[Objective] Soil organic matter (SOM) is a fundamental indicator for evaluating soil fertility and soil quality. In mountainous counties characterized by complex terrain and pronounced environmental heterogeneity, SOM exhibits strong spatial variability even over short distances, which often results in limited prediction accuracy for conventional digital soil mapping (DSM) models. With the nationwide implementation of the Third National Soil Census, the demand for high-resolution and high-accuracy SOM mapping at the county scale has become increasingly urgent. Against this backdrop, Yiyuan county in Shandong province was selected as the study area to assess the applicability of the geographically weighted random forest (GWRF) model in SOM mapping within complex terrain regions. Furthermore, it sought to systematically compare the predictive performance of GWRF with several commonly used models, thereby providing technical support for soil resource surveys, census result compilation, and county-level land management. [Methods] The dataset consisting of 1 565 measured topsoil SOM samples was utilized, along with nineteen environmental variables representing five categories: topography, climate, vegetation, soil properties, and land use. Through correlation analysis and collinearity diagnostics, twelve key variables were retained for model construction. The GWRF model, which integrates localized spatial modeling with nonlinear machine-learning capability, was developed to generate high-resolution SOM predictions across the study area. An adaptive bandwidth strategy was employed, and the optimal bandwidth of 500 was determined. Grid search combined with cross-validation was used to identify the optimal mtry value of 4 for the random forest component. In addition to GWRF, four reference models were constructed for comparison: ordinary kriging (OK), multiple linear regression (MLR), geographically weighted regression (GWR), and random forest (RF). Model performance was evaluated using two commonly adopted accuracy metrics: the coefficient of determination (R²) and root-mean-square error (RMSE). [Results and Discussions] Overall, SOM levels in Yiyuan county were relatively low, with a mean value of 15.62 g/kg. The spatial variation was moderate and exhibited a clear pattern: SOM values were higher in the central area and lower in the northeastern and southwestern areas. Considerable differences were observed in prediction accuracy among the five models. The GWRF model achieved the best overall performance, with an R² of 0.48 and an RMSE of 5.12 g/kg. This accuracy clearly surpassed that of RF (R²=0.41) and GWR (R²=0.35), and its advantage over MLR and OK was even more pronounced. A paired-sample t-test further confirmed that the accuracy improvements of GWRF over the other four models were statistically significant, supporting the robustness and reliability of the model's enhanced performance. According to the mapping results, the OK model produced an excessively smooth surface, making it difficult to reveal local details. While the MLR and GWR models could characterize certain environmental effects, they exhibited significant biases such as underestimation of high values and overestimation of low values. In contrast, the GWRF model performed prominently in capturing both global trends and local subtle variations. The analysis of variable importance showed that soil type, annual evapotranspiration, slope, and sand content were the most influential factors governing SOM distribution in the study area. Moreover, their spatially varying importance revealed notable heterogeneity. [Conclusions] This study demonstrated that the GWRF model possesses significant advantages in county-scale SOM digital mapping within mountainous areas. Its prediction accuracy markedly exceeded that of RF and conventional linear models, owing to its ability to simultaneously capture nonlinear environmental relationships and localized spatial variations. The enhanced mapping precision and improved representation of spatial details highlight the strong potential of GWRF for applications requiring high-accuracy soil information. GWRF is well-suited for SOM prediction under complex terrain conditions and can serve as an effective technical tool for county-level soil property estimation. Future research may incorporate human-activity-related variables, employ localized variable-selection strategies within the GWRF framework to further refine model performance, and explore the application potential of more advanced deep learning models in soil property mapping.

