BACKGROUND:Accurate assessment of disease severity is essential for evaluating fungicide performance and breeding disease-resistant crop varieties. Manual scoring of infection on individual leaf discs is labor-intensive and variable, while traditional computer vision methods require manual parameter tuning and lack robustness. Existing deep learning approaches often struggle to simultaneously localize leaf discs and accurately segment disease symptoms, limiting their practical application. RESULTS:We developed SporaScan, an automated pipeline combining YOLO v8n for leaf disc localization, Mobile SAM for background removal, and UNet for sporulation segmentation. It achieved high accuracy (mAP@50 >99%, mIoU@50 >96%), with background removal reducing misclassification (0.21% for sporulation and 2.75% for leaf discs). Severity estimates showed strong agreement with manual annotations (R2 = 0.99). In a blind test, technicians selected SporaScan as superior in 37.2% of cases, manual annotation in 26.2%, and equal performance in 36.6% (P < 0.001). Compared with Mask R-CNN, the sequential design improved accuracy and efficiency. CONCLUSION:These results demonstrate that SporaScan provides an efficient and practical approach for automated assessment of downy mildew severity, supporting applications in disease evaluation, breeding, and fungicide assessment (http://116.10.197.212:9060/segment/#/). © 2026 Society of Chemical Industry.

Underground mining robots are increasingly modeled for planning, operator training, and digital-twin workflows, where reliable actuator-level kinematics is needed to reduce hazardous in situ trials. Unlike typical open-chain industrial manipulators, representative mining machines are often linear-actuator-driven closed-chain mechanisms with planar four-bar linkages, making reusable kinematic modeling and real-time FK/IK solving challenging. We present MineRobot, an actuator-centered framework for modeling and solving the kinematics of this representative mechanism class. MineRobot introduces the Mining Robot Description Format (MRDF), a domain-specific representation that parameterizes mining-robot kinematics with native semantics for actuators and loop closures. It then contracts planar four-bar substructures into generalized joints and extracts, for each actuator, an Independent Topologically Equivalent Path (ITEP) classified into four canonical types. Based on this decomposition, per-type solvers are composed into a sequential forward-kinematics (FK) pipeline, while inverse kinematics (IK) is formulated as a bound-constrained actuator-length optimization solved by a Gauss–Seidel-style update scheme. By converting coupled closed-chain kinematics into small topology-aware solves, MineRobot reduces robot-specific hand derivations and supports efficient repeated FK/IK computation without treating each query as a full coupled constraint-solving problem. Experiments on representative underground mining robots demonstrate real-time FK performance and robust IK convergence within the tested operating ranges, supporting the use of MineRobot as an actuator-centered kinematic layer for planning, training, and digital-twin workflows.

Effective crop management decisions, such as fertilization, irrigation, and crop protection, are closely tied to the crop growth stages. Precise identification of development stages is essential to optimize management practices in line with crop needs. While deep learning has shown promise in identifying growth stages, existing models often face challenges due to limited data availability and reduced accuracy in complex field conditions. To overcome these limitations, this study proposes a semi-supervised image classification method built on an enhanced ResNetRS50 architecture, named CO-ResNetRS50-SSL. This model leverages ResNetRS50 as its backbone, integrating Coordinate Attention (CA) for improved positional feature extraction and Omni-Dimensional Dynamic Convolution (ODConv) to enhance the adaptability of convolutional kernels to varying targets. Additionally, a semi-supervised learning framework is employed to boost generalization while minimizing dependence on labeled data. Ablation experiments show that semi-supervised learning boosted ResNetRS50's accuracy from 88.58 % to 89.36 %. Adding Coordinate Attention further increased accuracy to 89.89 %, while incorporating ODConv in the final CO-ResNetRS50-SSL model achieved 90.38 % accuracy, 90.59 % precision, and 90.19 % F1 score (with 65.38 M parameters). Comparisons reveal that CO-ResNetRS50-SSL outperforms state-of-the-art models (FasterNet-T1, ShuffleNetV2, Swin Transformer, Vision Transformer, ConvNeXt-base) with highly significant differences (p < 0.001) and delivers robust performance across rice growth stages, with an optimal tradeoff at 224 x 224 resolution. CO-ResNetRS50-SSL can accurately detect rice growth stages with limited labeled data, and its improvements in accuracy and generalization are expected to enhance decision-making in precision agriculture, optimizing resource allocation, reducing inputs, and advancing progress in the field of digital agriculture. Future work will focus on improving efficiency in utilizing unlabeled data, ensuring more balanced performance across different growth stages, and enhancing the model's adaptability to other crops and more complex agricultural scenarios.

