Accurate and efficient tomato ripeness detection is essential for automated harvesting and yield estimation in greenhouse environments. However, this task remains challenging due to three main factors: complex illumination conditions in greenhouse scenarios lead to noticeable color variation, clustered growth results in fruit occlusion, and high-accuracy models often fail to meet real-time requirements. To address these issues, the study proposes a high-performance detection model, CICE-YOLO, introduces four innovations. First, we introduce a Color Shift Estimation-and-Correction (CSEC) module to generate pseudo-normal exposure feature maps and separately estimate color deviations in bright and dark regions, thereby mitigating color distortion caused by illumination variations. Second, the YOLOv10n backbone is changed to integrate the Cross-stage Heterogeneous (CHet) module with heterogeneous convolutions to enhance fruit feature extraction. Third, an Improved Bidirectional Feature Pyramid Network (IBiFPN) with learnable fusion weights is adopted to achieve adaptive multi-scale feature aggregation, improving the detection of occluded and partially visible fruits. Finally, the CICE-YOLO detection head relies on the Efficient Intersection over Union (EIoU) loss to decouple width and height errors for more accurate bounding-box regression. Experiments conducted on a self-built dataset containing supplemental nighttime lighting, occlusion, forward lighting, and backlighting scenarios demonstrate that CICE-YOLO achieves an mAP50 of 85.9%, showing a 4.2 percentage points improvement over the baseline. The model size is 6.2 MB, and the inference time per image after offline illumination correction is 15.1 ms, which meets the real-time detection requirements. Furthermore, evaluations on a yellow-tomato dataset confirm the model robustness ability of the proposed method.

Rail welded joints are prone to geometric defects that compromise operational safety and ride quality. Traditional grinding strategies often rely on fixed rules, which fail to adapt to the diverse and irregular profiles of weld defects, leading to either excessive material removal or inefficient maintenance. To address this issue, this paper develops a reinforcement learning-based multi-objective optimization framework using a dual-critic Deep Deterministic Policy Gradient (DDPG) to generate efficient grinding strategies. The proposed model simultaneously minimizes grinding volume and the number of grinding passes, addressing both material conservation and operational efficiency. A continuous-action actor network is used to predict the optimal grinding depths and locations, while two separate critic networks evaluate the trade-offs between the competing objectives. The model is trained on high-resolution field data collected from 55 real-world rail welded joints across high-speed, conventional, and subway lines. After 6000 training steps, both critic networks converged with stable critic losses and smooth policy gradients, ensuring accurate value estimations under different grinding conditions. Across 300 sampled trade-off weights, the model generates over 220 feasible grinding strategies per case, with defined Pareto-optimal fronts. Case studies demonstrate that the optimized strategies reduce grinding volume and passes, improving profile trough restoration from -0.41 to -0.18 mm. Therefore, the proposed method can serve as an interpretable and adaptive tool to support rail maintenance decision-making by offering context-specific grinding strategies-from aggressive single-pass interventions to multi-pass, low-depth approaches that prioritize rail longevity.

To address the issue of inconsistent sowing depth caused by the forward direction surface slope (FDSS) in hilly and mountainous areas, this study developed a sowing depth control system (SDCS) based on FDSS (FDSS-SDCS), formulated a mathematical model relating FDSS, operation speed (OS), and optimal downforce, and proposed a sowing depth control strategy based on FDSS segmentation. The accuracy and response time of the shaft pin sensor and FDSS-SDCS were calibrated through indoor experiments. The experimental results demonstrate a strong correlation between the actual downforce and the response voltage output from the shaft pin sensor, with an R2 value of 0.9995; The response time of the FDSS-SDCS reaches a maximum of 0.65 s, with a maximum steady-state error of 1.70 N and a maximum overshoot of 3.63 %, all of which meet the requirements for sowing depth operations under FDSS conditions in hilly and mountainous areas. The field experiment results of the FDSS-SDCS indicate that, compared to traditional mechanical profiling springs (with spring initial increments of 10 mm, 30 mm, and 50 mm), the FDSS-SDCS with constant downforce stability control (output downforces of 300 N, 800 N, and 1300 N) reduced the maximum difference in average downforce at adjacent sampling points by 359.88 N, 435.71 N, and 467.72 N, respectively. The maximum difference in the average downforce of the FDSS-SDCS at different OSs ranged from 366.49 N to 430.75 N, indicating a significant improvement in the stability of the actual downforce under active control by the FDSS-SDCS; The range of the average sowing depth (ASD) under active control was reduced from 50 +/- 3.22 mm to 50 +/- 0.88 mm. Additionally, the qualified rate of sowing depth (QRSD) increased by 0.20 % to 17.61 %, while the coefficient of variation in sowing depth (CVSD) decreased by 0.27 % to 3.16 %. The FDSS-SDCS developed by this institute is suitable for sowing depth operations under FDSS conditions in hilly and mountainous areas, bringing the sowing depth operation performance under FDSS closer to that of non-tilting state surface.

