This study proposes a multi-sensor fusion framework to address the limitations of single-sensor Unmanned Aerial Vehicle (UAV) systems in rice leaf area index (LAI) estimation. Three UAV platforms equipped with multispectral, RGB, or LiDAR sensors were employed to collect data during three critical rice growth stages. Ordinary Least Squares (OLS) regression models were first constructed to estimate LAI by using the single feature that exhibited high correlation with rice LAI. Results revealed that models based on Laser Penetration Index (LPI; R2 = 0.649, RMSE = 0.876) and Height Percentile Area 0-50 (HPA0-50; R2 = 0.601, RMSE = 0.933) derived from LiDAR sensor exhibited the best performance among all single features. Three machine learning algorithms (Gradient Boosting Regression, Random Forest, and Stacking) were further employed to construct LAI estimation models by fusing vegetation indices, colour indices, texture features, and spatial structural features extracted from three sensors. The Random Forest model based on multi-sensor fusion data achieved optimal performance (R2 = 0.861, RMSE = 0.600), demonstrating a 32.66% accuracy improvement over the single-feature OLS model. Permutation importance analysis was conducted to quantify sensor contributions, revealing the LiDAR sensor as the dominant contributor (67.78% of total importance), followed by multispectral (27.46%) and RGB sensors (4.75%). It was also demonstrated that a single-LiDAR sensor achieved higher accuracy (R2 = 0.812) for LAI estimation, much higher than the single-multispectral or single-RGB sensor, which benefited from spectral-texture fusion. These results demonstrated that multi-sensor fusion was an efficient approach to achieve higher accuracy for rice LAI estimation, with LiDAR-driven spatial structural features serving as the key role for high-accuracy LAI estimation.

Optimizing water and fertilizer management is crucial for improving cotton yield and quality. However, reliable and generalizable models for quickly and accurately estimating cotton canopy leaves water and nutritional status at a low cost throughout the entire growth stage are scarce. Therefore, this study aims to construct the generalization and adaptation retrieval model of cotton canopy leaf nitrogen content (LNC) and equivalent water thickness (EWT) based on PROSAIL, hyperspectral reconstruction with UAV multispectral imagery and module transfer learning. In the hyperspectral reconstruction module, the new hyperspectral reconstruction model (swinT-HSCNN) based on multispectral showing superior performance in reducing pixel-scale systematic errors and effectively captured spectral variations than HSCNN+ and MST++ model. In the PROSAIL module, the proposed Original-E2DCOS method demonstrated greater sensitivity to spectral response characteristics, especially for parameters and bands with low correlation values, and three bands (702 nm, 762 nm, and 938 nm) were selected as the sensitive bands corresponding to chlorophyll content (Cab) and equivalent water thickness (Cw) of cotton. The improved PROSAIL with hyperparameter optimization based on full spectrum and multispectral band shown better fitting performance than the model based on sensitive bands, and achieved high accuracy on simulated data, with R2 values exceeding 0.98 for both Cab and Cw. Moreover, the new developed modular transfer learning retrieval model of cotton canopy water and nitrogen content through PROSAIL model and hyperspectral reconstruction with UAV multispectral imagery achieved good inversion accuracy with R2 of 0.83, 0.85, RMSE of 0.0048, 0.0052, for LNC and EWT, respectively after verifying with actual experiment data. In summary, the proposed modular transfer learning retrieval model of cotton canopy water and nitrogen content integrates physical constraints into retrieval models, which enhancing their accuracy and generalization capability, and providing valuable technical support for precision agriculture in cotton production across different regions. (c) 2026 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Sweetpotato is a crucial food crop globally, valued for both its economic significance and health benefits. However, the prevalence of sweetpotato virus diseases (SPVD) poses a serious threat to the industry, leading to reduced yields and economic losses for farmers. Efficient diagnostic techniques are essential for ensuring food security and consumer health. Traditional diagnostic methods are effective but suffer from complexity, time consumption, and high costs. To address these challenges, a novel real-time end-to-end detector called SPVD-DETR based on the Transformer architecture is proposed in this study. By utilizing the unmanned aerial vehicle (UAV) orthomosaic image, SPVD can be diagnosed in real-time at the field scale. First, aerial survey tasks are customized with automated drone tools to rapidly scan sweetpotato fields, generating high-resolution orthophotos and a stitched orthomosaic image for analysis. Then, the object detector is enhanced by incorporating efficient backbones and hybrid encoder modules such as cascaded group self-attention, attention-based scale fusion, and dynamic upsampling. Extensive ablation studies and comparative results show that SPVD-DETR achieves a good balance between real-time performance and accuracy. Next, the model is fine-tuned on the SPVD image tiles and achieves a detection accuracy of 31.3% mean average precision (mAP) with the fastest inference speed of 90 frames per second (FPS). Finally, the prediction results are mapped back to the orthomosaic image, estimating an overall SPVD incidence rate of 15% with a misdiagnosis rate of 14%. This study introduces a novel paradigm for detecting SPVD at the field scale, promoting automatic and intelligent plant disease detection for large-scale high-throughput phenotyping in precision agriculture.