Unmanned roadheader-based tunneling in complex underground environments represents a significant advancement in mining technology, offering the potential to enhance safety and efficiency, but it poses considerable challenges in terms of precise control, navigation, and real-time decision-making. In response to these challenges, we have developed a comprehensive framework that integrates software, communication protocols, hardware, and newly developed algorithms. The proposed system architecture employs a laser-based total station for high-precision surveying, coupled with a programmable logic controller (PLC) integrated with multiple sensors to control the movement of the roadheader. To resolve key issues such as command issuance, robust positioning, autonomous navigation, real-time visualization of the working face, and trajectory planning for both the roadheader and cutter head, specialized algorithms have been developed. These components are integrated into a client/server software system, which includes a user-friendly interface that ensures efficient and safe tunneling operations. The client, deployed remotely, generates, and delivers operational instructions while providing real-time visualization of the working face, whereas the server, stationed on-site, executes these instructions, and ensures precise control and positioning of the roadheader. Comprehensive experimentation and system deployment have demonstrated the framework’s effectiveness, successfully resolving key issues such as accurate positioning, unmanned navigation, cutting control, and multiple-trajectory planning. This system not only facilitates successful unmanned tunneling but also offers substantial benefits to the mining industry in China, with potential applications in other unmanned tunneling machines and civil construction projects.

Deep learning (DL) models have shown exceptional accuracy in plant disease identification, yet their practical utility for farmers remains limited due to a lack of professional and actionable guidance. To bridge this gap, we developed CDIP-ChatGLM3, an innovative framework that synergizes a state-of-the-art DL-based computer vision model with a fine-tuned large language model (LLM), designed specifically for Crop Disease Identification and Prescription (CDIP). EfficientNet-B2, evaluated among 10 DL models across 48 diseases and 13 crops, achieved top performance with 97.97 % +/- 0.16 % accuracy at a 95 % confidence level. Building on this, we fine-tuned the widely used ChatGLM3-6B LLM using Low-Rank Adaptation (LoRA) and Freeze-tuning, optimizing its ability to deliver precise disease management prescriptions. We compared two training strategies-multi-task learning (MTL) and Dual-stage Mixed Fine-Tuning (DMT)-using a different combination of domain-specific and general datasets. Freeze-tuning with DMT led to substantial performance gains, achieving a 33.16 % improvement in BLEU-4 and a 27.04 % increase in the Average ROUGE F-score, surpassing the original model and state-of-the-art competitors such as Qwen-max, Llama-3.1-405B-Instruct, and GPT-4o. The dual-model architecture of CDIPChatGLM3 leverages the complementary strengths of computer vision for image-based disease detection and LLMs for contextualized, domain-specific text generation, offering unmatched specialization, interpretability, and scalability. Unlike resource-intensive multimodal models that blend modalities, our dual-model approach maintains efficiency while achieving superior performance in both disease identification and actionable prescription generation.

