Returning straw to fields is a fundamental agricultural practice that enhances soil structure, mitigates erosion, improves fertility and productivity, and supports subsequent crop cultivation. To address environmental imbalances, inconsistent growth, and pest, disease, and weed proliferation caused by the uneven distribution of existing straw return machinery, this study introduces a uniform straw cover detection and control system utilizing machine vision and deep learning segmentation. An optimized YOLOv11s-seg network was developed to extract geometric and positional data from scattered straw. Using an upper computer, the system calculates the straw center deviation, coverage ratio, and preset differences. Simultaneously, a uniform coverage control system was designed around a single-chip microcomputer and a hydraulic system. Utilizing visual inspection data, the system regulates the straw scattering angle and width in real time. The optimized segmentation model demonstrates exceptional performance, achieving an average accuracy of 96.7% for chaff detection. Field tests indicated that, across various operational speeds, the qualified scattering width rate exceeds 80% following system optimization. Furthermore, the average scattering unevenness surpassed the national standard requirement of 20%, confirming highly effective closed-loop control. The enhanced YOLOv11s-seg model satisfies the rigorous accuracy and efficiency demands for real-time field detection. This vision-based control system accurately regulates straw spreading direction and width, significantly improving uniformity and operational efficiency. These findings provide essential technical support for the automation and intelligent development of straw shredding and returning machinery.

Accurate identification and quantitative assessment of fish feeding intensity are pivotal for enhancing aquaculture production efficiency. Currently, feeding intensity is mainly assessed based on fish school feeding images with a single feature, overlooking the interdependencies between individual fish and the fish school’s behavior. Therefore, this paper presents a method based on detecting individual fish heads to characterize the feeding aggregation speed and the average swimming speed of the fish school, thereby quantifying the fish school’s feeding intensity. First, the improved YOLOv11n-ALL model was employed to detect individual fish heads, resulting in improved detection performance, increasing inference speed, and reducing computational complexity. Additionally, feeding aggregation speed and average swimming speed indices for fish schools were constructed by combining the YOLOv11n-ALL model with the ByteTrack algorithm to track and extract the centers of individual fish heads’ detection boxes. Finally, the fish school feeding kinetic energy was assessed using the feeding aggregation speed and average swimming speed dual indices, and the fish school feeding intensity levels were classified according to the feeding kinetic energy. Experimental results reveal that the improved YOLOv11n-ALL model achieved an average detection precision (mAP50) of 94.13% for detecting fish heads, reduced the parameter count by 22.09%, and exhibited a computational complexity of 6.4 GFLOPs. Furthermore, the classification model of fish school feeding intensity, quantified by the dual indices of average swimming speed and feeding aggregation speed, achieved a detection accuracy of 97.41%. This method digitizes detection results, enabling rapid classification of fish school feeding intensity and demonstrating its effectiveness for feeding intensity assessment and the development of scientific feeding strategies.

Tomato lateral branches are frequently occluded by leaves and fruits, leading to difficulties in feature extraction. Moreover, due to the differences in occlusion positions, a unified method for pruning point localization and pose estimation cannot be well adapted to complex and variable occlusion scenarios, often causing localization deviations in the operation of pruning robots. To address these challenges, this study proposes a lightweight keypoint detection model (YOLO-LB) for tomato lateral branches, along with differentiated pruning point localization and pose estimation methods for different occlusion scenarios. Specifically, YOLO-LB integrates the mobile multi-query attention and spatial and channel reconstruction convolution, significantly improving the feature extraction capability and keypoint localization accuracy for occluded lateral branches. By determining the occlusion status of each keypoint via confidence threshold, a classification system for lateral branch occlusion scenarios is established. Differentiated processing strategies are designed for each occlusion scenario, enabling precise calculation of pruning point coordinates and pose using 3D point cloud data. Experimental results demonstrate that YOLO-LB achieves 98.39% accuracy, 96.06% mean average precision, and 94.47% keypoint average precision, with only 3 million parameters and 8.9 giga floating-point operations per second, outperforming comparison models in comprehensive performance. After deploying this method on an embedded computing platform for field experiments, under various occlusion scenarios, the average Euclidean distance error for 3D pruning points is 8.01 +/- 3.04 mm; the mean absolute errors for pitch and yaw angles are 11.31 +/- 6.18 degrees and 10.97 +/- 6.29 degrees, respectively. The proposed method provides reliable visual support for tomato pruning robots.