[Objective] In the eel (Monopterus albus) farming system used in feed distribution research of mechanical arm, the challenges included slow path planning speeds, excessive trajectory redundancy, and suboptimal obstacle avoidance success rates within confined operational spaces. To mitigate these issues, an improved path planning algorithm, based on the bidirectional rapidly-exploring random tree star (BI-RRT*) algorithm was proposed. The primary aim was to significantly enhance the motion efficiency and task success rate of robotic arms operating in complex, constrained environments. [Methods] The proposed improved BI-RRT* algorithm integrated an adaptive goal-biased strategy with an enhanced artificial potential field (APF) method. The algorithm's framework comprises three core components: a high-quality sampling strategy, an efficient search strategy, and a path optimization algorithm. For the high-quality sampling strategy, an adaptive goal-biased approach was introduced to overcome the limitations of inefficient random sampling and slow convergence rates characteristic of traditional BI-RRT algorithms in complex environments. This strategy dynamically adjusted the generation of sampling points, moving beyond purely random selection. Instead, it prioritized sampling regions in the vicinity of the target, guided by the target direction and a predefined bias probability. This mechanism substantially augmented the growth propensity of the search tree towards the target area, effectively reducing the stochasticity of random sampling and consequently accelerating the path search process. To enhance search efficiency and prevent the algorithm from converging to local optima, an improved APF was incorporated into the node expansion process. The APF was refined to achieve superior integration with the BI-RRT framework. During each new node expansion, in addition to considering the inherent random exploration characteristics of BI-RRT, a directional attractive field was superimposed. This attractive field not only originated from the ultimate target point but also factored in the current growth orientation of the search tree and localized environmental information. Specifically, a composite attractive function was devised, which synergized the attractive force exerted by the target point on the current node with the attraction from potential "guide points". Concurrently, the computation of the repulsive field was optimized to more precisely delineate the geometry and proximity of obstacles, thereby circumventing common issues such as "oscillation" and "deadlock" prevalent in traditional APF. Through this methodology, the algorithm was able to more effectively steer the search tree to circumvent obstacles and rapidly converge towards the target region, significantly bolstering the directness of the search and successfully preventing the algorithm from becoming ensnared in suboptimal local solutions. For the path optimization algorithm, following the generation of an initial feasible path, a greedy optimization strategy was employed for path pruning and smoothing. This was executed to yield an optimal path characterized by reduced length, enhanced smoothness, and improved conformity with the kinematic properties of the robotic arm. Path pruning was initially applied to eliminate redundant nodes; if a collision-free direct connection existed between two non-adjacent nodes, intermediate nodes were excised, thereby substantially abbreviating the path length. Subsequently, path smoothing techniques, such as B-spline curves or cubic spline interpolation, were introduced to enable the robotic arm to execute movements with greater stability and efficiency during actual operation, mitigating impact and vibration. This two-stage optimization procedure ensured that the final generated path was not merely feasible but also optimal across metrics of length, smoothness, and motion efficiency. [Results and Discussions] To comprehensively validate the performance of the proposed algorithm, a two-stage experimental verification was conducted. Initially, comparative simulations were performed in both two-dimensional (2D) and three-dimensional (3D) environments utilizing the Matlab platform. These simulation scenarios were meticulously engineered to encompass three archetypal environments—simple, complex, and narrow passages—thereby emulating the diverse obstacle configurations potentially encountered in industrialized eel aquaculture. The results demonstrated that, concerning both path planning speed and quality, the improved BI-RRT* algorithm significantly surpassed RRT, APF-RRT*, and traditional BI-RRT* algorithms across all tested environments, substantiatingthe theoretical superiority and inherent robustness of the improved BI-RRT* algorithm proposed in this study across varying complex environments. To further ascertain the engineering applicability and practical potential of the algorithm, an eel feeding robotic arm simulation system was meticulously constructed based on the robot operating system and MoveIt frameworks. This system precisely emulated the kinematics, dynamics, and obstacle distribution pertinent to an industrialized eel aquaculture environment. During simulated continuous feeding tasks, the improved BI-RRT* algorithm consistently exhibited impressive and outstanding performance. Its average running time was merely 2.1 s, representing a substantial 41.6% reduction compared to the traditional BI-RRT*. The average length of the planned path was recorded at only 1 680 mm, with an average of 180 nodes, indicating a significant reduction in path redundancy. Furthermore, the algorithm achieved an impressive obstacle avoidance success rate of 96% in complex confined spaces. These empirical findings not only validated the algorithm's effectiveness but also underscored its immense potential for practical engineering applications. [Conclusions] The experimental results conclusively demonstrated that the improved BI-RRT* algorithm significantly enhanced the path planning efficiency and trajectory quality of robotic arms operating within confined spaces. It also exhibited exceptionally high reliability in obstacle avoidance, thereby effectively addressing the automated feeding requirements of industrialized eel aquaculture. The algorithmic framework possessed considerable generality, offering valuable theoretical insights and technical precedents for resolving analogous robotic arm path planning challenges in other agricultural automation contexts.