China holds the largest apple cultivation area globally, yet yields per hectare remain relatively low. Despite substantial government investment in modern orchard technologies, adoption remains limited among farmers. This study investigates the economic and sociological drivers of technology uptake, focusing on how information sources shape adoption behavior. Based on 382 farmer surveys across major apple-producing provinces, the study examines (1) farmers’ preferences for agricultural information sources, (2) the influence of demographic characteristics on those preferences, and (3) the differential effects of specific sources on the adoption of key technologies, including dwarf rootstocks and virus-free seedlings. Results show that agri-chemical dealers (ACDs) and farmer peers (FPs) are the most commonly used information channels. Access to advice from local experts (EXPs) significantly increases the likelihood of adopting dwarf rootstocks, while information from ACDs promotes the use of virus-free seedlings. In contrast, reliance on personal farming experience is negatively associated with technology uptake. These findings highlight the need to strengthen formal information dissemination systems and better integrate trusted local actors like ACDs and EXPs into agricultural extension. Targeted information delivery can improve adoption efficiency, promote evidence-based decision-making, and support the modernization and sustainability of China’s apple sector.

BACKGROUND: Crop diseases can lead to significant yield losses and food shortages if not promptly identified and managed by farmers. With the advancements in convolutional neural networks (CNN) and the widespread availability of smartphones, automated and accurate identification of crop diseases has become feasible. However, although previous studies have achieved high accuracy (>95%) under laboratory conditions (Lab) using mixed data sets of multiple crops, these models often falter when deployed under field conditions (Field). In this study, we aimed to evaluate disease identification accuracy under Lab, Field, and Mixed (Lab and Field) conditions using an assembled data set encompassing 14 diseases of apple (Malus x domestica Borkh.), potato (Solanum tuberosum L.), and tomato (Solanum lycopersicum L.). In addition, we investigated the impact of model architectures, parameter sizes, and crop-specific models (CSMs) on accuracy, using DenseNets, ResNets, MobileNetV3, EfficientNet, and VGG Nets. RESULT:Our results revealed a decrease in accuracy across all models from Lab (98.22%) to Mixed (91.76%) to Field (71.55%) conditions. Interestingly, disease classification accuracy showed minimal variation across model architectures and parameter sizes: Lab (97.61-98.76%), Mixed (90.76-92.31%), and Field (68.56-73.81%). Although CSMs were found to reduce inter-crop disease misclassifications, they also led to a slight increase in intra-crop misclassifications. CONCLUSION: Our findings underscore the importance of enriching data representation and volumes over employing new model architectures. Furthermore, the need for more field-specific images was highlighted. Ultimately, these insights contribute to the advancement of crop disease identification applications, facilitating their practical implementation in farmer's fields. (c) 2024 Society of Chemical Industry.

Monitoring spores is crucial for predicting and preventing fungal- or oomycete-induced diseases like grapevine downy mildew. However, manual spore or sporangium detection using microscopes is time-consuming and labor-intensive, often resulting in low accuracy and slow processing speed. Emerging deep learning models like YOLOv8 aim to rapidly detect objects accurately but struggle with efficiency and accuracy when identifying various sporangia formations amidst complex backgrounds. To address these challenges, we developed an enhanced YOLOv8s, namely, AFM-YOLOv8s, by introducing an Adaptive Cross Fusion module, a lightweight feature extraction module FasterCSP (Faster Cross-Stage Partial Module), and a novel loss function MPDIoU (Minimum Point Distance Intersection over Union). AFM-YOLOv8s replaces the C2f module with FasterCSP, a more efficient feature extraction module, to reduce model parameter size and overall depth. In addition, we developed and integrated an Adaptive Cross Fusion Feature Pyramid Network to enhance the fusion of multiscale features within the YOLOv8 architecture. Last, we utilized the MPDIoU loss function to improve AFM-YOLOv8s’ ability to locate bounding boxes and learn object spatial localization. Experimental results demonstrated AFM-YOLOv8s’ effectiveness, achieving 91.3% accuracy (mean average precision at 50% IoU) on our custom grapevine downy mildew sporangium dataset—a notable improvement of 2.7% over the original YOLOv8 algorithm. FasterCSP reduced model complexity and size, enhanced deployment versatility, and improved real-time detection, chosen over C2f for easier integration despite minor accuracy trade-off. Currently, the AFM-YOLOv8s model is running as a backend algorithm in an open web application, providing valuable technical support for downy mildew prevention and control efforts and fungicide resistance studies.