Accurate and rapid identification of corn diseases is a prerequisite for precision spraying in the field. Yet the highly variable outdoor environment, together with the computational cost of deep models, still constrains real-world deployment. To address these challenges, we propose Dycorn-YOLO11, a deployable corn disease detector built on a dual-dynamic feature capture architecture design. First, the online convolutional reparameterization block (OREPA) is embedded in the backbone to reduce training and inference complexity. Second, a dual-dynamic feature capture architecture is constructed, where v3-Dyhead and Dysample work in concert to achieve deformable alignment and content-adaptive sampling, strengthening the representation of small, low-contrast lesions. In addition, a “channel pruning + knowledge distillation” lightweight pipeline compresses the model substantially while preserving the backbone’s feature expressiveness. Experiments show that the improved model attains a precision (P) of 88.2 %, recall (R) of 73.9 %, and mAP of 81.5 %, representing gains of 0.8, 2.1, and 3.4 percentage points over YOLO11n. The parameter count and model size are reduced by 41.5 % and 40.0 %, respectively, and the inference speed increases to 113.2 f/s, achieving a balanced trade-off among accuracy, compactness, and efficiency. The model also maintains high stability on an external, complex dataset, and delivers strong real-time performance in both static and moving field experiments. Overall, Dycorn-YOLO11 demonstrates robust, deployable performance and provides a feasible technical foundation for precision spraying in corn fields.

[Objective]To enhance the application performance of unipolar contact electrostatic spraying technology in agricultural plant-protection unmanned aerial vehicles(UAVs),this study investigates droplet deposition characteristics under different crosswind and height conditions,aiming to optimize the spray system design,improve pesticide utilization,and reduce environmental pollution.[Method]A computational fluid dynamics(CFD)discrete phase model combined with user-defined functions(UDF)was developed to simulate the electrostatic spray field.A charge-to-mass ratio measurement platform was used to obtain droplet charge-to-mass ratios under different voltages as the initial model parameters.The deposition characteristics of droplets for the unipolar contact-type electrostatic spraying system,under different wind speeds and spraying heights were analyzed through simulation,and the model was validated by outdoor experiments.[Result]Simulation results showed that within the spraying height range of 1.5-1.8 m,electrostatic spraying achieved more uniform deposition and smaller crosswind drift compared with conventional spraying.At 1.5 m,the liquid film distribution was stable and less affected by wind speed.At 1.6 m,when wind speeds were 2 and 3 m/s,the liquid film peak decreased by 25 and 34 μm,respectively,compared with 112 μm at 1 m/s.At 1.8 m,the liquid film peak under all wind speeds decreased by 25.9%,26.3%and 17.3%,respectively,compared with that at 1.5 m.Outdoor experiments further demonstrated that the center-of-mass distance of electrostatic spraying was reduced by 13.05%compared with conventional spraying.[Conclusion]The unipolar contact electrostatic spraying system applied to agricultural UAVs exhibits excellent droplet deposition performance.The simulation results are consistent with outdoor experimental deposition trends,providing theoretical basis and practical guidance for the optimization of aerial electrostatic spraying technology and precision agriculture.