[Objective] Accurate and timely crop mapping is fundamental for agricultural management, yield forecasting, and food security assessment. However, in mountainous and hilly regions characterized by frequent cloud cover and highly fragmented farmland, crop classification methods relying solely on optical remote sensing data are severely constrained. Persistent cloud contamination introduces data gaps and temporal inconsistencies in optical image time series, significantly degrading classification accuracy and robustness. To address these limitations, a robust and adaptive deep learning framework is developed capable of effectively integrating multi-modal remote sensing data. The primary objective is to enhance crop classification accuracy and stability under complex conditions where optical observations are scarce or unreliable, thereby supporting reliable agricultural monitoring in cloudy and fragmented landscapes. [Methods] A novel deep neural network architecture named 3D convolutional neural network based on attention mechanism (Attention-3DCNN) was proposed, designed to jointly exploit multi-temporal optical and synthetic aperture radar (SAR) observations. The model integrated Sentinel-2 multispectral time-series imagery with weather-insensitive Sentinel-1 SAR data through a dedicated cross-modal fusion strategy driven by a triple-attention mechanism. The network adopted a dual-branch feature extraction architecture. For the Sentinel-2 data, a hybrid module combining three-dimensional and two-dimensional convolutional neural networks (3D-CNN and 2D-CNN) was employed to capture discriminative spatiotemporal features and crop phenological dynamics across the growing season. This design enabled effective modeling of the spectral-temporal interactions inherent in crop development. For the Sentinel-1 SAR data, depthwise separable convolutions were utilized to efficiently extract spatial and textural features related to crop structure and surface scattering characteristics while reducing computational complexity. Features extracted from both modalities were subsequently integrated using a custom-designed attention-based fusion module. This module consisted of three complementary attention mechanisms: channel attention, temporal attention, and spatial attention. Residual connections were incorporated throughout the network to facilitate stable training and effective gradient propagation. The proposed model was evaluated on two datasets to assess both its performance and generalizability. The first was the publicly available panoptic agricultural satellite time series (PASTIS) benchmark dataset from France, which contained dense time-series observations and multiple crop classes. The second was a real-world dataset constructed for Yishui county, Shandong province, China, which was characterized by high cloud frequency (approximately 33%), highly fragmented farmland (average parcel size < 0.5 hm²), and a relatively simple crop rotation system. Comparative experiments were conducted against several state-of-the-art models, including 3D-ConvSTAR, UNet++, Self-Attention 3D, CNN-LSTM dual-stream network, and TGF-Net. Ablation studies were also performed to quantify the contribution of each attention component. [Results and Discussions] Experimental results demonstrated that Attention-3DCNN consistently outperformed all baseline methods on both datasets. On the PASTIS benchmark, the model achieved an overall accuracy (OA) of 97.5%, confirming its strong classification capability under favorable observation conditions. On the more challenging Yishui county dataset, Attention-3DCNN attained an OA of 93%, outperforming the other comparison models. Ablation experiments confirmed the effectiveness of the proposed triple-attention mechanism, as removing any attention component resulted in a clear reduction in classification performance. Under heavy cloud coverage, Attention-3DCNN exhibited the smallest accuracy degradation, with an OA drop of only 3.6 percentage points, indicating its ability to adaptively rely on SAR information when optical data quality deteriorated. In regions with highly fragmented farmland, the proposed model also maintained the highest accuracy and the smallest performance decline (2.8 percentage points), benefiting from the spatial attention mechanism. Moreover, attention visualization provided meaningful interpretability. Temporal attention peaks aligned with key crop phenological stages, while channel attention highlighted spectrally and physically informative optical bands and SAR polarizations, which was consistent with established agronomic and remote sensing knowledge. [Conclusions] This study presents the Attention-3DCNN model for accurate and robust crop classification in regions affected by persistent cloud cover and fragmented agricultural landscapes. By fusing Sentinel-2 optical and Sentinel-1 SAR time-series data through a channel-temporal-spatial triple-attention mechanism, the proposed framework enables adaptive integration of complementary multi-modal information. The model achieves outstanding performance on both benchmark and real-world datasets, demonstrates strong robustness under adverse conditions, and offers enhanced interpretability. Overall, the proposed approach provides a reliable and practical solution for crop mapping in complex agricultural environments.