V-belts in combine harvesters are prone to burnout failures under high-load continuous operations, compromising transmission reliability. However, the underlying thermomechanical coupling mechanisms remain poorly understood. To address this limitation, this study presents a dynamic energy dissipation model that decouples multiple heat generation mechanisms—bending hysteresis, compressive deformation, tensile strain, and external sliding friction—in V-belt drives. Finite element analysis is used to characterize the spatiotemporal loading behavior of internal friction sources, and energy dissipation quantification methods are developed based on viscoelastic theory and Coulomb friction law. A steady-state heat transfer framework is established, and critical thermal boundary parameters are identified through inverse modeling. The integrated thermomechanical model is validated by comparing simulated surface temperature fields with experimental data obtained from a custom-built test rig. The results show that predicted steady-state temperatures are marginally overestimated, with relative errors below 5%, exhibiting clear dependence on driving pulley speed and driven pulley load. The analysis reveals that high-speed operation intensifies internal hysteresis heating, whereas high-load conditions amplify external friction losses—synergistically leading to elevated thermal stress and increased burnout risk. Thus, the study elucidates the failure mechanisms under combined dynamic loads, enabling physics-driven design optimization and predictive maintenance for agricultural machinery powertrains.

[This corrects the article DOI: 10.3389/fpls.2025.1673202.].

Agricultural robots can alleviate the challenges of increasing food demand and worsening labor shortages. However, agricultural environments are typically cluttered, challenging robotic perception to find all relevant objects, as these are often hidden from view. For instance, for leaf removal and whole-truss tomato harvesting, a robot needs to efficiently detect and accurately localize the petiole and peduncle nodes of a tomato plant, which are often hidden behind leaves or tomato trusses. In this study, an integration of multiview active vision (MAV) and multiple object tracking (MOT) algorithms is proposed to improve the detection and localization of tomato plant nodes in an occluded greenhouse environment. This paper details how the system is able to collect information in a cluttered 3D environment, to generate a 3D representations and based on that effectively pinpoints plant nodes through an active-vision strategy. This study tested the MOT-MAV perception system on ten randomly selected tomato plants in an unmodified cluttered tomato greenhouse. The integration perception system was compared to a baseline experiment, which uses predefined sets of camera viewpoints and a clustering method to combine node observations from different viewpoints. Three different versions of experiments were evaluated: "MOT Exp", "MAV Exp" and the combined "MOT-MAV Exp". Under the same camera viewpoints and routes, the "MOT Exp" improved recall by 0.09 over the "Baseline Exp," confirming MOT's effectiveness in enhancing object detection. After six viewpoints, the "MAV Exp" achieved a PCO of 77%, 11 p.p. higher than the "Baseline Exp," demonstrating faster object detection with active vision. The "MOT-MAV Exp" outperformed all setups in detection efficiency and accuracy. After six viewpoints, it reached 90% of the final detection result (10 viewpoints), 17 p.p. higher than the "Baseline Exp," highlighting its faster detection speed. Additionally, its PCO reached 84%, exceeding the "Baseline Exp" by 18 p.p., showcasing improved detection accuracy. These results demonstrate that integrating MOT with next-best-view (NBV) planning enables robots to handle cluttered greenhouse environments more effectively, which improving the perception capabilities of agro-food robotic systems is essential for promoting the development of efficient tomato de-leafing and harvesting robots.