Yield monitoring is crucial for the agricultural sector, as it can be used to inform decisions on harvesting, storage, and transportation. Traditionally, several statistical methods and visual inspection techniques are employed to get an early estimate of the final yield of citrus, with the downside of being inaccurate, costly, and time-consuming. In recent years, there have been a lot of advancements in the fields of Artificial Intelligence (AI) and computer vision, providing opportunities to automate plenty of things in different domains, including agriculture. This research proposes a deep learning-based framework that leverages multiple Convolutional Neural Networks (CNN) to efficiently and effectively operate in real-world environments, using field data to provide accurate, improved yield estimates. A high-quality dataset, consisting of citrus tree images, is obtained from orchards at the university research farm Koont and the National Agriculture Research Center (NARC). Afterwards, the CNN-based models are trained thoroughly with various configurations and data augmentation techniques. All the models are rigorously tested and evaluated on the basis of a number of performance metrics. Experiments have shown that YOLOv8m performs with the highest mean average precision, reaching up to 90% with an inference time of a few milliseconds, making it worthy to be deployed for fruit detection, counting, and yield estimation tasks.

The critical growth stages of crops have short observation windows and are easily affected by environmental factors, which often limits the availability of field measurements for remote sensing inversion. This limitation is more pronounced in the synergistic inversion of leaf area index (LAI) and leaf chlorophyll content (LCC), where canopy structure and chlorophyll related traits may show similar spectral responses in the visible region. Such spectral overlap can lead to parameter confusion and estimation bias, highlighting the need for methods that can expand training samples while preserving the relationships among spectral features and crop traits. With recent advances in generative modeling, Generative Artificial Intelligence (GAI) provides a potential solution by learning data distributions and generating synthetic samples that preserve these relationships. To address these issues, the study developed a conditional inversion for Remote Sensing model based on a Denoising Diffusion Probabilistic Model (CIRS-DDPM). In this model, LAI and LCC were embedded as continuous conditional variables and fused into the noise prediction network during reverse denoising, allowing the generated samples to better preserve the joint distribution among crop traits and remote sensing features. A Fully connected residual network for One-Dimensional Sequence Features (1D-ResMLP) was introduced as the backbone to improve noise prediction for tabular remote sensing data. The samples generated by CIRS-DDPM were compared with those generated by the Tabular Denoising Diffusion Probabilistic Models (Tab-DDPM) and PROSAIL. The results showed that CIRS-DDPM generated data had feature distributions highly consistent with measured samples. Models trained with CIRS-DDPM generated data also showed strong validation performance using real measurements. The one-dimensional residual network (1D-ResNet) model achieved R2 = 0.822, RMSE = 1.314, and MAE = 1.087 for LAI, and R2 = 0.735, RMSE = 10.397, and MAE = 8.251 for LCC. Compared with models trained using PROSAIL and Tab-DDPM generated data, the R2 values increased by 20.0 % and 57.0 % for LAI, and by 18.4 % and 31.8 % for LCC, respectively. These results indicate that CIRS-DDPM can improve data support for the synergistic inversion of LAI and LCC and provide an effective solution for data limited remote sensing modeling in agriculture.

