Wheat seedling counting is one of the most crucial field surveys at the emergence stage. However, the challenges remain, such as small size, dense distribution, and mutual occlusion of wheat seedlings. This study aims to accurately detect and count the wheat seedlings in complex field environments. An improved algorithm (named DSM-YOLOv10) was proposed using YOLOv10n. Firstly, global annotation was adopted rather than only local features during data labeling. The overall features of wheat seedlings were learned and extracted to effectively avoid the feature loss caused by overlapping and occlusion. Secondly, a DCC2f module was constructed to replace the C2f module with the dual convolution (DualConv) in the backbone network. Grouped and pointwise convolutions were concurrently deployed to combine with the residual connections. The gradient vanishing was alleviated in deep-layer training. A more complete transmission of cross-layer features was realized to minimize the redundant convolution. Computational efficiency was improved to facilitate subsequent real-time deployment on mobile edge devices. More efficient features were provided for the subsequent processing. Furthermore, a semantics and detail infusion (SDI) module was introduced to explicitly inject the high-level semantic information into low-level detail features using cross-layer attention and Hadamard products. Detail enhancement was obtained under semantic guidance to avoid simple feature stacking without deep information interaction in the conventional concatenation modules. Consequently, the performance was improved to decouple and reuse features in the overlapping regions of wheat seedlings. Finally, a multi-scale dilated attention (MSDA) mechanism was adopted to incorporate the multiple dilation rates. Multi-scale semantic information was effectively aggregated to integrate the local details with contextual information. Furthermore, the intersecting and overlapping wheat seedlings were enhanced after extraction. Moreover, the complex redundancy or additional computational cost was efficiently reduced in the self-attention mechanism. Experimental results demonstrate that the DSM-YOLOv10 model achieved precise detection of wheat seedlings under multiple complex scenarios. The mean average precision (mAP), precision (P), recall (R), and F1 score were 91.4%, 85.2%, 81.7%, and 83.4%, respectively, which was improved by 5.0, 2.6, 5.4, and 4.1 percentage points, respectively, compared with the original YOLOv10n model. Furthermore, the parameter count and floating point operations (FLOPs) were reduced by 4.7% and 10.7%, respectively. The real-time detection was also realized with an inference time of 13.2 ms (approximately 76 frames per second). Compared with the RetinaNet, Faster-RCNN, SSD, and YOLOv8n models, the DSM-YOLOv10 model exhibited the best performance in detection. The mAP values were 35.8, 24.5, 30.7, and 8.6 percentage points higher, respectively; whereas, the number of parameters was reduced by 87.4%, 96.2%, 90.0%, and 15.2%, respectively; and the FLOPs were decreased by 86.3%, 96.0%, 88.1%, and 7.4%, respectively. The DSM-YOLOv10 model also demonstrated excellent recognition of the small objects, with a mAP of 86.3%, particularly in overlapping seedling scenarios. Both its missed detection rate and false detection rate remained at the low levels, at 9.1% and 3.8%, respectively. The coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) were 0.92, 5.68 plants, and 4.33 plants, respectively, in the seedling counting task using the DSM-YOLOv10 model. Compared with the YOLOv10n, the R² increased by 6.98%, while the RMSE and MAE decreased by 35.23% and 40.44%, respectively, indicating the higher counting accuracy and robustness. This finding can effectively improve the detection and counting performance of wheat seedlings in complex scenarios. The improved model was also featured by fewer parameters and lower FLOPs. The finding can provide support for field data acquisition in smart agriculture.