Complex environments featuring variable lighting and backgrounds similar in color to the target objects present challenges for the rapid and accurate detection of tobacco leaves, which is critical for the development of automated tobacco leaf harvesting robots. This study introduces a depth filtering approach to filter out complex regions based on distance information, thereby simplifying the detection task, and proposes a lightweight detection method based on an enhanced YOLOv5s model. Initially, the YOLOv5s backbone network is substituted with a more lightweight MobileNetV2 to reduce the model size. Subsequently, sparse model training combined with the scaling factor distribution rules of batch normalization layers is utilized to identify and eliminate inconsequential neural network channels. Finally, fine-tuning and knowledge distillation techniques are employed to achieve a model accuracy close to the YOLOv5s baseline. Experimental results indicate that the depth filtering method can improve the model's precision, recall, and mean Average Precision (mAP) by 11.2%, 29.6%, and 17.1%, respectively. The optimized lightweight model achieves a precision of 91.1%, a recall of 90.8%, and an mAP of 91.6%, with a memory footprint of only 1.4MB. It delivers a detection frame rate of 112 fps on desktop computers and 21 fps on mobile devices, which is approximately 3.5 and 4 times faster, respectively, compared to the baseline YOLOv5s tobacco leaf detection model. The precision, recall, and mAP experience a marginal decrease of 3.8, 1.6, and 2.8 percentage points, respectively, while the memory consumption is merely 10% of the pre-optimization amount. In summary, the proposed method enables the accurate detection of tobacco leaves against near-color backgrounds. Simultaneously, it achieves effective lightweighting of the model without compromising its performance, thereby providing technical support for deploying tobacco leaf detection on mobile platforms.

The 3D reconstruction of plants is challenging due to their complex shape causing many occlusions. Next-Best- View (NBV) methods address this by iteratively selecting new viewpoints to maximize information gain (IG). Deep-learning-based NBV (DL-NBV) methods demonstrate higher computational efficiency over classic voxelbased NBV approaches but current methods require extensive training using ground-truth plant models, making them impractical for real-world plants. These methods, moreover, rely on offline training with pre-collected data, limiting adaptability in changing agricultural environments. This paper proposes a self-supervised learning- based NBV method (SSL-NBV) that uses a deep neural network to predict the IG for candidate viewpoints. The method allows the robot to gather its own training data during task execution by comparing new 3D sensor data to the earlier gathered data and by employing weakly-supervised learning and experience replay for efficient online learning. Comprehensive evaluations were conducted in simulation and real-world environments using cross-validation. The results showed that SSL-NBV required fewer views for plant reconstruction than non-NBV methods. It achieved IG prediction in 0.0038s, making it over 800 times faster than a voxel-based NBV, and an online learning iteration in 0.099s. SSL-NBV reduced training annotations by over 90% compared to a baseline DL-NBV. Furthermore, SSL-NBV could adapt to novel scenarios through online fine-tuning. Also using real plants, the results showed that the proposed method can learn to effectively plan new viewpoints for 3D plant reconstruction. Most importantly, SSL-NBV automated the entire network training and uses continuous online learning, allowing it to operate in changing agricultural environments.

Accurate segmentation of key tobacco structures is essential for enabling automated harvesting. However, complex backgrounds, variable lighting conditions, and blurred boundaries between the stem and petiole significantly hinder segmentation accuracy in field environments. To overcome these challenges, we propose an enhanced instance segmentation approach based on YOLOv8-seg, incorporating depth-based background filtering and architectural improvements. Specifically, depth information from RGB-D images is employed to spatially filter non-target background regions, thereby enhancing foreground clarity. In addition, a Hybrid Dilated Residual Attention Block (HDRAB) is integrated into the YOLOv8-seg backbone to improve boundary discrimination between petioles and stems, while a Lightweight Shared Detail-Enhanced Convolution Detection Head (LSDECD) is designed to efficiently capture fine-grained texture features. Experimental results demonstrate that depth filtering increases mAP50bb and mAP50seg by 7.9% and 6.3%, respectively, while the architectural enhancements further raise them to 89.5% and 91.1%, surpassing the YOLOv8-seg baseline by 5.2% and 10.0%. Compared with mainstream models such as Mask R-CNN and SOLOv2, the proposed method achieves superior segmentation accuracy with low computational cost, highlighting its potential for practical deployment in automated tobacco harvesting