Objective Citrus Huanglongbing (HLB) transmission is influenced by the coupling of multiple dynamic factors. Traditional optimal control methods face the limitations in practical applications due to their high computational complexity and reliance on precise models. To address this problem, this paper proposes an intelligent dynamic prevention and control method for HLB based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm.MethodFirstly, based on the transmission dynamics of HLB, a HLB propagation dynamics model of the interaction mechanism between host and vector was established. On this basis, the HLB transmission control dynamic model was discretized to construct a Markov Decision Process environment suitable for deep reinforcement learning. Subsequently, the TD3 algorithm was introduced, and a multi-objective reward function compatible with biological constraints was designed. Finally, an HLB prevention and control strategy was proposed.ResultSimulation experimental results demonstrated that the proposed dynamic prevention and control strategy for HLB based on TD3 exhibited the significant advantages over traditional algorithms across multiple key performance indicators. Compared to DDPG and PPD, the speed of system state convergence to the disease-free equilibrium point increased by 26.59% and 20.99% respectively, the cumulative control cost reduced by 23.79% and 19.90% respectively, and the peak pesticide usage decreased by about 35.57%. Numerical analysis further showed that timely spraying insecticide during the early stages of HLB outbreak played a critical role in interrupting the transmission chain and preventing large-scale epidemics. Compared with constant control strategies, dynamic control strategies had more advantages in suppressing the spread of diseases and reducing the cost of implementing control measures.ConclusionThe HLB prevention and control method based on TD3 proposed in this study provides a new perspective for the efficient control of HLB transmission, and demonstrates the potential of deep reinforcement learning methods in agricultural disease prevention and control.

ObjectiveTo address the issues of hallucinations, contextual logical inconsistencies, and inability to run on low-resource devices when large language models perform named entity recognition (NER) in agriculture.MethodUsing DeepSeek with 671 billion parameters (DeepSeek-671B) as the teacher model, domain knowledge was transferred to student models with fewer parameters. The student models selected were low-parameter versions of DeepSeek, Qwen, and Llama (1.5 billion, 7.0 billion, and 14.0 billion parameters, abbreviated as 1.5B, 7.0B and 14B respectively), which underwent distillation and counterfactual reasoning training. Model performance was experimentally validated on the CropDiseaseNer dataset, a specialized agricultural disease dataset.ResultBy comparing the performance of a series of distilled student models, the results showed that DeepSeek-14B achieved an entity recognition F1 score of 89.60% while requiring only 2.08% of the parameters of the teacher model. Its performance significantly outperformed both the general-purpose large model GPT-mini-14B (F1 score: 57.64%) and the domain-adapted model GLiNER (F1 score: 82.96%) based on a general LLM. Further analysis revealed that the DeepSeek student model, sharing the same architecture, demonstrated superiority over models with different architectures in recognizing long-tail categories such as disease entities and pathogen genus names, owing to its parameter alignment advantage.ConclusionThis study validates the effectiveness of knowledge distillation in NER tasks within the agricultural domain, offering a novel solution for entity recognition technology in resource-constrained scenarios.

As an important agricultural product, the quality and commercial value of soybeans largely depend on their geographical origin. Addressing the issues of low detection efficiency, complex operation, and long detection times in traditional soybean origin traceability, this study proposes a novel method for the rapid and non-destructive detection of soybeans from different producing areas by the synergetic application of near-infrared (NIR) spectroscopy and electronic eye (EE) combined with an improved CNN-VMamba mixed network (CVMNet). First, the spectral signal and visual image of soybean samples are collected by an NIR spectroscopy and EE, respectively. Subsequently, the Gramian Angular Difference Field (GADF) is implemented to convert the 1D spectral signal into a 2D spectrogram for effectively characterizing inter-wavelength correlation patterns and nonlinear spectral dependencies within the signals. The spectrograms and EE images are then input into the CVMNet model to perform pattern recognition. This model first utilizes hybrid multi-scale dilated dynamic convolution (H-MDDC) to effectively capture local detailed features within the images. The VMamba structure is then employed to exploit long-range dependency among pixels and optimize computational efficiency. A modality interaction module (MIM) is then introduced to facilitate information interaction between the bi-modal features. Finally, a co-attentional modality fusion (CoAMF) mechanism is adopted to perform information fusion and classification on the NIR and EE information. The experimental results indicate that, compared to using NIR spectroscopy or EE individually, the proposed method yields higher recognition accuracy for detecting soybean origin, with its detection accuracy reaching 98.67