Leaf area index (LAI) is one of the key parameters to describe canopy development and growth status in apple orchards. In this study, a multimodal unmanned aerial vehicle (UAV) framework was developed for LAI estimation in the three phenological stages of apple trees, including the end-flowering, spring shoot growth, and fruit expansion. Multispectral imagery and light detection and ranging (LiDAR) point cloud data were synchronously acquired during ground measurements. A feature dataset was then fused with spectral, texture, and three-dimensional canopy structural information. Specifically, 15 spectral indices, 40 texture features, and 27 structural descriptors were extracted to characterize canopy reflectance, spatial heterogeneity, and geometric architecture, respectively. Seven regression models were evaluated, including Backpropagation neural network (BPNN), Deep neural network (DNN), Gaussian process regression (GPR), Random forest (RF), Ridge regression (RIDGE), and Support vector regression (SVR). Model performance was assessed at each phenological stage after repeated random validation. The coefficient of determination, root mean square error, and mean absolute error were used as evaluation metrics. In addition, ablation experiments were conducted to quantify the individual and combined contributions of spectral, textural, and structural information. Shapley additive explanations were used to interpret feature importance at the growth stages. The results showed that the deep neural network achieved the best overall performance and the most stable accuracy among all models. The coefficient of determination reached 0.86, 0.87, and 0.89, respectively, at the end-flowering, spring shoot growth, and fruit expansion stage after validation. Deep Neural Network was highly effective for LAI estimation under different canopy conditions and phenological stages. Furthermore, the better performance was achieved in the lower fluctuations and higher robustness to the stage-dependent canopy. Ablation analysis further demonstrated that canopy structural information was the dominant source for LAI estimation in apple trees. When used alone, structural features consistently outperformed or matched the spectral and textural feature groups. Especially, there was a strong sensitivity of canopy geometry to leaf area changes during rapid vegetative development at the new shoot growth stage, with the coefficient of determination of 0.79. Meanwhile, the structural and textural information was fused for the best dual-source performance at the new shoot growth stage. The coefficient of determination increased to 0.84, whereas the root mean square error was reduced to 0.16. A similar pattern was observed during the fruit expansion stage. Specifically, the combination of structural and textural information also produced the strongest two-source performance, with a coefficient of determination of 0.85 and a root mean square error of 0.13. Once spectral, textural, and structural information were fully integrated, the estimation accuracy further improved, with the coefficients of determination of 0.86, 0.87, and 0.89 at the three stages, respectively, while the root mean square errors were 0.14, 0.14, and 0.12, respectively. The interpretation analysis also confirmed the dominant role of canopy structural information in all growth stages. Structural descriptors with canopy density and height distribution contributed most to prediction, whereas spectral and texture information served as complementary inputs to enhance model discrimination under complex canopy conditions. Overall, multimodal feature fusion with deep learning can also provide accurate and interpretable LAI estimation for apple orchards at phenological stages. The findings can offer strong support to accurately monitor the orchard canopy.

Eco-unmanned farms aim to achieve precise, intelligent, and sustainable production under unmanned operations, thus integrating ecological practices, information science, and agricultural engineering. However, the heterogeneous elements included in the complex systems are the soil, crops, meteorology, hydrology, and machinery. The dynamic interactions among these components exhibit strong nonlinearity and spatiotemporal variability. However, conventional approaches can rely mainly on remote sensing and static models. Digital twin technology can be expected to construct high-fidelity virtual replicas, and then map, simulate, and interact with physical farms in real time. This review was presented on the digital twin applications in the eco-unmanned farms, covering the frameworks, key technologies, core functionalities, and closed-loop mechanisms. The challenges were also proposed for the future directions. 1) A three-layer conceptual framework comprised of the physical entity layer, the virtual model layer, and the data interaction and service layer. a) The physical entity layer was involved in intelligent agricultural machinery (e.g., unmanned tractors, plant protection UAVs, and spraying robots), an air-space-ground integrated sensing network (including satellite remote sensing, UAV remote sensing, and ground-based IoT sensors), and farmland infrastructure. This layer served as both the data source and the execution terminal for the control commands. b) In the virtual model layer, a multi-scale model library was established to integrate the mechanistic models (e.g., crop growth and soil water-salt transport models), data-driven models (e.g., convolutional neural networks and machine learning regression), and hybrid models to combine physical principles with artificial intelligence. c) The data interaction and service layer were used for the seamless data flow and closed-loop decision-making using edge computing, cloud computing, standardized interfaces, and "What-If" scenario simulation. 