[Objective] Pears are a common fruit rich in vitamins and minerals. Traditional pear grading primarily relies on manual inspection, which is not only laborious but also susceptible to subjective factors, leading to unstable and inaccurate results. Furthermore, manual operations may cause varying degrees of physical damage to pears, affecting their appearance and market value. Therefore, developing an automated, efficient, and reliable pear grading technology has become an urgent demand in the industry. To address the current problem of poor detection accuracy caused by the small scale of surface defects in Dangshan pears, a lightweight high-precision model was proposed based on an improved Mamba-YOLO architecture, aiming to balance detection accuracy and efficiency. [Methods] The dataset comprised 1 000 images, which were partitioned into training, validation, and test sets in an 8:1:1 ratio. The following improvements were made to the network architecture. Firstly, a dynamic upsampling (Dysample) module was adopted. Compared to the existing upsampling module in Mamba-YOLO, the Dysample module featured fewer parameters and floating-point operations (FLOPs). Its design eliminated complex dynamic convolution kernels, requiring only a small number of linear layers and grouping operations, thereby preserving computational efficiency while enhancing the retention of defect details. Secondly, regarding pear surface defect detection, defects often exhibited high-frequency local features, whereas traditional convolutional neural networks (CNNs) suffer from insufficient feature capture and imbalanced frequency response. As the dilation rate increased, the frequency response of the convolution kernel decreased and its bandwidth narrowed, consequently limiting its ability to process high-frequency information. Therefore, a frequency-adaptive dilated convolution (FADC) module was proposed, which dynamically adjusted the convolution kernel size, enabling the network to adaptively select matching kernels based on local input features. Smaller kernels were used in high-frequency regions, and larger kernels in low-frequency regions, thereby achieving collaborative optimization of multi-band features and enhancing the ability to extract defect features. Finally, considering that using only single-scale depthwise convolutions to capture local features might lead to insufficient perception of input feature information, and that traditional gating mechanisms may lack adequate global context information modeling, the squeeze-and-excitation module was fused with a channel mixer based on the convolutional gated linear unit (CGLU). This combination was extended into a multi-scale version termed MS-CGLU. By incorporating convolutional kernels of different sizes to extract multi-scale features, followed by weighted fusion, stronger feature representation was achieved. [Results and Discussions] The proposed method was rigorously evaluated on the dangshan pear test set. Ablation experiments demonstrated that introducing the CGLU, FADC, and Dysample enhanced detection performance, confirming the effectiveness of these modules. Compared to YOLOv8n, Gold-YOLO-N, and YOLOv12n, the mean average precision (mAP) was higher by 4.7, 5.3, and 6.3 percentage points, respectively. Compared to the baseline Mamba-YOLO-T, the mAP increased by 3.4 percentage points and the frames per second improved by 10.8 percentage points. Furthermore, in comparative experiments with larger-scale models from the same Mamba-YOLO series, the proposed algorithm still demonstrated significant advantages, i.e., its parameter count was only 41.7% of Mamba-YOLO-B and 15.7% of Mamba-YOLO-L, and its FLOPs was merely 57.1% and 18.1% of the respective models, yet it achieved increases in mAP@0.5 of 3.2% and 1.4%, and increases in mAP@0.5:0.95 of 3.1% and 2.6%, respectively. [Conclusions] This research developed a high-precision and lightweight algorithm for detecting surface defects on Dangshan pears. It achieved a superior balance between detection accuracy and inference speed, significantly outperforming relevant lightweight benchmarks and even larger models within its own family in terms of efficiency. This work can provide reliable algorithmic support for lightweight detection research of pear surface defects.