Cadmium (Cd) contamination in vegetables poses a serious threat to food safety, making the development of an efficient, nondestructive detection method crucial. This study integrates hyperspectral technology, feature selection, and machine learning to propose a nondestructive method for assessing Cd contamination levels in cabbage leaves. Hyperspectral data were collected from the adaxial and abaxial surfaces of cabbage leaves under eight different Cd stress conditions. After preprocessing the data and selecting feature wavelengths, machine learning models were constructed. Results indicate that spectral data from the adaxial surface of cabbage leaves outperformed data from the abaxial surface in distinguishing Cd concentration levels. After applying feature selection methods, all performance metrics of the models significantly improved across both the calibration and test datasets. Among these, the Uninformative Variable Elimination (UVE)-Competitive Adaptive Reweighted Sampling (CARS) combination method demonstrated superior filtering performance compared to UVE and CARS alone, successfully reducing the number of original bands from 137 to 25 while preserving key spectral information related to Cd contamination. Among all model combinations, the Light Gradient Boosting Machine (LightGBM) model constructed using adaxial leaf spectral data and UVE-CARS-filtered features demonstrated the highest discrimination performance. Its test set achieved an accuracy of 0.992, a precision of 0.990, a recall of 0.992, and an F1 score of 0.992. This study confirms that the combined approach of leaf adaxial surface spectroscopy and UVE-CARS-LightGBM can effectively distinguish cabbage leaves under different Cd stress concentrations, providing a new pathway for efficient nondestructive detection of heavy metal contamination in vegetables.

ObjectiveLitchi is one of the most representative characteristic fruit trees in the Lingnan region, and the management of its autumn and winter shoots as well as nutrient regulation is directly related to fruit trees yield. Affected by climatic conditions and the vertical canopy structure of the tree crown, litchi trees are prone to apical flushing and asynchronous shoot emergence of autumn and winter shoots, resulting in nutrient waste. Therefore, achieving accurate segmentation of litchi autumn and winter shoots provides a critical basis for subsequent precision management. MethodFirst, high-resolution images of litchi autumn and winter shoots were acquired via low-altitude UAVs over two years. Second, the SMA-DeepLab model was proposed for accurate segmentation of litchi autumn and winter shoots. In this model, the backbone network of DeepLabv3+ was replaced with SMANet. The main network of SMANet improved feature quality through StarNet and integrated features with the adaptive spatial feature fusion (ASFF) module for multi-scale feature fusion. Meanwhile, a receptive field aggregator (RFA) was introduced to enhance boundary precision. ResultIn terms of accuracy performance, the mean pixel accuracy (mPA) and mean intersection over union (mIoU) were 93.46% and 87.84%, respectively, representing improvements of 2.74 and 2.75 percentage points compared with the baseline model. In terms of efficiency, the floating-point operations per second (FLOPS) and frames per second (FPS) were 111.44 and 31.63, respectively, and the number of parameters was reduced by 51.3% compared with the baseline model. In addition, visualization of segmentation results showed that the proposed model achieved accurate segmentation of slender shoots and motion-blurred regions when facing interfering factors like complex backgrounds. ConclusionThe SMA-DeepLab model proposed in this study provides an effective solution for the segmentation of litchi autumn and winter shoots and serves as a technical reference for other objects segmentation tasks in the field of smart agriculture.

To address the reduction in nozzle row-following accuracy caused by sensing–execution latency during corn row-following operations, a delay compensation method based on machine vision and dynamic region of interest (ROI) adjustment was proposed. The method integrates real-time forward-velocity information from a global navigation satellite system/inertial measurement unit (GNSS/IMU), decomposes the delays associated with image processing, command transmission, and actuator motion, and calculates a visual look-ahead distance from the total response delay and robot forward velocity. Inverse-perspective mapping was used to establish the relationship between pixel and world coordinates, and the ROI position was dynamically shifted to synchronize the sensing–execution process. Indoor bench tests showed that, under variable conveyor-belt speeds ranging from 0 to 0.25 m/s, the algorithm achieved a row-following accuracy of 93.75% and a lateral mean absolute error of 0.019 m; compared with the average result of the three fixed-ROI tests, the lateral mean absolute error was reduced by 24.8%. Whole-machine tests showed that, under random platform forward speeds of 0–1.00 m/s, the row-following accuracy remained above 85.71%, with a lateral mean absolute error of 0.034 m. The results indicate that the proposed method effectively compensates for system delay under different speed conditions and reduces lateral tracking errors caused by longitudinal spatiotemporal mismatch, providing technical support for the development of precision corn row-following spraying equipment.