Clustered tomato maturity is a key factor in the harvesting decision-making process of automated harvesting robots. To enable all-day operation, accurate maturity recognition under different lighting conditions is essential. This study explores a deep learning-based machine vision approach for clustered tomato maturity recognition in three lighting scenarios: daytime, nighttime without supplementary lighting, and nighttime with supplementary lighting. To enhance recognition performance, a nighttime image processing method based on Zero-DCE for illumination enhancement and deep white-balance for color correction was proposed. Additionally, an improved object detection model, YOLOv5s-DRSN, was developed by incorporating a deep residual shrinkage network (DRSN) to suppress noise interference and improve detection accuracy. Experimental results show that the mean absolute error (MAE) under daytime, nighttime with and without supplementary lighting was 2.88%, 10.67%, and 4.63%, respectively. The harvesting decision accuracy reached 100% in both daytime and nighttime without supplementary lighting, while it slightly decreased to 83.3% under nighttime with supplementary lighting. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)

Accurate perception in complex agricultural environments is challenging due to significant plant occlusion, primarily from leaves, which hinder data collection and increase uncertainty in robotic operations. Deep-learning-based Next-Best-View (DL-NBV) methods address this by using neural networks to predict information gain (IG) for potential camera views and actively repositioning the camera to maximize data collection with minimal views. However, training DL-NBV models requires extensive IG-labeled data. A self-supervised learning-based NBV method, SSL-Global-NBV, enables robots to collect training data autonomously and improve themselves for global NBV planning. Despite its advantages, SSL-Global-NBV has two key limitations: (1) it requires a fixed number of views, limiting scalability across different plant sizes, and (2) it selects views globally, resulting in inefficient view transitions across the entire view space, reducing trajectory efficiency. To overcome these limitations, this paper introduces SSL-Local-NBV, which incorporates local view planning for scalable and efficient view selection. To prevent redundant visits to the same views, a View Trajectory Network (VTN) was proposed to memorize the view trajectory information of visited views. Comprehensive evaluations in simulation and real-world plant reconstruction demonstrated that SSL-Local-NBV reduced trajectory distance by 56%-70% per reconstruction cycle, achieving 267%-300% higher trajectory efficiency than global NBV methods. Compared to SSL-Global-NBV, SSL-Local-NBV improved plant reconstruction efficiency by 5.2% across varying plant sizes, demonstrating greater scalability. For real plants, SSL-Local-NBV achieved over 80% reconstruction, confirming its feasibility in practical applications. Notably, SSL-Local-NBV fully automated training through self-supervised learning, enabling continuous and lifelong robotic learning.

Efficiently, accurately, and realistically reconstructing large-scale 3D orchard scenes in a virtual world is an immensely challenging task. This complexity stems from the intricate and expansive of real orchard scenes. Traditional 3D reconstruction and rendering methods have encountered limitations in terms of modeling efficiency and computational costs, hindering the ability to provide users with immersive experiences. In response to these challenges, this study introduces a strategy for 3D scene reconstruction and rendering grounded in implicit neural representation: the NeRF-Ag model. Building upon the baseline NeRF, this model integrates a multiresolution latent feature encoding technique, notably heightening training efficiency and elevating modeling precision. Furthermore, by means of environmental factor embedding, the model's robustness and practical applicability are further enhanced. The experimental outcomes illustrate that NeRF-Ag attains photo-realistic rendering outcomes across small, medium, and large scales. Moreover, it surpasses NeRF concerning the evaluation metrics of PSNR, SSIM, and LPIPS. Notably, the training speed of NeRF-Ag is roughly 39 times faster than NeRF. In 3D reconstruction tasks, NeRF-Ag showcases enhanced texture detail representation and higher modeling accuracy compared to the COLMAP-based 3D reconstruction method. Additionally, this study accomplishes free-viewpoint rendering of 3D scenes employing NeRF-Ag and provides evidence substantiating the connection between the quantity of training images and the precision of 3D rendering. The conclusions of this study will contribute to supporting and referencing the implementation of immersive visual interactive features within agricultural digital twin systems.