In orchard inter-rows, weeds often grow unevenly in patches owing to terrain cluttered with stones, branches, and debris. Empirical operating parameters can cause cutting and drive motors of mowing robots to stall under sudden loads, reducing efficiency and quality and risking damage. This study introduces a load-adaptive control strategy that links blade rotation and chassis forward speed. A fuzzy controller regulates forward speed based on blade motor deviation and its rate of change, while a sliding-mode controller incorporating a torque-feedforward state observer controls the cutting motor. The chassis employs a Smith-radial basis function (RBF)-proportionalintegral-derivative (PID) controller to suppress phase lag, with parameters optimized by an improved chimp optimization algorithm (ICOA) employing tent reverse learning, distance-weighted updating, and Gaussian mutation. Simulations indicate that ICOA converges in 15 iterations with a target function value of 0.2413, reducing error by 13.57 % over the baseline. The ICOA-Smith-RBF-PID controller achieves a rise time of 0.0335 s (step), with maximum speed errors of 0.0039 m center dot s- 1 (ramp) and 0.0038 m center dot s- 1 (sinusoidal). The blade speed controller responds in 0.0163 s (step) and 0.0246 s (load step), with a maximum speed error of 15.33 r center dot min- 1 under continuous disturbance. Field tests across five rows of a single orchard yielded mean mowing, grassbreaking, and cutting stability rates of 92.67 %, 89.41 %, and 88.98 %, respectively, with no motor stalling. The proposed strategy ensures stable, efficient, and precise mowing under dynamic orchard conditions; demonstrates robust applicability in hilly orchards; and provides a reference for coordinated control in agricultural robotics.

To mitigate tillage-depth fluctuations induced by surface undulations and actuator nonlinearities in horticultural electric tractors, this study develops a predictive depth control method. A novel MSCNN-BiLSTM-Attention model is constructed to forecast pitch-angle variations. It utilizes multi-scale convolution to extract disturbance features across different temporal scales, while bidirectional temporal modeling combined with a multi-head attention mechanism enhances temporal representations. Model hyperparameters are optimized using an improved beluga whale optimization algorithm. The predicted pitch angle is integrated as feedforward compensation into the control loop. By fusing this prediction with a rotary-tillage kinematic model, future tillage depth is estimated, enabling the design of an RBF-based active disturbance rejection controller (RBF-ADRC) for adaptive compensation of system nonlinearities and time-varying disturbances. The proposed prediction model achieved MAE and RMSE of 0.0526 degrees and 0.0686 degrees, respectively, with a test-set R2 of 0.9931, and errors primarily within +/- 0.03 degrees. Simulations demonstrated that the proposed prediction-assisted control strategy outperforms RBFPID, optimal sliding mode, and model predictive controllers under step, sinusoidal, and slope disturbances, achieving zero overshoot, a steady-state error of 0.001 cm, and reduced dynamic errors. Greenhouse field experiments further validated the method, yielding average tillage depths of 4.03, 8.14, and 11.59 cm against targets of 4, 8, and 12 cm, with maximum fluctuations of approximately 0.6 cm. Overall, this deep learning-based predictive feedforward strategy significantly improves tillage-depth stability and control precision under complex surface conditions, providing a practical solution for intelligent electric tractors in protected horticulture.