2) Key technologies. The framework was used to construct the digital twins in eco-unmanned farms. The air-space-ground integrated sensing network captured the multi-source data from satellites, UAVs, and ground sensors. A three-dimensional perception network was selected to cover the entire farm at multiple scales. The model library was developed to integrate the mechanistic models with physical interpretability and data-driven models, indicating the high predictive accuracy. A hybrid modeling was realized to balance the precision and computational efficiency. 3) Core functionalities. a) "What-If" ecological simulation enabled the virtual experimentation of different strategies, such as irrigation scheduling, variable-rate fertilization, and pest control. Decision support was then provided to reduce the cost and risk of field trials. b) Visualization and interaction were used to enhance user immersion and operational efficiency, including 3D reconstruction using UAV LiDAR and multi-view photography, neural radiance fields, virtual dashboards, and VR/AR interfaces. c) Optimization and collaboration were used for the global scheduling of resources, equipment, and workflows after multiple objective optimization, prediction maintenance, and energy-carbon footprint. Digital twin applications were elevated from single-point optimization to farm-wide collaboration. 4) Closed-loop application mechanisms were further examined from the bidirectional driving and data feedback. The "virtual-to-physical" driving mechanism translated the optimal decisions from the virtual space into precise execution commands, such as variable-rate prescription maps and path planning instructions for unmanned machinery. Conversely, the "physical-to-virtual" driving mechanism was used to feed the real-time monitoring data back to the virtual model, enabling dynamic state and parameter calibration. Data feedback and model optimization mechanisms were used for the digital twins with continuous self-evolution, including data assimilation, online learning, and adaptive control strategies. Several key challenges were remained on the widespread adoption of digital twins in eco-unmanned farms, including the trade-off between model accuracy and computational complexity, the technical bottleneck of multi-source heterogeneous data fusion, the dilemma of real-time performance and reliability in complex field environments, and the absence of full-lifecycle model evolution mechanisms. Future research can also be expected to focus on four directions: frameworks, key technologies, functionalities, and closed-loop mechanisms. Digital twin technology can also revolutionize the intelligent and sustainable eco-unmanned farms.

Climate change, resource limitations, and increasing global food demand are accelerating the need for efficient and sustainable agricultural management practices. Unmanned aerial vehicles (UAVs) have emerged as a transformative technology in precision agriculture (PA) because of their capability to provide high-resolution, real-time, and site-specific crop monitoring. This review critically examines recent advancements (2016–2025) in UAV-assisted PA, focusing on UAV platforms, sensing technologies, data acquisition systems, information fusion methods, and artificial intelligence (AI)-driven analytical frameworks. Particular emphasis is placed on applications including crop monitoring, disease and pest detection, weed mapping, irrigation management, soil assessment, yield estimation, phenotyping, and precision spraying. The review highlights that integrating RGB, multispectral, hyperspectral, thermal, and LiDAR sensors with machine learning (ML) and deep learning (DL) algorithms substantially improves monitoring accuracy, operational efficiency, and agricultural decision-making compared with conventional practices. Algorithms such as Random Forest (RF), Support Vector Machine (SVM), convolutional neural networks (CNNs), and YOLO-based models have demonstrated strong effectiveness in yield prediction, disease recognition, and weed discrimination. Despite these advancements, several challenges continue to limit large-scale implementation, including restricted flight endurance, payload limitations, environmental sensitivity, data-processing complexity, interoperability issues, and limited AI model transferability across different agricultural environments. Furthermore, model performance remains highly dependent on sensor configuration, dataset quality, and field-specific environmental conditions. Recent developments indicate rapid commercialization of UAV technologies together with emerging trends in edge AI, explainable AI (XAI), UAV–IoT integration, cloud-based analytics, and autonomous multi-UAV systems. Overall, this review identifies major technological advancements, key operational limitations, and future research directions required to support scalable, reliable, and climate-resilient UAV-assisted agricultural systems.