In robotic harvesting, maneuvering around obstacles to position manipulators is challenging, especially in unstructured environments. This study proposes a method to detect the relative position of tomato bunches to the main stem position using the BlendMask-BiFPN algorithm. Initial comparative tests between full-stem and partial-stem labelling strategies revealed that the latter produced more complete peduncle masks, which guided our choice for subsequent experiments. Significant modifications to the BlendMask algorithm included the integration of a ResNet-101-BiFPN backbone, which improved the feature fusion network of the model. The revised model demonstrated high efficiency in pinpointing the relative positions of clustered tomatoes, achieving 91.3 % ARmask 50 and 84.8 % APmask 50 for the detection of tomato bunches. Comparisons with Mask RCNN, YOLACT, YOLACT++, and YOLOv8 showed that the BlendMask-BiFPN model outperforms these alternatives, suggesting its potential for more effective robotic harvesting in complex agricultural scenarios.

Possibility of application of the low frequency inductive heating for heating ferromagnetic has been considered and discussed. On the basis of simulations and investigated results the heater structure for railway rails was developed.

As the material living standards of residents rapidly improve, the demand for fruits and vegetables continues to rise. Agricultural greenhouses, as a critical means of increasing fruit and vegetable yields, require effective control of temperature, humidity, and light intensity to ensure rapid crop growth. Traditional manual monitoring and control methods are inefficient and labor-intensive. Against the backdrop of rapid developments in electronic technology, this paper innovatively designs a greenhouse IoT intelligent control system based on a microcontroller, achieving real-time monitoring and automated control of greenhouse environment parameters. The innovations of this system are evident in several aspects: Firstly, it employs an STM32 microcontroller as the main control chip, integrating YL-69 soil moisture sensors, GL5506 photodiodes, and DS18B20 temperature sensors to achieve high-precision detection of soil moisture, light intensity, and temperature. Secondly, an LCD1602 display is used to timely showcase real-time environmental parameters, and alarm thresholds can be set via buttons. If these thresholds are exceeded, a buzzer sounds an alarm. Concurrently, the system utilizes water pumps and supplementary LED lights to intelligently adjust humidity and light intensity within the greenhouse. Most importantly, by incorporating the ESP8266 module, the system achieves remote data transmission, allowing users to view real-time environmental parameters via a mobile app and set thresholds and control operations. This greenhouse IoT intelligent control system not only features alarm functions for exceeded environmental parameters but also enables automatic adjustment through intelligent devices and supports remote control, greatly enhancing operational convenience and system practicality. Overall, this system innovatively integrates various sensing technologies and IoT communication, providing an efficient and intelligent solution for the modern management of agricultural greenhouses. It exhibits broad market prospects and holds significant importance for advancing agricultural modernization.