Accurate weed models are critical for elucidating mechanical weeding interactions. However, key parameters and targeted models for flexible weed-body fracture are lacking. We propose a novel, discrete-element model (DEM) for simulating flexible weed stress fracture using the Hertz–Mindlin with Bonding V2 contact model. The proposed model facilitates high-fidelity root-architecture reconstruction and accurately reproduces root mechanical responses, including fracture, fragmentation, tension, and shear. A full-chain multi-index constrained parameter calibration system was established via physico-mechanical testing of barnyard grass (Echinochloa crus-galli) roots, with a simulated-measured contact-angle deviation of just 2.23%. The key calibrated parameters included a normal stiffness and shear stiffness of 1.98243 × 1014 N·m−3, normal strength and shear strength of 2.03399 × 1010 MPa, bonding V2 radius of 5.771 × 10−2 mm, and inter-particle contact radius of 6.627 × 10−2 mm, ensuring model accuracy and biological authenticity. Moreover, we developed an adaptive evolutionary genetic hippopotamus optimisation algorithm (AEG-HOA) for the global optimization of DEM parameters, with the ultimate tensile and shear strengths as response metrics. Comparative results revealed that AEG-HOA achieved 1.11% and 1.08% calibration errors, far below those of conventional methods (4.23% and 3.23%, respectively), attaining convergence within nine iterations with robust stability. This work provides a high-precision standardised DEM for mechanical weeding interaction research and efficient plant-simulation-parameter calibration.

Dust accumulation on greenhouse roofs remains a significant challenge, significantly reducing light transmission and crop productivity, and creating a need for efficient and automated cleaning solutions. Current manual or semi-automated methods are labor-intensive, inconsistent, and often unsuitable for large greenhouse structures. This study presents the design and simulation of a solar-powered autonomous cleaning system for plastic greenhouse roofs. The proposed system integrates a photovoltaic (PV) energy supply with an automated cleaning mechanism composed of rotating brushes, flexible wipers, and a water pump to effectively remove dust and contaminants from the greenhouse surface. A lightweight mobility unit is incorporated to enable stable movement and operation on inclined greenhouse roofs. Structural and performance analyses were conducted using three-dimensional modeling and finite element simulations to evaluate stress and deformation in the frame, brush, and wiper. The maximum stress occurred at the upper joints, while total deformation remained below 2.1 mm, confirming adequate structural safety under operational loads. The cleaning efficiency was tested at two-wheel speeds (25 rpm and 50 rpm) and three roof positions (top, middle, bottom), demonstrating stable operation and uniform cleaning performance. The proposed system offers a practical solution for maintaining greenhouse transparency, reducing manual labor, and supporting sustainable agricultural management.

Abstract Traditional greenhouse cleaning methods are labor-intensive, prone to human error, and inefficient, often compromising light transmittance and productivity. To address these challenges, this study proposes an autonomous robot designed to clean greenhouse roofs efficiently and reliably. The robot features an integrated cleaning system with adjustable brushes, wipers, and water sprinklers, ensuring optimal performance and significantly improving light transmittance. Powered by a 500 W PV system, it utilizes electric wheels for smooth, stable movement and incorporates a replaceable brush-wiper mechanism for enhancing durability and maintenance efficiency. The design process involved SolidWorks modeling for mass properties, CFD simulations with the k-ε turbulence model to evaluate wind load conditions, and ANSYS structural analysis to confirm durability under extreme wind speeds of up to 126 km/h (ten times greater than normal conditions). Structural tested at different robot’s rotational speeds 25 rpm and 50 rpm confirmed optimal performance at 25 rpm, balancing cleaning efficiency and long-term durability. Additionally, the robot incorporates advanced control unit with sensors for autonomous operation, real-time light transmission monitoring, and navigation capabilities, distinguishing it from traditional manual or semi-automated methods. The results demonstrated robust performance in extreme conditions, surpassing existing systems limited to standard weather. The robot’s performance is limited by speed (0.35 m/s), battery life, roof complexity, maintenance, adaptability, and cost, indicating areas for improvement. Future developments will integrate AI for autonomous decision-making, GPS for precise navigation, and a smart cleaning system to optimize performance based on real-time data, further reducing maintenance costs and ensuring optimal greenhouse lighting.