[Objective] As a major crop in protected horticulture, cluster tomatoes grow in clusters with dense overlapping fruits. In greenhouse environments, light conditions are complex and variable, and the fruit color transitions continuously from green to red across different ripening stages, showing continuous gradation characteristics. These factors result in the low efficiency and strong subjectivity of traditional manual recognition methods. Meanwhile, deep learning-based detection models often suffer from decreased detection accuracy, large localization errors, and slow inference speed when facing complex backgrounds and color interference, making it difficult to meet the dual requirements of real-time performance and high precision in practical applications. Therefore, to meet the practical application requirements of high accuracy, high real-time performance, and strong robustness for cluster tomato ripeness detection, this paper proposes a lightweight target detection model for cluster tomato ripeness, namely LampCT-YOLO (Cluster Tomato YOLO with LAMP pruning), which is based on improved YOLOv10. Through structural optimization and lightweight transformation of the baseline model, the detection accuracy, inference speed, and robustness are effectively improved, providing a novel technical solution for cluster tomato ripeness detection. [Methods] Taking YOLOv10 as the baseline model, first, the issue of insufficient feature extraction capability in complex scenarios was addressed by introducing the SegNeXt attention mechanism into the backbone network. By adaptively adjusting attention weights and calculating the correlation matrix between different feature channels, the mechanism automatically identified color channels strongly associated with the three ripeness levels of cluster tomatoes and assigned them higher attention weights, while suppressing feature responses from irrelevant background channels such as greenhouse frames, soil, and irrigation pipes. To achieve lightweight deployment of the model and meet the real-time detection requirements of edge devices, a gradient-based global channel importance method—LAMP channel pruning technology—was introduced after model training. The core principle of this technology was to evaluate the contribution of each channel to the model's detection performance by calculating the gradient magnitude of channels in each network layer, thereby eliminating redundant channels. This significantly reduced the model size and computational complexity while effectively maintaining the model's high detection performance for the three-category ripeness classification of cluster tomatoes. [Results and Discussions] Experiments showed that under the environment of NVIDIA A100 graphics card, for 240 cluster tomato images in the test set, the LampCT-YOLO model exhibited excellent detection performance. The mean average precision at 50 intersection over union (mAP50) for the early ripe, mid-ripe, and late ripe stages of cluster tomatoes was 84.6%, 89.5%, and 88.4%, respectively, which represented increases of 5.5, 7.7, and 0.9 percentage points compared with YOLOv10. The average mAP50 for the three ripeness categories of cluster tomatoes reached 87.6%, a 4.7 percentage points improvement over YOLOv10, demonstrating outstanding performance in both detection accuracy and stability. In addition, the model was found to maintain high recognition accuracy when facing variations in light intensity, fruit occlusion ratio, and background complexity, indicating good robustness and environmental adaptability. Regarding the lightweight effect, after applying the LAMP channel pruning technology, the number of model parameters and computational complexity were reduced by 63.07% and 50.06%, respectively, while the inference speed was improved by 23.1%. This effectively met the requirements of edge computing devices for real-time detection and low power consumption, alleviating the trade-off between model accuracy and inference speed. To verify the practical application value of the LampCT-YOLO model, the model was deployed on a self-developed fruit and vegetable inspection robot, which conducted field tests on 456 clusters of tomatoes in a real greenhouse environment. The results showed that the inspection robot successfully identified 78, 61, and 248 clusters of early ripe, mid-ripe, and late ripe cluster tomatoes, respectively, with detection accuracies of 84.8%, 87.1%, and 84.4%, and an average accuracy of 85.4%. Meanwhile, there were 5, 7, and 10 false detections, as well as 9, 2, and 36 missed detections for the early ripe, mid-ripe, and late ripe stages respectively, which to a certain extent reflected the practical application potential of the model. [Conclusions] The optimized LampCT-YOLO model not only significantly improves the recognition accuracy of cluster tomatoes at different ripening stages but also greatly reduces the model complexity, successfully achieving efficient deployment in resource-constrained scenarios. This model effectively balances the dual requirements of detection accuracy and real-time performance for inspection robots, and further constructs a reusable technical framework for the ripeness detection of protected horticultural fruits and vegetables. It provides strong support for the transformation of protected agriculture from labor-intensive to technology-intensive, and injects key innovative impetus into the large-scale and diversified implementation of smart agriculture.

