches up to 99.98%. For disease recognition, the optimized lightweight model achieves an average detection accuracy of 98.76%, with only 12.64 M parameters. The average single-image inference latency is as low as 9.1 ms, demonstrating excellent potential for embedded deployment. Moreover, the system exhibits high stability in complex environments, maintaining measurement deviation of crop height within ±2 cm. This precision allows for seamless integration into low-altitude monitoring platforms, such as unmanned aerial vehicles (UAVs), to satisfy the real-time data processing requirements of low-altitude agricultural operations. In summary, the proposed system delivers excellent scalability, real-time efficiency, and robustness, effectively overcoming traditional bottlenecks in multi-task synergy, lightweight deployment, and adaptation to complex environments. It provides an integrated solution for multi-dimensional image recognition in smart agriculture empowered by the low-altitude economy, thereby facilitating the large-scale implementation of precision agriculture while improving agricultural production efficiency.

Accurately obtaining annual cropland area and spatiotemporal change information is crucial for safeguarding the agricultural land redline and improving the efficient, economical, and intensive utilization of water resources. It also provides a scientific basis for navigation and intelligent scheduling of agricultural machinery, thereby effectively enhancing the precision and efficiency of mechanized agricultural operations. Taking the Aksu river basin in the Xinjiang Uygur Autonomous Region as an example, this study proposes a method for extracting cropland spatiotemporal change information based on time-series index features. First, based on NDVI data derived from time-series remote sensing images covering the entire crop growth phenological period, annual composite vegetation index images are generated using the maximum value compositing method, and these images are stacked into bands to construct a time-series index feature image. Second, an object-oriented segmentation approach is applied to perform multi-scale segmentation on the feature image to obtain segmentation results. Finally, the change patterns of cropland across different years revealed by the feature image are analyzed, and extraction rules for different cropland change types are established. A decision tree classification method is then employed to achieve rapid extraction of annual cropland spatiotemporal change information. Experimental results demonstrate that the proposed method can extract annual cropland change information rapidly and accurately. Analysis of the annual change information shows that the cropland area in the Aksu river basin exhibited an overall increasing trend from 2021 to 2024, with relatively high growth rates of 1.24% and 1.28% in 2022 and 2023, respectively, while the growth rate slowed to 0.42% in 2024. To validate the accuracy of the extraction results, a comparison between the statistical yearbook data of selected counties and the remote sensing monitoring data was conducted, revealing a relative error of approximately 5%. This confirms that the proposed method attains satisfactory accuracy in extracting cropland area.

Apples, as a high-yield fruit, often suffer significant losses during transportation and storage due to collisions and improper temperature control. Traditional manual inspection methods for apple surface damage are not only costly and inefficient but also heavily reliant on operator experience and subjective factors, leading to limited detection accuracy. To address these issues, this study proposed a cost-effective and efficient apple damage detection method based on laser speckle imaging. The proposed approach integrated the RS_FUJII speckle feature extraction algorithm with the RECONANOM-AE damage detection model to improve detection speed and accuracy. Shandong Fuji apples were used as the experimental samples. Speckle images were first preprocessed, followed by feature extraction. To overcome the limitations of the traditional FUJII algorithm in spatiotemporal feature representation and noise robustness, this study introduces a spatiotemporal weighting mechanism and robustness enhancement strategies to develop an improved RS_FUJII algorithm for high-precision damage feature extraction. In addition, an unsupervised neural network model, RECONANOM-AE, is designed to automatically identify damage regions via reconstruction error. Experimental results demonstrate that the proposed model exhibits favorable convergence and stability during training, with both training and validation errors gradually decreasing with increasing epochs. Comprehensive performance evaluation showed that the model achieved a true positive rate of 86.67% and a false positive rate of 6.67%. The overall recognition Accuracy reached 90%, with a Precision of 92.86%, Recall of 86.67%, Specificity of 93.33%, and an F1 Score of 89.68%. These results demonstrated that the proposed method achieved good training stability and strong generalization ability. The method enabled non-destructive quality assessment of apples and provided a reliable technical solution for intelligent fruit quality inspection.

Accurate localization of keypoints for animal body measurement is of great importance for intelligent livestock health monitoring and automated measurement systems. However, existing keypoint detection models often suffer from two major limitations: a large number of parameters which are unfavorable for deployment on edge devices, and insufficient localization accuracy in some lightweight models that fails to meet the precision requirements of body measurement tasks. To address these challenges, this paper proposes a lightweight keypoint detection model based on an improved YOLOv8-Pose architecture, which effectively enhances detection accuracy while maintaining real-time performance. In the proposed model, coordinate attention mechanisms are integrated into the shallow layers of the backbone network and before the three heads, enabling spatial coordinate information to be effectively embedded into channel attention. This design enhances the model ability to perceive the spatial distribution of keypoints, thereby improving the accuracy and stability of keypoint localization. In addition, a forced sharpening module is introduced in the shallow backbone to strengthen edge and contour features, facilitating a better balance between global semantic representation and local detail preservation. Furthermore, a high-resolution feature branch is incorporated to provide fine-grained spatial information to the three detection heads, compensating for the loss of spatial resolution in deeper feature maps. Experimental results demonstrate that the proposed model achieves a precision of 82.6%, a recall of 81.3%, and an mAP50-95 of 58.1% on complex horse keypoint detection tasks. Compared with YOLOv8-Pose, the proposed method achieves improvements of 21.0%, 39.2%, and 45.3% in precision, recall, and mAP50-95, respectively, exhibiting superior robustness and localization capability under complex poses and diverse environmental conditions. In terms of model complexity, the number of parameters increases from 3.361 M to 4.611 M, corresponding to an increase of approximately 37.2%, while the floating-point operations rises from 9.6×10⁹ to 10.4×10⁹, an increase of only 8.3%. Despite this limited computational overhead, the proposed model delivers substantial performance gains without significantly affecting inference speed. These results indicate that the proposed approach achieves a favorable trade-off between accuracy and efficiency by incurring a relatively small model cost for considerable performance improvement, demonstrating its strong potential for intelligent livestock farming and automated body measurement applications.