Hypervolatile species such as carbon monoxide (CO) and molecular nitrogen (N-2) have been detected in comets, and could be used to constrain comet formation temperature conditions if their presence is due to freeze-out and/or entrapment. Here, we instead explore another plausible origin of cometary hypervolatiles: photodissociation of less volatile species. We characterize CO and N-2 formation following ultraviolet (UV) irradiation and electron bombardment of carbon dioxide (CO2), ammonia (NH3), H2O:CO2, H2O:NH3, and H2O:CO2:NH3 cometary ice analogs. We find that CO and N-2 form in all photoprocessed ices at temperatures between 10 and 100 K, resulting in 0.4%-0.9% CO and 0.03%-0.7% N-2 relative to water, and CO/CO2 and N-2/NH3 mixing ratios of 2.5%-62% and 0.7%-9%, respectively, across the experiments. Because our initial ices are reasonably well matched to interstellar ices and we use a UV exposure similar to a dark cloud, we can compare the resulting ratios directly to cometary abundances. Such a comparison shows that, while only a few of CO observations in comets are readily explained by photodissociation, almost all observed cometary N-2 can be accounted for by photodissociation of NH3 embedded in water ice. The latter result is also consistent with observed similarly elevated isotopic ratios of N-2 and NH3 in 67P. Taken together, our results suggest that N-2/H2O ratios <1% should be used cautiously when inferring a comet's formation location, while the more substantial CO abundances seen in many comets do likely imply entrapment at low ice temperatures.

To address the challenges of hyperspectral data redundancy, small-target segmentation difficulty, and insufficient model real-time performance in high-precision online detection of foreign matter and kernel breakage in machine-harvested soybeans, this study proposes a collaborative detection method based on "feature wavelength optimization + MobileNetV4-Unet-SGCPNet hybrid network". First, 18 key feature bands were screened from 400 - 1000 nm hyperspectral data using successive projection algorithm (SPA) and competitive adaptive reweighted sampling (CARS), constructing a multi-source feature spectral image dataset. Subsequently, a MobileNetV4-Unet-SGCPNet hybrid network was designed, with a lightweight MobileNetV4 as the encoder, combined with the symmetric encoder-decoder structure of Unet and the spatial detail-guided context propagation module (SGCP) to achieve high-precision segmentation of broken grains, complete grains, and impurities. Finally, pixel-wise voting was employed to fuse multi-band feature information, enhancing the model's generalization capability. The results demonstrate that: on the test set, the model achieves an average intersection over - union of 89.69 % for soybean component recognition, with a mean precision average of 94.55 %, a mean precision of 93.63 %, a frame rate of 4.77 FPS, a parameter count of only 2.86 MB, and a computational load of 35.78 GFLOPs. Compared to mainstream models, this method reduces parameters by 97.2 % and computational cost by 97.8 %, while the average intersection - over - union drops by only 6.2 %, with a frame rate improvement of over 5 times, striking a significant balance between detection accuracy and real-time performance. Crossvariety and cross-device validations further confirm that the model effectively adapts to morphological and spectral variations across different soybean varieties, exhibiting strong generalization ability. This study provides a core algorithmic foundation for online monitoring systems of intelligent harvester operation quality, offering critical support for enhancing the commercial value of machine-harvested soybeans and advancing the intelligence level of agricultural machinery.

Rice is one of the major staple crops in China, and its yield is closely tied to national food security and farmers’ economic returns. Lodging in rice not only reduces the efficiency of mechanical harvesting but also severely impacts yield and grain quality. Therefore, accurately identifying lodged areas is of great importance. This study proposes a rice lodging detection method based on UAV-acquired multispectral remote sensing imagery. High-resolution, multi-temporal images were collected over paddy fields in Yuhang District, Zhejiang Province, using DJI Mavic 3 M and M300 UAVs. A dataset was constructed via image cropping and data augmentation. Two deep learning models—U-Net with a VGG-16 backbone and DeepLabv3+ with a MobileNetv2 backbone—were compared for semantic segmentation performance. Experimental results show that the U-Net model achieved superior performance on the validation set, with a mean Intersection over Union (MIoU) of 91.57 %, mean Pixel Accuracy (MPA) of 95.83 %, Precision of 95.27 %, Recall of 95.83 %, and training/validation losses of 0.106 and 0.151, respectively, outperforming the DeepLabv3+ model. Additionally, the impact of different training-validation data split ratios was examined. The U-Net model showed better generalization and stability when trained with a 9:1 split compared to an 8:2 split. Furthermore, based on the semantic segmentation results, the area of lodged rice was estimated and compared against ground-truth measurements. The U-Net model produced minimal relative error, with a maximum deviation of <3 %, demonstrating strong practical applicability. These findings suggest that the U-Net model not only offers high accuracy and stability but also provides a reliable technical foundation for agricultural disaster monitoring and precision management using high-resolution UAV imagery.