[Objective] Maize is one of the most important staple crops in the world and serves as a cornerstone of food security and agricultural sustainability. Accurate and timely prediction of maize yield is essential for optimizing agricultural management practices, supporting market regulation, and guiding policy decisions related to food supply and climate adaptation. In recent years, data-driven yield prediction methods based on machine learning and deep learning have achieved notable improvements in predictive accuracy. However, most existing approaches primarily rely on statistical correlations among variables and often treat influencing factors as independent predictors, without explicitly addressing the complex causal mechanisms and time-lagged interactions that govern crop growth processes. This limitation may lead to reduced model interpretability and compromised robustness under changing environmental conditions. To address these challenges, a novel maize yield prediction framework that integrates causal inference with a hybrid deep learning model was proposed, aiming to improve both predictive performance and mechanistic understanding. [Methods] Multi-source heterogeneous datasets collected across the maize growing season were utilized, including remote sensing-derived vegetation indices, meteorological variables (such as temperature and precipitation), soil profile moisture measurements at multiple depths, and crop observation data corresponding to key phenological stages. First, the Peter-Clark and momentary conditional independence (PCMCI) causal discovery algorithm was applied to systematically identify causal relationships between maize yield and its potential driving factors. The PCMCI method enables the detection of both contemporaneous and time-lagged causal links while effectively controlling for confounding effects in high-dimensional time series data. Through this process, the causal structure of yield formation was explicitly characterized, and key variables with statistically significant causal impacts were selected as inputs for the prediction model. Subsequently, a hybrid moving average, convolutional neural network-long short-term memory (MA-CNN-LSTM) model was constructed to capture the complex spatiotemporal patterns in the causally screened input variables. Specifically, a moving average module was employed as a preprocessing step to suppress high-frequency noise and enhance signal stability. A CNN was then used to extract latent correlation features among multiple variables, reflecting their joint influence on yield formation. Finally, an LSTM network was adopted to model temporal dependencies and cumulative effects across the growing season, enabling effective representation of dynamic yield responses. [Results and Discussions] The causal analysis revealed that soil moisture at depths of 10 cm and 50 cm exerted a significant positive influence on maize yield (P < 0.01), with deeper soil moisture showing a stronger and more persistent time-lagged effect. This finding highlighted the critical role of subsurface water availability in sustaining crop growth during later developmental stages. In addition, vegetation indiced such as the modified chlorophyll absorption ratio index and the normalized difference vegetation index exhibited significant short-term causal relationships with yield during the mid-growth stage of maize, indicating their sensitivity to canopy structure and photosynthetic activity during this period. Comparative experiments conducted against traditional statistical models and conventional machine learning approaches demonstrated that the proposed PCMCI-MA-CNN-LSTM framework consistently achieved superior predictive performance. On the test dataset, the coefficient of determination (R²) reached 0.955, while the mean absolute error (MAE) and root mean square error (RMSE) were reduced to 1.201 kg/mu and 1.474 kg/mu (1 hm²=15 mu). These results indicated that incorporating causal variable selection effectively enhances model accuracy and stability by reducing redundant and spurious correlations. [Conclusions] The results confirm that incorporating causal analysis into yield modeling provides a robust basis for identifying key driving variables and effectively enhances the accuracy and interpretability of maize yield prediction. The proposed framework offers a promising approach for precision agriculture and decision support in crop yield forecasting, particularly under complex and dynamic agro-environmental conditions.

[Objective] In the agricultural domain, user demand texts serve as essential primary sources for agricultural extension, production management, and policy services. However, these texts typically contain highly specialized terminology, exhibit non-standard, colloquial, and diverse linguistic expressions, present fragmented semantics, and rely heavily on contextual reasoning. Such characteristics make them difficult to parse accurately using traditional rule-based approaches or shallow machine learning models. Consequently, these limitations often lead to biased demand classification and incomplete extraction of key factors, thereby constraining the quality of data available for intelligent agricultural decision-making. To address these challenges, the aim of this research is to develop a robust, domain-adapted, and highly interpretable structured analysis method for agricultural user demands. [Methods] Agri-NeedAgent, an agricultural user demand analysis framework, was proposed based on a "three-stage training + multi-agent collaboration" paradigm. First, during the domain knowledge pretraining stage, 80 000 agriculture-related texts, including crop cultivation manuals, pest and disease control guides, agricultural policy documents, and farmer consultation records, were used to construct domain-specific semantic understanding, thereby enhancing the model's capability to interpret agricultural terminology, dialectal expressions, contextual logic, and implicit semantics. Second, in the instruction fine-tuning stage, 6 320 annotated samples in an "instruction-input-output" format were employed to establish an explicit mapping from raw demand texts to structured outputs. Third, in the agricultural knowledge low-rank adaptation stage, Low-rank Adaptation (LoRA) was applied to perform lightweight parameter tuning on task-specific agents, enabling targeted adaptation for demand classification and key-factor extraction tasks. Built upon the above training process, a multi-agent collaborative framework was constructed, in which the manager agent was responsible for task scheduling and quality control, while task agents were designed to perform demand classification, key-factor extraction, and explanation generation, respectively. Through this division of labor and collaborative mechanism, the framework achieved efficient and structured analysis of agricultural user demands. [Results and Discussions] Experimental results demonstrate that the proposed Agri-NeedAgent achieved a demand classification accuracy of 84.6%, a key-factor extraction F₁-Score of 85.2%, a