Accurate estimation of crop water status is essential for monitoring plant senescence and enabling intelligent agricultural management. This study proposes a pixel-aligned co-registration and DSM-grid fusion framework that integrates high-resolution point clouds, multispectral (MS) images, and thermal imagery acquired by Unmanned Aerial Vehicles (UAVs) to enable three-dimensional prediction and visualization of cotton canopy leaf water content (LWC) and equivalent water thickness (LEWT). To address the low spatial resolution of thermal imagery, a downsampling-upsampling simulation framework was developed to evaluate interpolation errors. This framework quantitatively compares three common interpolation methods-nearest neighbor, bilinear, and bicubic interpolation-using RMSE and PSNR metrics. Results show that bicubic interpolation performs best in preserving spatial details and minimizing errors, and is therefore adopted in the subsequent image fusion process. A 3D grid was constructed based on the digital surface model (DSM), enabling grid-cell (pixel-aligned) spectral and thermal features to be mapped onto point-cloud units. Vegetation and thermal indices extracted from the mapped features were used as input variables. Combined with recursive feature elimination (RFE) and random forest (RF) models, the prediction of LEWT and LWC achieved R2 values of 0.792 and 0.752, and rRMSE values of 13.84% and 9.68%, respectively. These results significantly outperformed those of partial least squares regression (PLSR), support vector machine (SVM), and extreme learning machine (ELM) models. By integrating the predicted results with the point cloud data, a 3D representation of canopy water parameters was constructed, revealing a typical top-down gradient of water loss. The experiment also revealed that nitrogen treatment significantly influenced the vertical distribution of water content. High-nitrogen application delayed water loss in the middle and lower canopy layers, highlighting the coupled regulation between nitrogen and water. Parameter comparisons showed that LEWT exhibited higher sensitivity than LWC across both temporal and spatial scales, making it a more robust indicator for canopy water monitoring. Additionally, point clouds generated from Cross-circling oblique (CCO) photogrammetry outperformed UAV LiDAR systems in terms of point density, structural completeness, and image fusion potential. In summary, this study validated the feasibility and effectiveness of integrating point cloud, MS, and thermal imagery via the proposed pixel-aligned co-registration and DSM-grid fusion framework for 3D crop water monitoring. The proposed method provides a reliable technical foundation for drought detection, irrigation management, and yield prediction in precision agriculture.

Accurately detecting tea buds under top-view imaging is essential for automating tea harvesting. However, this task is particularly challenging due to the buds’ extremely small size, blurred boundaries, high visual similarity to background leaves, and low inter-class variance—all of which result in sparse features and frequent mislocalization. Existing CNN-based detectors often struggle to preserve fine spatial details while modeling global context in such scenarios. To address these issues, we propose a Transformer-based framework for robust top-view tea bud detection, leveraging its global attention mechanism to enhance feature discrimination for tiny and ambiguous objects. Specifically, a Gradient-guided Class Activation Attention module enhances spatial sensitivity by generating attention maps from mid-level CNN features via gradient signals, allowing the model to focus on weak yet semantically relevant regions. An Explicit-Guided Decoder reinforces inter-layer consistency and structural modeling by leveraging feedback from previous predictions and stage-wise feature injection. In addition, a Spatial Position-Sensitive Enhancement Loss introduces a Gaussian-based weighting mechanism to emphasize central object regions, improving localization accuracy for tiny targets. Experiments on a top-view tea bud dataset demonstrate that our method outperforms state-of-the-art baselines (YOLOv11, DETR, RT-DETR), achieving a Precision of 0.8371 and mAP@0.5 of 0.8013. Moreover, evaluations on three public tiny-object detection datasets further confirm its generalization across domains. These results highlight the effectiveness and transferability of the proposed approach in both agricultural and general tiny-object detection scenarios.