Intelligent recognition and location of crosswalks is a crucial element of autonomous driving. Achieving accurate real-time crosswalk detection with limited computational resources and multiple interference scenes has been troubling researchers. To address these issues, a lightweight convolutional neural network based on YOLOv5s, namely YOLO-LCD, is proposed. First, a lightweight network incorporating a coordinate attention(CA) mechanism is proposed as a replacement for the backbone network of YOLOv5s, which leads to a significant reduction in the parameter count of YOLOv5s by using depthwise convolution operations. Secondly, for better extraction of crosswalk features, Reparameterized Generalized-FPN(RepGFPN) is adopted as the neck of YOLO-LCD. Finally, for better localization of crosswalks, we use a dynamic label assignment strategy to better define positive anchors and use the EIoU loss rather than the CIoU loss. Additionally, the detection algorithm that we propose is verified on the crosswalk dataset containing complex scenarios. The experimental results demonstrate that the YOLO-LCD algorithm obtains a surprisingly higher F1 score of 97.8% with 80.2% fewer parameters and 82.9% fewer FLOPs than YOLOv5s, and the FPS on an old CPU device is 186.4% faster. This study provides a solution for accurate real-time crosswalk detection under limited computing resources.

The extraction of entities and relationships from unstructured text is not only a critical issue in information extraction, but also an essential component of constructing knowledge graphs. Most existing models for relation extraction first extract all subjects, and then proceed to extract objects and relations based on the identified subjects. This method is highly dependent on the extraction of subjects, and the quality of subject extraction significantly impacts the extraction of objects and relations. During relation processing, most models only use the simplest classifier, with the difference lying in the input for relation processing, which sometimes leads to the mutual relationship between the subject and the object being overlooked.To avoid overlooking the inherent features of input sentences, we generate three distinct token representation sequences for subject, object, and relation at the encoding layer. Additionally, combining explicit injection of context features from the encoding layer, entity extraction in both directions can mutually promote each other, aiming to obtain as many entity pairs as possible. This approach aims to address the limitations imposed by the relationship being subject to the extracted subject. While most existing models use a simple classifier for handling relationships, our model employs the Biaffine model combined with an attention mechanism to assign all potential relationships for each entity pair. Its advantage over other simple classifiers is maintaining a matrix for each relation, accurately modeling the features of relations, and its probability calculation mechanism can precisely explore the interaction between subjects and objects.We evaluated the proposed model on the NYT and WebNLG datasets. Extensive experimental results demonstrate the high effectiveness of the proposed model, achieving significant performance across all datasets.

Aiming at the problem that the boom-type roadheader cannot quickly adjust the cutting swing speed to adapt to the hardness of the coal and rock when the coal hardness changes in coal mines, a control strategy for the driving swing speed is proposed. In this strategy, firstly, the PSO-BP neural network is used to construct a cutting load recognizer to provide a basis for adjusting the cutting swing speed of the roadheader; secondly, the PID control is optimized based on the fuzzy algorithm, and the fuzzy PID controller is established to improve the regulation of the cutting The efficiency of the swing speed; Finally, the roadheader swing speed simulation control system model is built in Matlab/Simulink, and the proposed roadheader cutting swing speed control strategy is simulated. The simulation experiment results show that the roadheader swing speed adjustment system using PSO neural network algorithm combined with fuzzy PID control has significantly improved response speed and control accuracy, and has good superiority and stability. The strategy based on particle swarm BP neural network algorithm combined with fuzzy PID control can provide certain theoretical guidance for stabilizing the cutting motor power of the roadheader and improving the efficiency of roadway work.

Intersections are one of the important factors affecting the overall traffic efficiency of road sections. To alleviate the pressure on road traffic and improve the traffic efficiency of intersections, this paper proposes an optimization model of signal light timing from three aspects of road, vehicle and environment, selects road capacity, average delay and vehicle carbon emissions as optimization goals. In addition, this paper proposes an improved particle swarm optimization to solve the model. Add adaptive weights to the particle swarm optimization, update the particle position by levy flight to improve the ability of the algorithm. The experimental results show that the timing scheme obtained by using the improved particle swarm optimization algorithm is superior to other methods in terms of convergence speed and convergence accuracy, which proves the feasibility and superiority of the algorithm.