structured interface compliance rate of 94.2%, and an interpretability score of 90.2.These results showed clear improvements over traditional deep learning models such as Bidirectional Encoder Representations from Transformers (BERT) as well as general-purpose large language models (LLMs) without domain adaptation. The findings confirmed the critical role of domain knowledge injection, explicit task alignment, and multi-agent specialization in enhancing semantic understanding and structured analysis of agricultural texts. Ablation experiments further validated the effectiveness of each component. Removing domain pretraining or LoRA fine-tuning resulted in substantial performance degradation in classification and key-factor extraction, indicating the necessity of domain adaptation and task-specific optimization for handling non-standard agricultural expressions. Moreover, eliminating the manager agent or the Reasoning and Acting (ReAct) mechanism significantly reduced structured interface compliance and interpretability, highlighting the importance of task coordination, intermediate verification, and multi-step reasoning for ensuring logical consistency and output completeness. Additionally, removing the external knowledge base reduced the interpretability score from 90.2 to 77.6, underscoring its essential role in providing theoretical grounding, reasoning support, and professional explanations. Although the multi-agent collaboration introduced an additional inference overhead of approximately 140 ms, the overall per-sample inference time remained within 225 ms, meeting the real-time requirements of agricultural consultation scenarios. [Conclusions] Supported by a "three-stage training + multi-agent collaboration" framework, LLMs can effectively address challenges posed by non-standard expressions, semantic fragmentation, and multi-factor reasoning in agricultural user demand texts. The proposed method demonstrated significant improvements in demand classification, key-factor extraction, structured output compliance, and interpretability, providing high-quality and traceable structured data for intelligent agricultural decision-making. After domain adaptation and task-specific tuning, the model not only gains enhanced capability for deep semantic analysis of agricultural user demands but also ensures the completeness and interpretability of outputs through multi-agent coordination. Although the current workflow still requires optimization in terms of data preparation, staged training, and knowledge-base updating, future work will focus on expanding region-specific and emerging-technology-related demand data, developing a dynamically updated agricultural knowledge system, improving multi-agent coordination efficiency, and exploring cross-lingual agricultural demand analysis to further promote the application and deployment of agricultural large models across broader scenarios.

[Objective] There are several critical challenges in automated safflower harvesting, particularly the inefficiencies in path planning, suboptimal route quality, and limited decision-making capability under dynamic and complex environments. To solve these issues, the problem was formulated as a three-dimensional traveling salesman problem and an enhanced reinforcement learning model named actor-critic reinforcement learning pointer network (AC-RL-PtrNet) was proposed, specifically designed for deployment on intelligent safflower picking robots in agricultural settings. [Methods] First, to address the inherent limitations of conventional attention mechanisms in dynamic environments with complex spatial structures, an enhanced attention module was proposed based on the dynamic exponential moving average framework. By combining multi-head attention, spatial distance encoding, and adaptive exponential smoothing, the improved design allowed the model to better capture long-range dependencies and spatial context among safflowers. Meanwhile, to minimize computational cost while preserving inference quality, a structured pruning approach was adopted, which selectively removed redundant connections in the long short-term memory gates and fully connected layers. In parallel, the critic network was redesigned to improve learning stability and accuracy. This was achieved through the inclusion of batch normalization, residual feature aggregation, and a multi-layer value estimation head, all of which contributed to a tighter actor-critic synergy during policy training. [Results and Discussions] To quantitatively assess the impact of each component, ablation experiments were conducted across various configurations. The results confirmed that each module contributed distinct benefits, while their combination yielded the highest improvements in both planning precision and inference efficiency. This coordinated actor-critic design effectively enhanced both trajectory quality and decision stability, which were critical in sequential robotic picking tasks. Experimental results also demonstrated that, compared with traditional swarm intelligence algorithms particle swarm optimization (PSO), ant colony optimization (ACO), and non-dominated sorting genetic algorithm, the proposed AC-RL-PtrNet model achieved a planning time improvement ranging from -2.63% to 61.87% on the 25-target dataset and from 22.93% to 59.1% on the 31-target dataset. Meanwhile, the optimized paths were significantly shortened across different planning instances, indicating robust generalization capability under varied problem scales. Furthermore, field experiments provided concrete validation of the model's practical applicability. When deployed on a mobile picking robot in real safflower fields, the AC-RL-PtrNet achieved a 9.56% reduction in path length and 5.43% time saved for a 25-target picking task, and a 20.17% path reduction and 29.70% time saving for a 31-target scenario involving a different safflower variety. Overall, these results all indicated that the proposed method exhibited significant advantages in enhancing path planning efficiency and optimizing path quality. [Conclusions] This study offers a practical solution for achieving efficient and robust automatic picking by safflower picking robots and provides new insights into solving 3D combinatorial optimization problems.