Unmanned aerial vehicle (UAV) wind field airflow is the main factor affecting spraying width and operational effect. Under the soybean and maize compound planting mode, the width of the spray droplets of UAV aerial spraying is easily sprayed on other crops around the target crop by mistake due to the influence of the wind field. In order to solve this problem, a narrow-width spraying UAV equipped with an airflow guidance device was developed in this study, which can achieve precise spraying on target crops within a narrow operation width and reduce the impact of droplet drift on surrounding non-target crops. The wind field of the flight platform was simulated via simulation software, and the wind field distribution characteristics corresponding to three sizes of the guidance device were analyzed. It was verified that the guidance device has a segmentation effect on the wind field, and the optimal size of the guidance device was determined as 0.92 m accordingly. Meanwhile, the installation position and quantity of nozzles were determined to be set at the position with the minimum airflow disturbance. The wind speed at measuring points with different angles on four diameters at five heights under the UAV hovering state was tested through bench tests, to further verify the simulation results and appropriately adjust the nozzle position according to the measured wind speed. Outdoor flight spraying tests were carried out, and the results showed that when the 0.92 m guidance device was applied at a flight altitude of 1.00 m, the effective spray width was only 1.46 m according to the droplet deposition density on the coated paper of three test collection belts. Under the maize-soybean composite planting pattern, the field spraying test yielded results of 18.0%-30.0% deposition rate in the maize area and 0.1%-1.6% in the soybean area with maize as the operational target. A predictive effect of wind field simulation on the installation position of nozzles where droplets suffer the least wind field disturbance under aerial spraying conditions was confirmed. The UAV wind field can be effectively segmented by the airflow guidance device, and the diameter of the droplet-laden airflow column can be reduced, thus realizing narrow-width spraying.

Electrostatic spraying has been widely used to enhance the droplet deposition under plant protection using UAV (unmanned aerial vehicles). The plant protection droplets can also move through the intense rotor under downwash and turbulence, when the droplets can carry sufficient electric charge. However, the CMR (charge-to-mass ratio) can attenuate markedly before reaching the crop canopy. Field-scale quantification of CMR has been constrained to deploy over a wide spray footprint, due to the conventional rigid Faraday cylinder collectors. In this study, a flexible AFC (aluminum foil container) was developed to collect the charged droplets under field conditions. The CMR was then quantified in contact charging UAV electrostatic spraying. A coupled attenuation model was further used to interpret the relative contributions of the charge loss and droplet mass loss during flight. A high-voltage generator module was integrated into a quadrotor spray platform and then configured for contact charging. The negative high voltage output was connected to a metal contact electrode inside the liquid tank in charge of the working solution. While the positive terminal was mounted on the landing gear and discharged to air, providing for the weak capacitive coupling that supported charge balance during hovering. The AFC collector followed the Faraday cage principle. But the rigid collectors were replaced with a flexible aluminum foil conductive layer supported by an insulating backing, allowing for the collector geometry to match the spray swath. Repeatability tests under identical spraying showed that the consistent CMR was comparable to a conventional Faraday cylinder in the field deployment. Field experiments were conducted at the charging voltages from 15 to 35 kV and the flight heights from 2.0 m to 5.0 m. Two spray adjuvants were evaluated at the same recommended concentration of 1‰, a surfactant-based product. Momentive Agrospred 910 and a modified plant oil product, Maifei, were used in the field test. Baseline CMR increased with the voltage and then reached 3.50 mC/kg at 35 kV. Field measurements indicated that there was an outstanding CMR loss in the rotor downwash, indicating the height effect. Once the flight height increased from 2.0 m to 5.0 m, the average CMR attenuation rates were 14.09 %, 24.23 %, 39.32 %, and 52.84% at 2.0, 3.0, 4.0, and 5.0 m, respectively. Rotor wind speed during measurements remained close to 14m/s. A shared background served as in the treatments. Model fitting showed that the exponential charge decay dominated the CMR attenuation from 2.0 m to 5.0 m, while there was a weak evaporation-driven mass loss. The charge attenuation coefficient, λ, was fitted approximately 0.16 at 35 kV, whereas the mass loss coefficient, K' prime, was approximately 4.0×10^-4, corresponding to less than 0.003 correction within 5.0 m. The λ increased from 0.08 to 0.16 over the voltages, as the voltage rose from 15 to 35 kV, with the intermediate values of 0.10, 0.12, and 0.14 at 20, 25, and 30 kV. Despite the high decay per unit distance at the higher initial charge, the higher voltage still produced the higher retained CMR in the typical operating heights. Both adjuvants increased baseline CMR for the retained CMR under UAV operation. Momentive increased the baseline CMR by about 17.71% at 35 kV, with the maximum of 4.12 mC/kg, whereas Maifei increased the baseline CMR by about 45.43 % with the maximum of 5.09 mC/kg. Attenuation rates were broadly comparable with/without adjuvants. The primary benefit of the initial CMR of the more charge remained after in-flight decay. Overall, the droplet charging model was established for contact charging, while the CMR attenuation model was for UAV electrostatic spraying. The AFC effectively captured the CMR variations dominated by flight height and rotor wind in the UAV crop protection spraying. The CMR can play an enhancing role in spray adjuvants. These findings can provide an experimental reference to optimize the UAV electrostatic spraying.

