Aiming at the problems of poor seed distribution uniformity and low sowing accuracy in traditional sowing operations, this paper studies and develops a variable-rate monitoring and control technology system based on an unmanned tractor-towed seeding machine. The system adopts a distributed controller architecture, including a cooperative control ECU, a seed-metering monitoring and control ECU, and a sowing depth monitoring and control ECU, achieving multi-source information interaction and cooperative decision-making through a CAN bus. In terms of hardware, the seed-metering device's electric drive scheme has been improved by replacing the original split-type chain transmission scheme with an integrated torque servo motor, enhancing the system integration and transmission efficiency. Field test results show that the optimized electric drive seed-metering device, at forward speeds of 4 km/h, 8 km/h, and 12 km/h, outperforms the traditional ground wheel-driven method in terms of seed spacing qualification index, standard deviation, and coefficient of variation. Especially under the medium-speed condition of 8 km/h, the performance is optimal, with the coefficient of variation as low as 6.45%. The system demonstrates good robustness and adaptability in complex field environments, providing reliable technical support for achieving precise, efficient, and intelligent sowing operations.

Effective disinfection and disease prevention are crucial for large-scale poultry farming operations. However, existing poultry house disinfection robots suffer from issues such as slow spray response, poor uniform coverage, and inconsistent manual operation. Therefore, this paper proposes a spray parameter optimization method based on response surface regression and Bayesian optimization. By optimizing key parameters, including spray flow rate, spray distance, nozzle height, and robot walking speed, the precision and stability of spray operations are enhanced. This method integrates response surface regression modeling, Bayesian optimization, and fuzzy PID control technology. First, a spray control system was designed centered on a microcontroller, employing fuzzy PID closed-loop control technology. This was combined with genetic algorithm optimization of spray control strategies to enhance control accuracy and precision. Next, a quadratic polynomial response surface model was constructed using Box-Behnken design. The model's significance was verified via analysis of variance (P<0.0001), with a goodness-of-fit R & sup2; of 0.9599. Furthermore, a multi-start Bayesian optimization algorithm was employed for global parameter optimization. Results indicated that at a spray flow rate of 2.82 L/min, spray distance 80 cm, walking speed 0.4 m/s, and nozzle height 61 cm above the chicken cage floor, spray coverage reached 95.00%.In the eight validation tests, the measured average coverage rate was 95.50%, with an overall relative error of 0.53%. Additionally, the coefficient of variation (CV) of the spatial distribution of spray deposition among different monitoring points in the validation tests ranged from 4.07% to 5.06%, indicating relatively low spatial variability and acceptable coverage uniformity under the optimal parameter combination. The results demonstrate that the proposed method can effectively improve coverage control accuracy and maintain relatively stable deposition distribution during poultry house spraying operations, providing a reference for parameter optimization and spray control in future disinfection applications. Comparative tests further demonstrated that, under identical operating conditions and the same optimal parameter combination, the optimized PID control method outperformed both traditional PID control and manual backpack spraying in terms of coverage accuracy and spatial distribution stability. Control system experiments demonstrated that under both constant and variable speed conditions, the flow adjustment response time was 2-3 seconds, with steady-state error below +/- 0.05 L/min, exhibiting excellent dynamic tracking and stability. These findings validate the effectiveness of Bayesian optimization and fuzzy PID control in parameter optimization for poultry house disinfection spray systems, providing theoretical foundations and technical support for developing intelligent epidemic prevention equipment in livestock and poultry farms.