[Objective] Large language models (LLMs) have demonstrated strong capabilities in natural language understanding, knowledge integration, and complex reasoning, offering new opportunities for intelligent decision-making in agriculture. However, their direct application in agricultural production and facility environment control remains challenging due to strong physical constraints and high operational risks. The lack of real-world interaction and executable decision grounding limits the practical effectiveness of conventional LLMs in such scenarios. To address these challenges, a tool-augmented LLM-based agricultural intelligent agent system, termed AgriAgent, was proposed, and a digital-twin-based evaluation platform for agricultural decision-making was developed. By integrating a high-fidelity digital twin environment with an end-to-end agent architecture, the decision-making performance of agricultural intelligent agents with different parameter scales was systematically evaluated across multiple crops and climate scenarios. [Methods] A high-fidelity agricultural digital twin evaluation platform was constructed using the decision support system for agrotechnology transfer (DSSAT) v4.8 crop growth model as the core simulation engine to model crop growth under diverse environmental conditions and management strategies. Meteorological driving data were obtained from the Seoul Historical Weather Data dataset. Through data cleaning, missing-value imputation, unit normalization, and time-series reconstruction, the raw meteorological data were transformed into standardized inputs compatible with DSSAT. Three climate scenarios representing different environmental complexities were designed, including a regular scenario, a perturbed scenario, and an extreme scenario. The regular scenario employed historical observations, the perturbed scenario introduced stochastic disturbances to simulate short-term climate variability, and the extreme scenario incorporated multi-factor coupled stresses such as high temperatures and excessive precipitation during sensitive growth stages. In total, 90 annual climate driving sequences were generated. Fixed soil profile parameters calibrated by domain experts were applied across all simulations to minimize confounding effects. Within this digital twin environment, a tool-augmented agricultural intelligent agent, AgriAgent, was implemented using a modular architecture consisting of a sensor module, memory module, retriever, large language model, and tool executor, forming a closed-loop decision-making framework. In each decision cycle, the agent perceived environmental and crop state information, including soil moisture and nutrient status, meteorological conditions, crop growth stages, and stress indicators. State summaries and historical decisions were stored in memory, while agronomic knowledge was retrieved through a retrieval-augmented generation mechanism. Based on integrated information, the LLM generated structured environmental control commands in JSON format, which were validated and constrained by the tool executor before updating the DSSAT environment. The system supported irrigation, supplementary lighting, ventilation, heating, fertilization, and CO₂ enrichment. Five representative crops: maize, millet, sugar beet, tomato, and cabbage were simulated under the three climate scenarios over complete growing seasons, resulting in 450 crop-scenario combinations. An unmanaged DSSAT simulation served as the baseline. AgriAgent models with three parameter scales (1.5B, 3B, and 7B), built on the Qwen2.5 series, were evaluated. Crop economic yield expressed as dry matter at physiological maturity was adopted as the evaluation metric. [Results and Discussions] The results showed that AgriAgent consistently outperformed the baseline across all crops and climate scenarios, with model scale exerting a significant influence on decision-making performance. AgriAgent-7B achieved the best overall performance under regular, perturbed, and extreme scenarios, demonstrating strong generalization ability and environmental adaptability. By dynamically adjusting water, nutrient, light, and thermal management strategies, the agent effectively mitigated environmental stresses even under multi-factor coupled extreme climate conditions. Under extreme scenarios, AgriAgent-7B increased yields by 463.60% for maize, 351.20% for millet, 125.40% for sugar beet, 1 537.46% for tomato, and 1 185.14% for cabbage compared with the baseline. Particularly large gains were observed for high-value crops such as tomato and cabbage, highlighting the advantages of the proposed framework for precision-controlled facility agriculture. In contrast, AgriAgent-1.5B exhibited performance comparable to the baseline, while AgriAgent-3B achieved moderate improvements but remained inferior to the 7B model. These findings indicate a clear scaling effect, suggesting that larger models possess stronger capabilities in multi-source information integration, long-term temporal reasoning, and adaptation to complex environments. [Conclusions] This study developed a digital-twin-based agricultural decision evaluation platform and proposed a tool-augmented, end-to-end agricultural intelligent agent named AgriAgent. Experiments across multiple crops and climate scenarios verified the effectiveness and robustness of the proposed framework for dynamic agricultural decision-making. The results demonstrate that integrating knowledge retrieval, reasoning, and tool execution within a closed-loop LLM-based agent enables stable, reliable, and adaptive environmental control, providing a feasible technical pathway and standardized evaluation paradigm for intelligent agriculture.