Meteorological data is a kind of time series data with obvious seasonal trends. Prediction of large-scale meteorological data can understand the stage weather change condition of the region, and accurate prediction of meteorological data is important to guide the life of residents and activities of foreign travelers. In order to accurately predict the weather data of Taishan region, the weather data prediction model of Prophet+LSTM is proposed. In order to construct this prediction model, firstly, the time series data of temperature in Taian City from January 2000 to December 2019 are modeled, and the prediction effect of the model on temperature is evaluated by using evaluation indexes, and secondly, the most suitable combination coefficients of the two models are found by linear weighted combination method to obtain the Prophet+LSTM combination model. Finally, by validating different meteorological data from two different regions, the experimental results show that the combined model has higher prediction accuracy than the traditional machine learning methods as well as its single model, and has better application prospects.

Image feature point and descriptor extraction is the basis of SLAM, SFM and 3D reconstruction tasks. In this paper, we study the SuperPoint network, which has good robustness in extracting feature points and descriptors, and introduces the idea of group convolution, replaces the normal convolution with group convolution, and introduces the Mish activation function to replace the ReLU activation function to solve the problem that some data fall into negative intervals. The experimental results show that the accuracy of single-strain estimation only decreases by 0.01 when the tolerance distance difference is 3, and the repetition rate of feature point detection increases by 0.3%, which has good robustness. In this paper, the network achieves lightness without excessive loss of accuracy.

As the main equipment of the coal mining industry, the boom-type roadheader's tunneling efficiency is mainly affected by the power of the cutting motor. In the traditional coal mining process, sudden changes in the hardness of the coal and rock will cause the power of the cutting motor to oscillate, thereby reducing the operating efficiency of the roadheader. In order to solve the problem of the power oscillation of the cutting motor when the coal hardness changes suddenly, a roadheader swing speed adjustment system is designed. The system mainly stabilizes the power of the cutting motor by adjusting the swing speed of the cantilever. The system first introduced the sliding window algorithm to solve the problem of BP neural network forgetting historical samples, and then used the BP neural network algorithm to optimize PID control, construct a BP-PID controller, and finally use the cutting current as the basis for sudden changes in coal hardness. The control performance of BP-PID control based on sliding window is simulated and analyzed. The results show that the BP neural network PID control system based on sliding window can quickly adjust the swing speed, stabilize the cutting motor power, reduce the oscillation amplitude, have better adaptability, and has more superior control performance than a single PID control system.

河流水位数据是洪涝灾害仿真模拟的重要依据,而精准的水位预测可以给洪水的淹没范围提供可靠的参考信息.单一BP神经网络的水位预测模型通常用于洪水水文模拟,但其准确度不高.为了精确分析洪水淹没范围并实现洪水淹没的仿真模拟,首先引入主成分分析方法(PCA)提取出影响水位变化的主成分,然后将主成分作为GA-BP神经网络的输入变量,河流水位数据为输出变量,建立PG-BP神经网络洪水水位预测模型.以大沽河流域为研究区域,使用该模型对汛期水位进行预测,根据模型预测的水位数据,可实现大沽河流域洪水淹没的仿真模拟.该模型水位预测的预报准确率均值达99.8％,预报效果较好,拟合精度较高,且可视化仿真也能够真实生动地显示出受灾地区,可以为防洪决策提供有力支撑.