Precise interventions in the management of fruit tree diseases and pests are of great importance. However, most existing studies have focused primarily on disease detection and identification, with relatively little attention given to the progression of affected trees and the effectiveness of subsequent interventions. In this study, we developed a deep survival analysis model—REK-Surv—based on the Kernel Attention Network framework to predict the prognosis of citrus trees infected with huanglongbing. During a 15-month follow-up, we applied three different intervention strategies to 56 trees and collected a range of potential prognostic factors, including the intrinsic health status of the trees, surrounding ecological conditions, and environmental variables such as climate and soil properties, in order to evaluate their effects on survival outcomes. Experimental results demonstrated that the REK-Surv model effectively captured the risk function of diseased trees, achieving a high prediction accuracy of 99%. Analysis of survival curves, CT values, and SHAP values for the different interventions revealed that chemical treatment had a certain inhibitory effect on huanglongbing progression, whereas pruning showed no significant effect. Contribution analysis of prognostic factors identified light intensity, wind direction, soil electrical conductivity, and wind speed as key determinants of tree survival. Furthermore, June and November were identified as high-risk periods associated with sharp increases in mortality rates. This study provides quantitative evidence supporting precise interventions for huanglongbing-infected citrus trees, expands the application of survival models in plant disease prognosis, and contributes to more scientific and forward-looking pest and disease management in orchards.

Spray drift during pesticide application remains a critical environmental concern, particularly under natural wind conditions. This study presents a field evaluation of drift behavior of knapsack and boom sprayers based on 15 standardized field trials conducted in accordance with ISO 22866:2005. Experiments were performed under three wind speed categories: low (0-3.3 m s-1), moderate (3.4-5.4 m s-1), and high (>= 5.5 m s-1). Drift deposition, droplet density, and spray coverage were evaluated using two hollow-cone nozzles on knapsack sprayers (Li Cheng KZ 80-06 and KZ 80-20) and two flat-fan nozzles on boom sprayers (Li Cheng VP110015 and VP11005). Results showed that boom sprayers were more susceptible to wind effects, with drift increasing significantly under moderate and high wind conditions. The boom sprayer with the VP11005 nozzle generated the highest near-field drift, reaching 108% at 1 m distance. Conversely, the knapsack sprayer with the coarser KZ 80-20 nozzle produced minimal drift, recording only 36.9% at 1 m under moderate winds. A strong positive correlation (R2 = 0.94) was observed between spray coverage and drift deposition. The findings recommend restricting Li Cheng VP11005 nozzle operation to low-wind conditions only, providing practical guidance for drift mitigation and for optimizing pesticide application strategies.

To address the plant protection challenges of high canopy closure and poor droplet penetration in high-density maize, this study systematically elucidated the synergistic regulatory mechanisms of UAV operational parameters and reduced nano-pesticide application. Integrating laboratory microscopic characterization with a two-phase field experiment, this study evaluated spray volume, droplet size, and pesticide dosage as core variables. Phase I quantified the effects of physical spray parameters on droplet deposition and penetration characteristics, while Phase II comprehensively revealed the field control efficacy under multi-factor synergy. Results showed that 22.5 L/ha achieved the optimal balance; it increased deposition density by 109.0% compared to 15.0 L/ha (p < 0.05) while maintaining comparable efficacy to 30.0 L/ha without sacrificing operational efficiency. The 200 mu m droplets demonstrated superior penetration, increasing lower-layer coverage by 33.9% compared to 300 mu m (p < 0.05), and overcoming the evaporation limitations of 100 mu m (which showed no significant efficacy improvement despite higher density, p > 0.05). Microscopic measurements confirmed that nano-pesticides exhibited excellent wetting dynamics, reducing the final contact angle on hydrophobic maize leaves to as low as 28.24 degrees (achieving a maximum dynamic reduction rate of 54.48%), significantly enhancing interfacial adhesion. This mechanism compensated for low-volume spraying limitations, allowing a 30% dosage reduction to maintain a robust field efficacy of 90.6%, whereas a 50% reduction significantly compromised efficacy (<72.5%). Ultimately, this study established "22.5 L/ha + 200 mu m + 30% reduction" as the optimal operational parameter combination, providing a solid theoretical foundation and crucial technological support for precision plant protection and "pesticide reduction with efficiency enhancement" in high-density maize cultivation.