Electrostatic spraying is expected to integrate with unmanned aerial vehicle (UAV) platforms in modern agriculture. This approach represents an innovative plant protection technology, offering promising potential for improving pesticide utilization efficiency, deposition uniformity in precision agriculture. Electrostatic spraying imparts electrostatic charges to droplets through the application of high-voltage electric fields, thereby modifying their trajectories and enhancing adhesion to plant surfaces. This is particularly beneficial for achieving deposition of the abaxial (underside) surfaces of leaves, which are typically difficult to reach using conventional spraying techniques. This review aims to examine the development history, theoretical foundations, and current research on electrostatic spraying for crop protection UAVs. It systematically outlines the three mainstream charging mechanisms—corona, inductive, and contact charging—highlighting their physical principles, advantages, limitations, and the applicable voltage ranges. Inductive charging dominates current applications due to its relative safety and engineering simplicity. However, contact charging also presents distinct advantages over other mechanisms. The droplet charging process is relatively mild, avoiding intense discharges that may occur in corona charging systems. Since the sprayed liquid comes into direct contact with the electrode, a stable charge-transfer field is established, leading to more sufficient and uniform charging of droplets. In addition, the structural designs of inductive charging nozzles are discussed, including various electrode configurations (e.g., ring-type, cone-type, and embedded parallel plates), selection of electrode materials (such as copper, nickel, and stainless steel), and the integration of air-assisted mechanisms. Furthermore, key evaluation techniques are reviewed, including charge-to-mass ratio (CMR) measurements, droplet size characterization (e.g., volume median diameter, VMD), and deposition detection methods. However, standardized testing protocols remain lacking, and significant discrepancies persist between laboratory measurements and field performance—especially under complex field conditions involving wind, temperature, and humidity variations. The droplet behavior of electrostatic spraying is further analyzed through droplet trajectory modeling, with emphasis on three dominant electrostatic field interactions: 1) induced fields between droplets and plant targets, 2) repulsive fields among charged droplets, and 3) externally applied fields between the nozzle and the target. Each mechanism contributes differently to droplet motion, distribution, and deposition efficiency. Evidence from recent studies suggests that combining electrostatic spraying systems with UAV platforms can effectively improve spray characteristics, such as deposition density, spray width, total deposition, and pest control performance. However, their trajectory control capability remains limited and challenging:1) rotor-generated airflow interferes with droplet trajectories, weakens electrostatic adhesion, and accelerates CMR decay; 2) variations in UAV flight height affect the electric field distribution between the nozzle and the crop canopy, reducing deposition accuracy; and 3) environmental factors such as wind, temperature, and humidity introduce uncertainties that compromise field performance. To address these limitations, several recommendations for future research are proposed: 1) advance high-voltage contact or corona charging systems with enhanced safety features; 2) develop electrostatically optimized-size nozzles and adjuvants; 3) refine evaluation metrics by integrating CMR with droplet size distribution; and 4) conduct large-scale, crop-specific field validations. With the advancement of UAV technology, high-voltage electrostatics, and system miniaturization, UAV electrostatic spraying is poised to become a key tool for next-generation precision pesticide application, offering strong potential for reducing pesticide use and supporting sustainable agriculture.