Attributes in datasets are usually not equally significant. Some attributes are unnecessary or redundant. Attribute reduction is an important research issue of rough set theory, which can find minimum subsets of attributes with the same classification effect as the whole dataset by removing unnecessary or redundant attributes. We use Chi-square statistics to evaluate the significance of condition attributes. It can reduce the search space of attribute reduction and improve the speed of attribute reduction. Conditional entropy of relative attributes is adopted as a heuristic function. Two decision table reduction algorithms, forward selection and backward deletion, are proposed to approach the optimal solution. Based on this, an efficient incremental attribute reduction method for dynamically changing datasets is proposed by preserving intermediate variables. The intermediate variable is the observation frequency matrix of joint events of each condition attribute and decision attribute. Experimental results show that the proposed algorithms can improve performance in terms of processing time.

We propose a fundamental theorem for eco-environmental surface modelling (FTEEM) in order to apply it into the fields of ecology and environmental science more easily after the fundamental theorem for Earth’s surface system modeling (FTESM). The Beijing-Tianjin-Hebei (BTH) region is taken as a case area to conduct empirical studies of algorithms for spatial upscaling, spatial downscaling, spatial interpolation, data fusion and model-data assimilation, which are based on high accuracy surface modelling (HASM), corresponding with corollaries of FTEEM. The case studies demonstrate how eco-environmental surface modelling is substantially improved when both extrinsic and intrinsic information are used along with an appropriate method of HASM. Compared with classic algorithms, the HASM-based algorithm for spatial upscaling reduced the root-mean-square error of the BTH elevation surface by 9 m. The HASM-based algorithm for spatial downscaling reduced the relative error of future scenarios of annual mean temperature by 16%. The HASM-based algorithm for spatial interpolation reduced the relative error of change trend of annual mean precipitation by 0.2%. The HASM-based algorithm for data fusion reduced the relative error of change trend of annual mean temperature by 70%. The HASM-based algorithm for model-data assimilation reduced the relative error of carbon stocks by 40%. We propose five theoretical challenges and three application problems of HASM that need to be addressed to improve FTEEM.

With the development of remote sensing and large-scale environmental modelling, large amount of environmental data are continuously becoming available. An intuitive and comprehensive visualization of these data could facilitate data exploration, communication and collaboration between the stakeholders for informed decisions making. In Poyang lake basin regions, we demonstrate how to develop a software platform that can visualize three-dimensionally environmental data layers including terrain, weather, river net, water level, land use changes and interrelations between these data layers for environmental system monitoring and decision supporting. The tool is built by combining several prevailing projects including Unity, Paraview, and hydrological models, etc. We develop an open standardized framework for the software tool to host environmental data layers permitting the application of the tool, which could be employed for the intuitive and comprehensive data visualization in other regions facing environmental challenges.

Airborne light detection and ranging (lidar) is becoming a widely adopted technique for capturing elevation data, which are mainly used for creating digital terrain models (DTMs). However, the large size of lidar datasets poses a serious computational challenge to the promising radial basis function (RBF) interpolation method. In this work, to reduce the huge computational cost and improve the interpolation accuracy, random Fourier features are first introduced to approximate the Gaussian kernel of RBFs in feature space, then a random features-based weighted RBF interpolation method is developed. Based on randomized Fourier features, the nonlinear kernel-based training and evaluation of the RBF method is transformed into simple linear operations in feature space, and with the help of weighted ridge regression, the negative effect of the non-Gaussian distribution of lidar datasets on DTM production is reduced. In other words, the combination of randomized Fourier features and weighted ridge regression improves the efficiency and accuracy of the RBF interpolation method. Experiments on simulated datasets indicate that the proposed method performs better than the classical or random features-based RBF method for dealing with non-Gaussian distributed samples, with the former being slightly less accurate than the iterative RBF method due to the low-dimensional random features. However, the computational cost of the new method is much lower compared with the classical or iterative RBFs. Interpolation of airborne lidar-derived points demonstrates that the new method has a computational cost similar to the inverse distance weighting and triangulated irregular network (TIN) approaches, and is significantly faster than the ordinary kriging (OK) or thin plate spline (TPS) methods. Quantitatively, for interpolation of 644,433 points, the proposed method is approximately 833 and 21 times faster than OK and TPS, respectively. Moreover, the new method avoids the surface discontinuity artifacts presented by the OK, TPS, and TIN methods.