Leaf area index, plant height, and above-ground biomass are key physicochemical parameters that reflect a crop's growth status. Accurate and efficient quantitative estimation of these indicators is crucial for developing production management strategies and predicting yields. UAV remote sensing imagery offers advantages such as mobility, flexibility, and high data collection efficiency, making it practical for monitoring maize growth indicators. However, existing parameter inversion models are significantly affected by background noise such as soil and shading. Furthermore, the asymptotic saturation phenomenon in optical sensor observations limits the ability of spectral information to assess key crop parameters. This study aims to investigate whether background pixels can be removed from UAV imagery and whether utilising canopy thermal information extracted from UAV thermal infrared images can improve the accuracy of crop parameter inversion. First, UAV images were segmented using vegetation indices. Vegetation indices, texture features, and temperature characteristics were extracted for AllPix and GreenPix within the experimental plots. Subsequently, Correlation analysis and random forest-based feature selection were employed to screen the characteristic parameters. Finally, we employed multiple linear stepwise regression (MLSR), partial least squares regression (PLSR), random forest regression (RFR), Gradient Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) to establish LAI, PH, and AGB inversion models for four critical growth stages of spring maize with a stratified validation strategy. The results indicated that spring maize LAI, PH, and AGB showed the strongest correlation coefficients with MRETVI, percentage of hot spot area, and mean temperature, respectively, at −0.856, −0.781, and −0.858. The results indicated that the model developed by integrating AllPix and GreenPix features exhibited the best performance. The optimal validation set for LAI yielded determination coefficients of 0.8579, with root mean square errors of 0.7861 m²/m². For PH, the determination coefficients of the optimal validation set is 0.9605, with root mean square errors of 20.8327 cm. For AGB, the optimal validation sets yielded determination coefficients of 0.9094, with root mean square errors of 83.1961g. Compared with models based on all pixels, the models based on spring maize pixel provided higher estimation accuracy for LAI, PH, and AGB. In particular, for LAI, the validation R² increased by 0.0270 and the RMSE decreased by 0.0714 m²/m². For PH, the validation R² increased by 0.0490 and the RMSE decreased by 9.8603 cm. For AGB, the validation R² increased by 0.1219 and the RMSE decreased by 36.6957 g. These findings provide technical support for monitoring spring maize growth and optimising production management.

This study addresses the problem of inconsistent and unstable sowing depth in conventional wheat planters. A novel measurement and control strategy based on mulch quantity regulation is proposed for a dual-axis rotary planter. A mathematical model of sowing depth is established and a corresponding adjustment unit is designed. An adaptive fuzzy PID controller is developed and compared with a conventional PID controller. The adaptive fuzzy PID controller achieves a rise time of 0.19 s and a stabilization time of 0.26 s. Compared with the conventional PID controller, these values represent reductions of approximately 26.2% and 14.9%, respectively. Simulation results confirm the controller's effectiveness and adaptability. The system is then implemented on a dual-axis rotary-tillage wheat planter and validated through field experiments. The proposed method enables precise and stable control of sowing depth, improving both the quality and consistency of wheat planting operations.

To address rising false negatives in automated cherry‑tomato sorting, multi‑scale detection challenges, and the difficulty of deploying advanced models on edge devices, we propose an improved YOLOv10m‑based model. First, the neck’s original concatenation (Concat) is replaced with an Adaptive Multi‑Scale Feature Fusion (AMSFF) module to enhance feature fusion. A dynamic scale‑selection mechanism is introduced to improve detection of challenging, small, and multi‑scale targets. Partial Convolution (PConv) is then applied to the neck’s downsampling operators to reduce computational cost while retaining essential performance. Finally, Ghost Depthwise Convolution (GhostDWConv) substitutes the baseline Depthwise Convolution (DWConv) to better balance efficiency and representational capacity. Compared with the baseline, the enhanced architecture increases recall by 7.06% and improves mean Average Precision by 2.12% (mAP@50) and 2.49% (mAP@[50:95]). Meanwhile, model parameters, model size, and FLOPs are reduced by 26.14%, 25.97%, and 28.55%, respectively. When deployed on the NVIDIA Jetson Orin Nano, the proposed model attains an inference throughput of 58.6 FPS, outperforming industrial camera frame rates and fulfilling real-time constraints for inline cherry tomato sorting. This advancement not only enhances detection precision but also contributes to model lightweighting, thereby offering tangible benefits for improving sorting efficiency and ensuring product quality in cherry tomato production lines.