To improve the driving stability and safety of electric drive mobile platforms (EDMP) for protected horticulture, it is essential to minimize the excessive slip of driving wheels. Therefore, a tire-soil model is established and a dynamic model of EDMP with implement considering the influence of wheel slip on longitudinal motion is developed. Subsequently, a linear parameter-varying model incorporating longitudinal speed and wheel force is established. A state estimator utilizing an improved adaptive strong tracking unscented Kalman filter (ASTUKF) algorithm is proposed to obtain real-time friction coefficients of four wheels and determine the optimal slip rate. Based on this, a robust model predictive controller (RMPC) with the employment of linear matrix inequality is designed to suppress EDMP slip. Finally, to validate the effectiveness of the proposed controller, a real test system for the EDMP is developed. It’s demonstrated that the slip rate of the EDMP can be significantly reduced through the implementation of the proposed skid control strategy. The average estimation errors of the ASTUKF are reduced by 95.5% and 81.6% compared to the KF and UKF, respectively. Under both straight and continuous steering conditions, the wheel slip rate errors are reduced by 48.33% and 55.63%, respectively.

Strawberries grown on elevated stands usually suffer from fruit occlusion issues, which severely limit the implementation of strawberry recognition and picking point localization, and the embedded devices carried by strawberry picking robots have high requirements for model lightweighting, posing a dual challenge to the efficient execution of automated picking tasks by robots. To address this issue, this study proposes a method for strawberry recognition and picking point localization in multi-occlusion scenes based on a lightweight keypoint detection model. Firstly, a strawberry dataset covering no, slight, moderate, and heavy occlusion scenes is constructed. Then, a lightweight strawberry recognition and keypoint detection network, LS-net, is proposed. LS-net improves the spatial relationship modelling capability between strawberries and stems by integrating the lightweight MobileNetv4 backbone with the Mobile Grouped-Query Attention mechanism; improves the feature pyramid network using depthwise separable convolutions and incorporates an anchor-free decoupled head network to reduce computational complexity while maintaining detection accuracy; and introduces the Matrix Non-Maximum Suppression to optimize the processing of overlapping strawberries, which effectively reduces the false negative detections. Based on the keypoint detection results from LS-net, the picking point coordinates and stem pose are calculated after a series of processes such as region-of-interest extraction, binarization, and depth data alignment. The experimental results show that the accuracy of LS-net is 91.07%, the mean average precision is 93.93 %, and the average pixel Euclidean distance is 4.79. By deploying LS-net to the embedded device, its frames per second reaches 78.2, and the success rates of 3D picking point localization and stem pose estimation are 84.07% and 81.32 %, respectively. LS-net and related methods provide a visual recognition solution adapted to embedded devices for strawberry picking robots. (c) 2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Precision seeding represents a key advancement in rapeseed mechanization, offering an effective strategy to minimize seed usage, reduce labor requirements, and improve efficiency. Accurate real-time seed-flow monitoring is essential for effectively maintaining sowing quality. Conventional detection techniques struggle with high-frequency seed discharge, leading to missed detections of overlapping seeds, limited resistance to interference, and diminished accuracy. This study introduces a novel rapeseed seed-flow detection method and sensor system based on microwave resonant cavity perturbation with absorptive damping. By constructing a unidirectional microwave field and incorporating absorbent materials, multipath reflections within the cavity were significantly mitigated, thereby improving detection stability and precision. CST simulations confirmed the directional damping and absorption characteristics of the materials, introducing negligible attenuation (mean 0.41 dB) along the main propagation path while providing strong suppression (mean 16.92 dB) of unwanted field components. The signal-processing circuit features two-stage intermediate-frequency amplification, RMS detection, and multithreshold comparators. When integrated with an overlapping-seed recognition algorithm and an STM32 microcontroller, the system achieved real-time counting of both single and overlapping seeds. Bench tests revealed that at a seeding frequency of approximately 31 Hz, the detection error of the device was below 2.3%, compared to over 4.1% for traditional photoelectric and fiber-optic sensors. Additional tests using multiple rapeseed varieties further verified robust single-seed detection accuracy (>97%). The recognition rates for double and triple overlapping seeds reached 76.3% and 42.0%, respectively. In field trials, detection accuracy remained above 95.5% with stable performance under operational speeds of 2.8-4.8 km/h and seeding frequencies of 12.7-30.7 Hz. The results offer an effective technical solution for precise seed-flow monitoring in rapeseed direct seeding, enhancing precision planter intelligence.