In agricultural artificial intelligence image classification tasks, existing models often exhibit redundant feature extraction and limited generalization capability. To address these practical demands and technical challenges, we propose a Cognitive-Inspired TriPercept CrossFusion Self-Distillation Network (CTNet). First, to counteract the adverse effects of environmental variations and motion blur inherent in agricultural image acquisition, we introduce a comprehensive data augmentation strategy specifically designed for agricultural imagery, thereby enhancing model robustness. Second, we present an efficient and lightweight TriPercept CrossFusion Network (TPCF), which concurrently captures multi-branch representations and emphasizes discriminative local features. A cross-branch fusion mechanism facilitates inter-branch feature interaction, enabling effective feature learning with minimal computational overhead. Third, we devise a novel end-to-end trainable self-distillation framework, termed Cognitive-Inspired Multi-Stage Self-Distillation (CISD). CISD operates without external pretrained teacher models; instead, it performs intra-model knowledge transfer across layers and employs a composite objective combining classification loss (cross-entropy), distillation loss (KL divergence), and hint loss (feature map alignment). Inspired by cognitive load theory, the framework dynamically adjusts the weighting of these loss components to achieve cognitive-load-aware multi-stage self-distillation, thereby improving generalization. We evaluated CTNet on a custom lychee disease dataset and six public benchmarks. On our augmented lychee dataset, CTNet achieves 99.03% accuracy. On three normal-collected datasets (PlantVillage, CottonWeedID15, and PotatoLeafDisease) the model attains accuracies of 99.74%, 91.81%, and 99.52%, respectively. On three controlled-collected datasets (PlantSeedling, RiceImage, and SoyBeanSeed) the corresponding accuracies are 97.45%, 99.97%, and 93.08%. CTNet consistently outperforms state-of-the-art classification methods across all benchmarks, demonstrating its effectiveness as a practical solution for agricultural image classification tasks.

Accurate quantification of rice aboveground biomass (AGB) is critical for crop monitoring but remains challenging due to the complex nonlinearity arising from the coupling of plant density, spatial structure, and internal dry matter distribution. To address the limitations of single-source remote sensing, this study proposed a Three-Dimensional Dry Matter Distribution Integration (3D-DMI) model, which establishes a physically interpretable framework decomposing AGB into dry matter density (ρ), horizontal projection distribution (S), and vertical cumulative distribution (hd) components. Guided by this framework, a core subset of six features (Red_650, MTCI, G_correlation, R_correlation, LPI, and HPA0_99) was extracted from UAV-based multispectral, RGB, and LiDAR data using a dual-step feature selection approach combining Maximum Information Coefficient (MIC) and Distance Correlation (dCor). A Random Forest (RF) regression model was then developed to estimate AGB across the entire growth season. The results demonstrated that the 3D-DMI model achieved excellent performance with an R2 of 0.920, an RMSE of 0.184 kg/m², and an RPD of 3.544, significantly outperforming any single-sensor approach. Single-feature analysis revealed that while LiDAR-derived structural features provided the fundamental basis for biomass estimation, they encountered inherent saturation bottlenecks during late growth stages. Feature contribution analysis based on SHAP further quantified that LiDAR-derived features dominated the estimation process (68.5% contribution), providing the volumetric basis, whereas RGB textures (18.3%) and multispectral features (13.3%) provided indispensable supplements. Ultimately, this study established a robust, physically grounded computational paradigm for high-precision UAV-based rice biomass monitoring across the entire growth cycle.