High-quality, real-time three-dimensional (3D) reconstruction and phenotyping of field crops are essential for advancing digital farm management. However, challenges such as diverse crop postures and fluctuating illumination in open-field environments significantly hinder reconstruction accuracy. To address these issues, this study proposes AgriGaussian, a novel 3D phenotyping method based on 3D Gaussian Splatting (3DGS), designed to enable low-cost, high-fidelity reconstruction and phenotypic analysis of agricultural fields, with chili pepper plants as an example. Specifically, a dataset encompassing three distinct growth stages of chili pepper plants was collected and preprocessed using a phenotyping robot (PR) equipped with oblique-view imaging. The AgriGaussian method integrates three core components: (1) a Depth-Supervised Strategy (DSS) for adaptively enhancing phenotypic features based on gradient variations; (2) a Crop Appearance Extraction (CAE) module that suppresses the effects of illumination and environmental interference while preserving fine phenotypic details; and (3) a Multi-scale Chunking Training (MCT) strategy that enables scalable reconstruction of large agricultural scenes. Experimental results demonstrate that AgriGaussian achieves high-fidelity reconstruction, with peak signal-to-noise ratio (PSNR) of 26.85 dB, structural similarity index (SSIM) of 0.86, and learned perceptual image patch similarity (LPIPS) of 0.24—surpassing existing baseline algorithms. Compared to manual measurements, the reconstructed canopy height and volume exhibit strong agreement, with coefficients of determination (R2) reaching 0.90 and 0.86, respectively, and average absolute percentage errors as low as 5.1%. Furthermore, the method accurately reconstructs fine-grained features such as foliar lesions and surface textures. In summary, AgriGaussian enables large-scale, high-fidelity reconstruction of field crop phenotypes and provides a promising foundation for advancing 3D phenotyping and digital twin technologies in precision agriculture.

The core objective of this study is to address critical challenges in the operational monitoring and fault early warning of wheat combine harvesters. To this end, this study designed a field-oriented multi-parameter detection system for wheat combine harvesters, which utilizes the CAN bus and virtual instrumentation. Key challenges in this field include three aspects: first, manual inspection is inefficient and lacks automated detection methods, making it difficult to meet the real-time requirements of large-scale operations; second, fault early warning accuracy is low, as single-parameter evaluation is prone to false positives and false negatives; third, monitoring parameters function in isolation, leading to significant data inconsistencies that hinder the early detection of potential faults. To address these issues, this study focuses on three key tasks: establishing a multi-parameter collaborative monitoring framework, optimizing hardware and communication protocols, and developing data processing methods for fault detection and warning. Specifically, sensors for fuel consumption, Hall-effect rotational speed, and strain-gauge torque are deployed at critical components of the harvester. The system then efficiently transmits operational status data via the CAN bus to a processing module, enabling remote real-time monitoring of the harvester's comprehensive operational conditions. For the designed fault warning algorithm, it dynamically adjusts warning thresholds by comparing characteristic parameters with historical data, thereby achieving accurate fault identification and timely warning responses. This study innovatively transmitted multi-source sensor data through the high-anti-interference CAN bus and developed a fault warning algorithm incorporating feature recognition and dynamic thresholds. In simulated experiments, the measurement errors of both instantaneous and cumulative fuel consumption were <= 5%, while the system achieved a warning accuracy of 97.3% and a response time of <= 180 ms. This represents a 15.3-percentage-point improvement in accuracy compared to traditional single-parameter warning systems. Overall, this study addresses the challenge of multi-parameter integrated monitoring for wheat combine harvesters and provides a scalable technical solution for hardware integration and comprehensive data analysis. It also offers a reference for the intelligent upgrading of Chinese harvesters, which is expected to accelerate the transformation of mechanization toward and informatization.