High accuracy and efficiency can often be required to assess the freeze damage of the winter wheat. However, the traditional field phenotypic identification can be confined to strong subjectivity, serious deviations between laboratory simulations and field data, and the lengthy cycle of physiological and biochemical detection. In this study, a freeze damage classification was proposed to fuse the vegetation indices, texture, and color features. A freeze damage model was also constructed to combine with the machine learning. The unmanned aerial vehicles (UAVs) were utilized to collect the multispectral and visible light imagery. Specifically, the UAVs were equipped with multispectral sensors and high-resolution visible light cameras. The optimal weather (clear skies, stable and light) was employed to ensure the data quality. A total of 15 vegetation indices were extracted, including the normalized difference vegetation index (NDVI), ratio vegetation index (RVI), excess blue index (ExB), and soil adjusted vegetation index (SAVI). The spectral reflectance was also selected from the specific wavelength bands in order to effectively represent the vegetation’s photosynthetic capacity and canopy health status. Furthermore, the gray-level co-occurrence matrix (GLCM) was employed in texture feature analysis. Some parameters were calculated to quantify the spatial distribution of the pixel gray levels in the imagery, such as the mean, variance, contrast, and homogeneity from the GLCM. These texture features were then captured for the subtle structural variations in the wheat canopy caused by freeze damage. Additionally, the color features were derived from the RGB color space of the visible light images. Statistical parameters were computed to represent the visual color of the winter wheat under freeze damage, such as the color mean, standard deviation, and color moment. Feature screening was conducted using the Pearson correlation coefficient (PCC) method and the minimum redundancy feature selection algorithm (mRMR). Firstly, the Pearson correlation analysis was used to measure the linear correlation between each feature and freeze damage levels, thus eliminating the features with a low correlation. Subsequently, the minimum redundancy feature selection was optimized the feature set, in order to balance the feature relevance and redundancy. The final features were selected for the informative and non-repetitive ones. Four classification models were then constructed: random forest (RF), support vector machine (SVM), backpropagation neural network (BPNN), and eXtreme gradient boosting (XGBoost). Each model was also trained with tunable parameters. For example, the XGBoost model was utilized to adjust the parameters, such as the learning rate, the number of estimators, and the maximum depth after cross-validation. The experimental results demonstrated that the multi-source feature fusion significantly improved the performance of the model. Among them, eXtreme Gradient Boosting (XGBoost) showed the optimal performance under the fusion of vegetation indices, texture, and color features. The accuracy, recall, precision, and F1-score reached 0.73, 0.68, 0.71, and 0.69, respectively. Further analysis revealed that the texture features shared the highest correlation with the manually surveyed levels of the freeze damage. The structural damages of the canopy were also sensitive to capture, as the freeze damage often disrupted the orderly arrangement of the wheat leaves and stems. The texture features were effectively detected after capture. The multi-source remote sensing data fusion was validated to assess the freeze damage. The spectral, texture, and color information were integrated to precisely monitor the freeze damage of the winter wheat for disaster prevention decision-making. Remote sensing and machine learning were also applied to assess agricultural disasters in smart agriculture. The finding can serve as a practical reference for crop stress monitoring in the disaster early warning of intelligent agriculture.

Here, the ZAL-YOLOv8n model was proposed to rapidly and accurately detect the apple appearance quality in natural scenes using the improved YOLOv8n model. The baseline model was selected as the YOLOv8n model, the lightest model in the YOLOv8 series. Subsequently, the improved model was deployed on the mobile edge devices, despite its relatively lower accuracy of recognition. The apple appearance quality was classified into three categories: Immature, Mature, and Fall ill, according to the influencing factors, such as the fruit maturity and health. When the fruit with the greenish-blue peel was marked as “Immature”. Once the peel was red or a mix of red and yellow, the fruit was marked as “Mature”. If there were disease spots on the peel, the fruit was marked as “Fall ill”. Two common apple diseases, namely “Ring rot” and “Scab”, were further identified from the disease category. Firstly, the partial convolution (PConv) and the efficient multi-scale attention mechanism (EMA) were integrated as the EP-C2f module, in order to replace the C2f module in the backbone network. The EMA attention mechanism was employed to redistribute the attention weights, and then group all channels with the same number of channels. Effective information on all channels was retained to realize the interaction between channel and spatial position. Thereby the target area was focused on enhancing the feature extraction for apples under complex occlusion. Simultaneously, the dynamic convolution of the PConv was used to filter out the effective feature data. The model size and parameters were reduced after extraction. Secondly, the MPDIoU was introduced as the boundary regression loss function, in order to realize the accurate positioning of apple skin lesion areas in diseased apples. The reason was that the original boundary box regression loss function (CIoU) of the YOLOv8 model failed to realize the rapid fitting, where the types and degrees of pathogen infections varied among diseased apples. The new position fitting was accelerated to distinguish and locate the lesion edges, in order to minimize the distance between the upper left and lower right corners of the predicted and the real box. Finally, the Slim-neck architecture was utilized to reconstruct the feature fusion network of the YOLOv8 model. The lightweight of the neck network was achieved to enhance the operating speed. The experimental results indicate that the accuracy, recall, and mean average precision of the ZAL-YOLOv8n model increased by 3.4, 1.1, and 1.3 percentage points, respectively. Meanwhile, the floating-point operations, parameter quantity, and model size were reduced by 22.2%, 17.7%, and 15.9%, respectively, fully meeting the deployment requirements for the mobile edge devices. The ZAL-YOLOv8n model was performed on the highly precise detection and quality identification of the apples with different appearance qualities in natural scenes. The high degree of lightweight was also achieved to balance between the accuracy and speed, in order to realize the real-time detection on the low-computing-power devices. Therefore, the improved model can also provide technical support to the research and development of intelligent robots for apple picking.