To address difficulties and inefficiencies in monitoring duck weight during the routine inspection of cage-raised meat ducks, this study proposes a precise body weight prediction method utilizing a tree-based machine learning algorithm, aiming to enhance the stability and accuracy of weight prediction in complex scenarios of large-scale meat duck production. Unilateral body size and age data were manually collected from meat ducks at Shuyang Xinfeng Future Farm to develop an extreme gradient boosting model predicting weight. The model integrated morphometric features and chronological age as predictors to evaluate their respective impacts on body weight estimation. The results indicate a distinct growth trajectory: for ducks 1-13 days, data collection is hindered by their small body size, and weight growth is relatively slow, making this period less critical for weight prediction modeling. Conversely, ducks aged 14 to 35 days undergo a rapid growth phase, representing a critical period in the production cycle during which age and unilateral body size serve as pivotal predictors. The Extreme Gradient Boosting Tree model effectively predicted the weight of meat ducks, achieving an optimal R² of 0.95, a minimum mean squared error (MSE) of 0.563 kg², and a minimum mean absolute error (MAE) of 0.044 kg. Furthermore, by evaluating the importance of unilateral body size data of meat ducks, it is concluded that body slant length, as the main parameter of meat duck development, contributes the most to the model among body size data, accounting for 20.7% (second only to age) while tibia length exhibited the weakest correlation, contributing only 4.8%. This study not only deepens the understanding of machine learning application in meat duck weight prediction but also provides a valuable method for selection and management in non-contact meat duck weight analysis scenarios. Additionally, it offers technical support for improving model accuracy and promoting the development of intelligent cage-raised meat duck farming, and provides an effective automated solution for weight monitoring via inspection robots.

Aiming at the three major bottlenecks faced by intelligent agricultural machinery, namely on-board software fragmentation, insufficient hardware-software collaboration, and lack of underlying platforms, this study proposes the agricultural machinery intelligent engine system (AMES). Its design follows three core principles: integrity, reusability, and data-driven. It adopts a layered architecture: integrating a real-time operating system (RTOS), general-purpose on-board software, and service-oriented architecture (SOA) communication middleware to ensure system compatibility and real-time performance, encapsulating the “perception-decision-control” closed-loop capability, providing a multi-domain control model library, a dedicated algorithm library, and a development toolkit to support three core functions: intelligent navigation, intelligent control, and intelligent management. In practical applications, AMES has achieved a pest and disease identification accuracy rate of ≥95% in field inspection robots, improved the overall operational quality by more than 20% in intelligent grain combine harvesters, and achieved a harvesting efficiency of 2 500 stalks per hour with a damage rate of ≤5% in asparagus harvesting robots. These results effectively promote the transformation of the agricultural machinery industry towards “integrated hardware-software services” and help China's intelligent agricultural machinery technology transition from “catching up” to “leading”.

To address the challenges in balancing accuracy and real-time performance for navigation path recognition methods of wheat harvesters, this study introduces a lightweight design based on the DeepLabv3+ model. Specifically, the original backbone network is replaced with MobileNetV3-Large, and the ReLU activation function is substituted with Leaky_ReLU. To further reduce computational load, the three dilated convolutions with different dilation rates within the atrous spatial pyramid pooling (ASPP) module are replaced with depthwise separable convolution (DSC) employing identical dilation rates. To ensure the generalizability of the wheat harvester navigation path recognition research, images of wheat harvest areas were captured under six typical environmental conditions: strong light, low light, back lighting, front lighting, shadows, and field edges. The road information within the collected images was annotated by the Labelme tool, thereby constructing a wheat harvest region data set. In the path extraction stage, key points were acquired from the segmentation mask maps using the horizontal scanning method, and the navigation path was fitted utilizing a piece wise cubic B-spline algorithm. Experimental results demonstrate that the improved DeepLabv3+ model achieved a segmentation accuracy of 98.04% and intersection over union (IoU) ratio of 95.20%, respectively, with a video image processing frame rate of 7.5 frames per second. The average pixel error and average distance error for navigation path recognition were 7.4 pixels and 37 mm, respectively. The wheat harvester operated at a travel speed of approximately 1.5 m/s, and the path recognition time per single frame was only 0.15 seconds. This performance effectively meets the real-time and accuracy requirements for wheat harvester operation. This research provides a theoretical foundation and technical support for enhancing the autonomous navigation capabilities of wheat harvesters.