Addressing the issues of severe resource wastage, low automation levels, and insufficient control precision in traditional orchard water and fertiliser management, this paper proposes a multi-dimensional, collaborative integrated water and fertiliser management system for orchards. This system combines IoT sensing technology, wireless communication technology, cloud platform computing, and intelligent control algorithms. Through a layered architecture design, this system innovatively enhances the perception layer, transmission layer, data layer, and application layer. It integrates diverse communication protocols such as ZigBee, LoRa and NB-IoT, combines sensors for irrigation, fertilisation, and other aspects, and merges intelligent algorithms including PID closed-loop control and fuzzy neural networks. This enables precise water-fertiliser ratio formulation, remote real-time monitoring, and adaptive regulation. Field validation at a demonstration orchard in Guangdong demonstrated water savings of 30%–65%, fertiliser reductions of 40%–50%, and crop yield increases of 20%–135%. Research findings indicate that the deep integration of IoT and cloud platforms significantly enhances water and fertiliser utilisation efficiency in orchards while reducing labour costs. This provides technical support for standardised orchard cultivation, aligning with the demands for precision and intelligent development in modern agriculture.

To address the difficulty of real-time detection of seed-filling performance in pneumatic suction seed metering devices under high-speed operation—where seed targets are tiny, prone to adhesion, and affected by motion blur—this paper proposes a lightweight online detection algorithm, YOLOv8n-MA. First, according to the seed adsorption characteristics of the suction holes, the detection targets are divided into three categories: none, one, and two. Second, based on YOLOv8n, the backbone network is replaced with MobileNetV1 to reduce computational cost, and an ACmix attention module is integrated into the Neck to enhance feature representation for the three suction-hole states. Finally, to meet the demand for low-latency inference on resource-constrained devices, the model is deployed on an edge computing controller to achieve real-time detection. Experimental results show that, compared with the original YOLOv8n, the parameters and FLOPs of YOLOv8n-MA are reduced by 34.4% and 59.8%, respectively, while the mean average precision (mAP) is improved by 2.0% to 96.8%, achieving a superior trade-off between accuracy and efficiency over other detection models of the same category, such as YOLOv5n, YOLOv9n, and YOLOv10n. In field tests, the detection accuracy reaches 95.02% at 12 km/h and 92.65% at 15 km/h. The proposed method provides effective technical support for the intelligent monitoring and control of precision seeding under high-speed operation.

To address the problems of unstable sowing depth and poor system coordination in corn precision sowing operations, an integrated monitoring and control system was developed. The system achieves closed-loop control of sowing depth by applying controllable downward pressure via a hydraulic circuit, combined with feedback from pin-type pressure and angle sensors. A coupled cooperative controller (SPC-SFMC-X2214A) was implemented to connect the tractor and planter CAN networks, enabling navigation data parsing and fault linkage. A CODESYS-based interface was developed for real-time data visualization and parameter configuration. Field tests showed that at operating speeds of 6-10 km/h, the sowing control error remained <= 2.00%. The response time of the seeding rate was 0.85 s (for 90-225 kg/hm2), exceeding the design requirement of less than 1 s. The developed system provides an intelligent and adaptive solution for improving the quality of corn precision planting.

Precise perception of crops and weeds is crucial for spot-spraying weeding (SSW). However, existing methods rely heavily on manually annotated data and are susceptible to lighting variations in agricultural environments, thereby limiting the robustness, stability, and real-time performance of SSW perception systems. To address these challenges, this study proposed Weedformer, a real-time perception frameworks that integrated crop contour extraction (CCEnet) with the weed index to enable precise segmentation and spraying of early-stage corn and weeds. First, CCEnet combined a lightweight pixel decoder (LPD) with a dual-core transformer decoder (DTD). This design preserved model compactness while substantially improving real-time contour extraction, reducing the model size to 31.5 MB while sustaining 57.2fps. In addition, we designed a Crop Contour Loss (CCL) to improve contour extraction by dynamically predicting key points. Overall, the model achieved an mAP of 98.49% and a recall of 96.32% for crop contour segmentation, outperforming existing models. Second, the proposed Weed Index converted the non-crop regions extracted by CCEnet to the YCrCb color space and incorporated VARI to extract weeds, thereby mitigating the effects of illumination variation on weed segmentation. Even without manual annotations, this method achieved a Dice score of 0.87 and an accuracy of 90.97% for weed segmentation. Experiments demonstrated that Weedformer achieved overall weed accuracy and HIT rates of up to 98.26% and 93.51%, respectively, in practical SSW systems, with pesticide savings reaching 42.44%. The proposed method provided a reliable technical foundation for achieving precise, targeted spot spraying with intelligent application equipment. The code for the proposed methods was open-sourced at: https://github.com/anqilin8/Weedformer.