Searching and detecting the task-relevant parts of plants is important to automate harvesting and de-leafing of tomato plants using robots. This is challenging due to high levels of occlusion in tomato plants. Active vision is a promising approach in which the robot strategically plans its camera viewpoints to overcome occlusion and improve perception accuracy. However, current active-vision algorithms cannot differentiate between relevant and irrelevant plant parts and spend time on perceiving irrelevant plant parts. This work proposed a semantics-aware active-vision strategy that uses semantic information to identify the relevant plant parts and prioritise them during view planning. The proposed strategy was evaluated on the task of searching and detecting the relevant plant parts using simulation and real-world experiments. In simulation experiments, the semantics-aware strategy proposed could search and detect 81.8% of the relevant plant parts using nine viewpoints. It was significantly faster and detected more plant parts than predefined, random, and volumetric active-vision strategies that do not use semantic information. The strategy proposed was also robust to uncertainty in plant and plant-part positions, plant complexity, and different viewpoint-sampling strategies. In real-world experiments, the semantics-aware strategy could search and detect 82.7% of the relevant plant parts using seven viewpoints, under complex greenhouse conditions with natural variation and occlusion, natural illumination, sensor noise, and uncertainty in camera poses. The results of this work clearly indicate the advantage of using semantics-aware active vision for targeted perception of plant parts and its applicability in the real world. It can significantly improve the efficiency of automated harvesting and de-leafing in tomato crop production.

Oats (Avena sativa L.) are rich in nutrients and bioactive compounds, serving as a roughage source for ruminants. This study investigated the impact of lactic acid bacteria (LAB), cellulase (M), and their combinations (LM) on the fermentation quality and metabolic compounds of oat silage. Results demonstrated that all additive treatments significantly increased lactic acid content compared to the control group (P < 0.05), with the lactic acid bacteria treatment group exhibiting the lowest pH value (P < 0.05). Analysis of antioxidant activity and metabolites in oat silage over 60 days revealed 374 differential metabolites with 113 up-regulated and 261 down-regulated, and all treatment groups showing higher antioxidant activity than raw oat materials (P < 0.05). Although no significant differences in antioxidant activity were observed among the various treatment groups in this experiment, notable changes in metabolic pathways were identified. Furthermore, two metabolites (carboxylic acids and derivatives and benzene and substituted derivatives) were identified through non-targeted metabolomics technology, both of which are strongly associated with the antioxidant activity of oat silage. This finding provides a theoretical basis for the efficient use of oat silage in animal husbandry.

Key performance of high-stability current source, indicators such as output stability and noise spectrum within an extended bandwidth have become crucial determinants in precision measurement systems. This paper introduces a novel design of a Voltage Controlled Current Source (VCCS) featuring bipolar output. This design integrates a cascaded transistor push-pull configuration with a DC Current Transformer (DCCT) based current feedback circuit, enabling current outputs up to 2 A. Rigorous experimental evaluations have substantiated proficiency in delivering minimal current ripple, exceptional stability, and versatility in output power range. The apparatus exhibits peak noise at $2.3 \mu \mathrm{A} / \mathrm{Hz}^{1 / 2}$ at 50 Hz and statistical uncertainty of 0.38 ppm at 2 A.

Greenhouse production of fruits and vegetables in developed countries is challenged by labor 12 scarcity and high labor costs. Robots offer a good solution for sustainable and cost-effective 13 production. Acquiring accurate spatial information about relevant plant parts is vital for 14 successful robot operation. Robot perception in greenhouses is challenging due to variations in 15 plant appearance, viewpoints, and illumination. This paper proposes a keypoint-detection-based 16 method using data from an RGB-D camera to estimate the 3D pose of peduncle nodes, which 17 provides essential information to harvest the tomato bunches. 18 19 Specifically, this paper proposes a method that detects four anatomical landmarks in the color 20 image and then integrates 3D point-cloud information to determine the 3D pose. A 21 comprehensive evaluation was conducted in a commercial greenhouse to gain insight into the 22 performance of different parts of the method. The results showed: (1) high accuracy in object 23 detection, achieving an Average Precision (AP) of AP@0.5=0.96; (2) an average Percentage of 24 Detected Joints (PDJ) of the keypoints of PhDJ@0.2=94.31%; and (3) 3D pose estimation 25 accuracy with mean absolute errors (MAE) of 11.38o and 9.93o for the relative upper and lower 26 angles between the peduncle and main stem, respectively. Furthermore, the capability to handle 27 variations in viewpoint was investigated, demonstrating the method was robust to view changes. 28 However, canonical and higher views resulted in slightly higher performance compared to other 29 views. Although tomato was selected as a use case, the proposed method is also applicable to 30 other greenhouse crops like pepper.