Green lawns can play a crucial role in the landscape and ecological environments with the rapid urbanization in China. However, there is an ever-increasing demand for the daily maintenance of lawns. Among them, the heterogeneous weeds can often compete with the native grass for nutrients and growing space in lawns, leading to the low overall quality of urban greenery. The aesthetic appeal of the lawn can also be diminished after premature aging. Thereby, the accurate and rapid detection of the weed can be required for deep learning and computer vision. This study aims to improve the efficiency and accuracy of the detection of heterogeneous weeds in natural environments. A lightweight algorithm was also proposed using the original YOLOv8n model. Firstly, the deformable convolutional network v2 (DCNv2) was employed to combine with the c2f convolutional layers in the backbone network. The offsets were then introduced to enhance the feature extraction of the improved model. Different regular shapes of the heterogeneous weeds were captured after optimization. Additionally, a modulation mechanism was incorporated to control the contribution of each sampling point to the output. The precision and robustness were then improved to focus more on the target regions, thereby reducing the interference from background noise. Secondly, a bidirectional feature pyramid network (BiFPN) was introduced in the neck network. Multi-scale features were efficiently fused using bidirectional cross-scale connections and weighted feature fusion. The targets were detected at varying scales. Furthermore, the BiFPN significantly reduced the computational overhead, compared with the traditional feature pyramid network (FPN). The redundant connections were also eliminated to incorporate the lightweight weighting mechanisms. The efficiency and generalization of feature fusion were also improved after feature fusion. Lastly, the traditional intersection over union (IoU) loss function was replaced with the inner-IoU (inner intersection over union) loss function. The objective function of the bounding box regression was then optimized to learn the target location information, thereby improving the convergence speed and detection performance. More precise guidance was realized for the improved model. The experimental results show that the improved YOLOv8-LDB model reduced the number of parameters, the computational cost, and the model size by 32.7%, 13.6%, and 31.8%, respectively, compared with the original YOLOv8n. While the mean average precision (mAP) increased by 3.2 percentage points. The performance of the improved YOLOv8-LDB model was better than that of the seven commonly used network models (including Faster-RCNN, SSD, YOLOv5s, YOLOv5n, YOLOv7n, YOLOv10n, and YOLOv11n), in terms of the precision, parameter count, computational cost, and model size. The mean average precision was improved by 21.6, 9.6, 1.6, 7.7, 2.4, 4.1, and 3.3 percentage points, respectively. Additionally, the detection speed increased from 80.4 to 87.1 frames per second. The inference efficiency was also enhanced for the real-time detection of heterogeneous weeds in natural environments. The YOLOv8-LDB algorithm demonstrated superior performance across multiple metrics. Automatic sprayers and weeding robots can be integrated to realize variable-rate precision spraying and targeted weed control. The automation level of lawn management can also be further advanced in the future. The lightweight improved model can be expected to find wide applications in smart city greening and precision agriculture. The findings can also provide technical support to protect the sustainable ecological environment in urban areas.