Straw coverage serves as a critical indicator in the realm of conservation tillage. This study aims to fulfill the detection needs for straw coverage on edge monitoring platforms by initially capturing straw images through an onboard terminal and subsequently creating a dataset via data augmentation. We opted for SegNext as the foundational model and incorporated ResNet101 as the backbone to enhance the extraction of features specific to straw. To achieve a lightweight model without sacrificing detection accuracy, ResNet101 was utilized as the teacher model to mentor ResNet18 as the student model, with the training outcomes quantified using QAT. In tests conducted under multifactorial field scenarios, the QSR101-18 model achieved mIoU of 85.78 %, mAP of 95.98 % and Kappa of 86.25 %, surpassing SegNext by 1.44 %, 1.57 % and 1.32 %, respectively. The QSR101-18 model FLOPs and Params are 0.71G and 0.45 M respectively, which is about 1/27 and 1/100 of SegNext. When deployed on edge platforms and analyzed across varying straw coverage rates, QSR101-18 demonstrated an overall error of only 1.3 %, well within acceptable limits. The inference speed for a single image was just 16.32 ms, meeting the speed requirements for field operations. Consequently, the proposed QSR101-18 model demonstrates several key advantages, including a lightweight architecture, minimal error rates, robustness, and high accuracy. It effectively addresses the challenges posed by unstructured, fragmented straw and various environmental factors in detecting straw coverage, all while adhering to the speed constraints required for field operations on edge monitoring platforms.

Accurate identification and localization of peduncle cutting points are crucial for the automated harvesting of tomatoes. Due to the slender nature of tomato peduncles, occlusions from surrounding fruits, stems, and other obstacles often occur, which can adversely affect the accuracy of harvesting point detection. An optimal observation viewpoint of the tomato clusters can significantly enhance the visibility of peduncles within the camera frame. This study presents a pose estimation method for tomato cluster observation based on semantic segmentation, aimed at improving peduncle recognition accuracy from the end-effector camera's perspective. A lightweight semantic segmentation network, Dual-Resolution Network with Convolutional Attention (DRCANet), is developed to efficiently identify tomatoes and stems in harvesting scenes. The DRCANet adopts a dual-branch structure that incorporates the Convolutional Attention (CA) Block in the low-resolution semantic branch to enable more efficient semantic feature extraction. Further optimization of model performance is achieved by integrating a Multi-Scale Convolution with Channel Excitation Module (MSCEM), the adaptive-weighted-fusion module (AWF), and shallow feature fusion. The proposed DRCANet predicts masks for both tomatoes and stems in the images. By combining these predicted masks with depth information, the spatial point cloud data of tomatoes and stems are extracted. The spatial relationship between each tomato cluster and its corresponding stem is then analyzed, leading to the final observation pose estimation for each tomato cluster. Experimental results demonstrate that the proposed DRCANet achieves mIoU and mPA values of 82.83 % and 91.37 %, respectively, with an average inference time of 11.42 ms. The proposed observation pose estimation method achieves an accuracy of 77.84 % with an average processing time of 68.25 ms. This study validates the effectiveness of optimizing the observation perspective in improving the recognition accuracy of tomato peduncle picking points, offering a novel approach to enhancing the harvesting success rate of tomato harvesting robots.