To accurately predict wear locations on the rotary tillage blade surfaces operating in wet-adhesive soil, a novel discrete element modeling method of rotary tillage blade is first proposed, and an improved contact model incorporating the wear effects between wet-adhesive soil and blades is developed in this work. Model parameters are calibrated using direct shear and penetration tests, and a discrete element interaction model of the wet-adhesive soil-blade coupled system is established. The proposed model is then employed to investigate the effects of the blade shaft orientation, shaft speed and cutting-edge angle on the wear and tillage performance, evaluated through surface pressure and friction force. Additionally, the corresponding model verification tests and blade wear tests are conducted. The results show that the relative error of the blade shaft torque between the simulated value and the experimental value is 3.9%, and the simulated results align closely with the actual wear position on the blade surface. Wear is most pronounced on the front cutting edge, particularly under the forward rotation condition. Furthermore, increasing the cutting-edge angle of the blade from 60 degrees to 120 degrees leads to a 106.5% rise in cumulative pressure on the front cutting edge, indicating that the cutting-edge angle is a key factor influencing blade wear.

During operation, horticultural electric rotary tillers encounter unequal soil resistance between tracks due to inconsistent soil compaction. On uneven terrain, lateral tilt further causes asymmetric tillage depths, resulting in dynamic load disturbances and bias torques that impair path tracking accuracy. To resolve these issues, we propose a hierarchical path tracking control method with rotary tillage resistance disturbance compensation. A variable universe fuzzy pure pursuit controller adjusts preview distance and turning radius based on path deviation. Target motor speeds are then derived using a kinematic model. A state observer estimates rotary and driving resistance torques and applies them as feedforward compensations. Sliding mode control (SMC) ensures precise motor speed control, with its parameters optimized via a whale optimization algorithm enhanced by chaotic reverse learning, Levy flight, and inertia weight strategies. Simulations show that the SMC-based improved whale optimization algorithm with state-observer-based feedforward compensation outperforms proportional-integral-derivative (PID), fuzzy PID, and traditional SMC in speed response, steady-state error, and robustness. The integrated controller achieves mean lateral absolute errors of 0.0808, 0.0174, and 0.0299 m under S-curve, parabolic, and hyperbolic paths, respectively. Greenhouse trials show mean lateral deviations of 0.0330, 0.0862, and 0.1033 m and heading deviations of 2.7965 degrees, 2.8839 degrees, and 3.4595 degrees at tillage depths of 4, 8, and 12 cm, respectively. These results confirm full coverage of the target area and provide theoretical support for intelligent navigation in facility tillage.

The performance evaluation of fertilizer spreaders is essential for improving the efficiency and accuracy of fertilizer application in modern agriculture. Traditional methods, such as manual collection boxes, are timeconsuming, labor-intensive, and unsuitable for intelligent agricultural systems. To address these limitations, this study proposes a deep learning-based intelligent detection system for analyzing granular fertilizer deposition distribution patterns. The system integrates image acquisition, data processing, visualization, and storage functions, enabling real-time detection and operational optimization. A novel segmentation model, SF-TPP, was developed by combining a swin transformer backbone with a path aggregation network (PANet), region proposal network (RPN), ROI alignment, and enhanced by CBAM and feature refinement modules. Comparative experiments demonstrated that SF-TPP outperformed conventional models (F-TPP, SOLO, YOLOv5), achieving 92.8% precision, 91.3% recall, and an F1 score of 0.920. A strong correlation (R2 = 0.9626) was established between pixel area and fertilizer mass, enabling accurate mass estimation. Field experiments showed that the system achieved low mean absolute errors in key metrics, including a CV MAE of 0.99 and a fertilization amount MAE of 0.76 g. These results demonstrate the system's high accuracy and practical potential for supporting precision agriculture.