While the root architecture of potted crop seedlings directly determines subsequent crop productivity and adaptability, these root systems remain challenging to quantify using conventional methods due to their structural complexity. To investigate the microscopic characteristics of the root systems of pepper seedlings within pots, Micro-CT was employed to scan the seedling pots. After three-dimensional (3D) reconstruction was conducted on the data acquired from the pot scans, the 3D model of the root system was segmented and extracted using the watershed algorithm. Vertically, the three-dimensional root model was divided from top to bottom into four equally spaced regions (a, b, c, and d), showing the volumetric distribution characteristics of pepper seedling roots within the pots. The results showed that region a had the largest average root volume proportion (29.72%), primarily due to the substantial volume contribution of the taproot. Region d followed with an average proportion of 27.26%, resulting from root coiling and entanglement at the pot bottom caused by the spatial constraints of the seedling tray. The middle regions of the pot, b and c, showed average root volume proportions of 23.14% and 19.89%, respectively. To further investigate the influence of root system characteristics on root injury during seedling gripping, the seedlings were categorized into three types based on their taproot growth positions. A gripping experiment was conducted on these three seedling types using spatula-equipped needles. The results showed that the greatest root injury (12.67%) was observed in Type 1 seedlings, which had taproots located closest to the needle insertion point. In contrast, the least injury (4.09%) was found in Type 3 seedlings, characterized by centrally positioned taproots. Type 2 seedlings, with their taproots growing on the side (laterally away from the insertion point), sustained intermediate injury (5.45%). This was because their lateral positioning led to an uneven distribution of mechanical stress during gripping compared with Type 3 seedlings. A validation experiment conducted on an automated seedling retrieval platform confirmed the root injury analysis. The experimental results showed maximum root injury in Type 1 seedlings (14.16%), followed by Type 2 (6.03%) and Type 3 (4.82%) seedlings, with a successful retrieval rate of 95.29%. These findings were consistent with the Micro-CT analysis. This study could provide a theoretical foundation for low-injury seedling gripping in fully automated seedling transplanters.

Against the backdrop of precision agriculture and the development of intelligent agricultural machinery, current domestic monitoring systems for wheat combine harvesters are plagued by limited functionality, low intelligence, significant errors in parameter monitoring, and yield estimation results prone to inaccuracies. Specifically, they lag behind mature international systems in terms of fault warning accuracy, data transmission efficiency, and yield visualization capabilities. This study seeks to realize comprehensive and precise monitoring, reliable fault early warning, and intelligent yield prediction for wheat combine harvesters across all operating conditions. To this end, it innovatively adopts CAN bus integration technology and impulse-type grain flow sensors to develop a comprehensive system for monitoring the operational status and warning faults of wheat combine harvesters, which covers the entire operational process. By integrating GPS positioning, multi-sensor parameter acquisition, and intelligent analysis modules through CAN bus integration, the system enables unified monitoring of geographic information, operational data, cleaning loss, and fault status. Additionally, it incorporates a yield measurement module based on an impulse-type grain flow sensor to generate the real-time yield distribution maps. Field experiments demonstrate that the system achieves an alarm accuracy of 97.3%, controls the fuel consumption measurement error within 5%, and limits the relative error of yield measurement accuracy to no more than 4%. Notably, the impulse-type grain flow sensor exhibits stable static detection accuracy and rapid, precise dynamic measurement performance-laying a solid foundation for the automation and intelligent advancement of combine harvester technologies.

To effectively solve the problem of high damage rates and low operating efficiency of clamping and conveying in the mechanized harvesting of Chinese cabbage, a "vertical clamping+flexible conveying" system was developed. Based on the measurement and analysis of the basic physical characteristics and static compression mechanical properties of Chinese cabbage, the "vertical clamping + flexible conveying" method was applied to arrange the clamping conveyor belts longitudinally. A combination of flexible feeding and soft clamping was used to achieve low-damage transportation. A dynamics coupling simulation model of the Chinese cabbage harvesting components was established. By adjusting the structural and operational parameters of the harvesting components, a simulation test of the Chinese cabbage harvesting operation was conducted to determine the kinematic and dynamic principles of the harvesting process. The designed and developed clamping conveyor device is installed on a cabbage harvester for field performance test verification. Field test results show that the field productivity of the cabbage harvester is 0.12 hm2/h, the average value of the clamping and conveying success rate is 96.38%, and the average value of the harvesting damage rate is 7.43%. The developed clamping and conveying device can effectively meet the requirements of high efficiency, and low energy consumption, low damage during harvesting, while also enhancing adaptability to Chinese cabbages of varying head diameters.

Transplanting represents a vital technique in contemporary vegetable cultivation; mechanized transplanting entails a complex process characterized by interactions among machine, soil, and plants, which ultimately constrains the quality of mechanized transplanting. Currently, due to the unclear mechanism of machinesoil-plant interaction, it is difficult to solve the problem of transplanting verticality. To this end, this study established a machine-soil-pot seedling coupled simulation model through EDEM-Recurdyn co-simulation to analyze the disturbance of the soil and pot seedling by the transplanter during the transplanting process. Simulation experiment results indicate that the trajectory eigenvalue of the transplanting point and the opening angle of the duckbill are the main factors affecting transplanting effects. Through regression modelling, it is predicted that when the eigenvalue lambda is 1.1, the duckbill opening angle gamma is 33.3 degrees, and the opening time t is 0.4 s, the transplanting effects are optimal, corresponding to the soil backflow ratio of 0.44, the soil disturbance ratio of 1.29, and the verticality of 79.83 degrees. By comparing the results of the simulation with field experiments, the average errors of soil backflow ratio, soil disturbance ratio, and verticality were 2.14 %, 2.46 %, and 1.86 %, respectively. The experiment results show that the simulation model can effectively reflect the physical changes during the transplanting process and can provide a reference for the optimization of the transplanting device design.

Highlights This research proposes a model based on DTs and BPNN and accurately predicts the growth indexes and state of lettuce. Abstract. This research proposed a full-space state prediction model based on Digital Twins (DTs) for intelligent prediction and optimization control of environmental parameters and crop growth in plant factories. Compared with traditional prediction models, this model significantly improved production efficiency and resource utilization in plant factories by dynamically adjusting environmental control strategies through real-time data collection and feedback. The model employed a Back Propagation Neural Network (BPNN) for accurate prediction of crop growth indexes, with experimental results showing a Root Mean Squared Error (RMSE) of 0.868 and a Mean Absolute Error (MAE) of 0.625 on the test dataset, indicating high prediction accuracy. The innovative aspect of this model lies its integration of DTs technology, enabling full-cycle monitoring and intelligent regulation of the crop growth process, addressing the limitations of existing models in dynamic feedback and real-time adjustment capabilities. Future extensive validation and optimization of the model across different crop types and environmental conditions will further enhance its potential for application in plant factory management. Keywords: Back propagation neural network, Digital twins technology, Lettuce, Plant factory, State prediction.

To efficiently detect and count maize seedlings in complex field conditions, this study first developed a sample dataset under diverse backgrounds and lighting scenarios and introduced a data augmentation technique called “M_AUG.” YOLOv5s was selected as the base model, enhanced with the Swin Transformer (Swin TR)to improve feature extraction across various scales and complex environments. The model also incorporated multi-scale attention (EMA)to enhance the representation of small samples and positive/negative samples, along with the Asymptotic Feature Pyramid Network (AFPN)to integrate seedling features at different levels. The results showed that the proposed SEA-YOLOv5 achieved mAP0.5 of 98.6 %, mAP0.5–0.95 of 73.2 %. and F1 of 97.1 %, with the parameters count of 5.55 million and a weight size of 11.7 MB. Compared to YOLOv5, SEA-YOLOv5 improved mAP0.5 by 5.8 %, mAP0.5–0.95 by 9.9 %, and F1 by 5.4 %, while reducing the parameter count by 1.46 million and weight size by 2.7 MB. SEA-YOLOv5 was compared with YOLOv7, YOLOv8s, Faster R-CNN, RetinaNet, YOLOv10s, DNE-YOLO, and YOLOv11s, and the results indicated that SEA-YOLOv5 outperformed the comparison models in overall performance. Upon deploying SEA-YOLOv5 on the Jetson Orin NX and conducting seedling detection and counting trials across eight plots, the model achieved a miss rate of just 0.63 % and a frame rate of 74.6 FPS. Thus, it can be concluded that the SEA-YOLOv5 model developed in this study provides high accuracy, a compact design, and strong portability, making it well-suited for real-time detection and counting applications in the field.