Design and Experimental Evaluation of an Artificial Intelligence-Guided Airflow Pollination System for Tomato Flowers
Abstract:
Pollination limitations in protected tomato cultivation can reduce fruit set when natural pollinators are insufficient, while manual pollination is labor-intensive. An artificial intelligence-guided, non-contact pollination system was therefore developed to automatically localize tomato flowers and deliver targeted airflow using a two-dimensional motion platform equipped with a fan. A You Only Look Once version 8 nano (YOLOv8n) object detector was trained using 710 annotated images acquired with an ESP32-based camera module (ESP32-CAM) under greenhouse and open-field conditions, with the dataset partitioned into training and validation subsets at a 90:10 ratio. On the validation subset, precision, recall, F1-score, and mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) were 0.841, 0.847, 0.84, and 0.906, respectively. Inference rates of 6.5 frames per second on a personal computer and 3.5 frames per second using the ESP32-CAM wireless image stream were achieved. Following planar calibration at a working distance of 50 cm, flower coordinates were estimated with errors of approximately ±2 mm in the horizontal direction and ±5 mm in the vertical direction under the calibrated experimental configuration. When integrated with the two-dimensional motion platform, the system achieved a mean operational success rate of 83.68 ± 6.07% across 20 plants per treatment. The mean processing time per flower was 4.46 ± 0.62 s for automated airflow pollination, compared with 4.14 ± 0.59 s for manual tapping, with no statistically significant difference detected between the two methods ($p$ = 0.097). The prototype was constructed at an approximate cost of USD 167. These findings demonstrate the technical feasibility of integrating vision-based flower localization with targeted airflow delivery for automated tomato pollination and indicate the potential of this low-cost approach for reducing dependence on labor-intensive pollination practices in resource-constrained protected cultivation.
1. Introduction
Protected cultivation is expanding in many regions because it improves crop control and reduces exposure to rainfall, pests, and seasonal shocks. In practice, however, greenhouse and high-tunnel tomato production still depends on reliable pollination at flowering. When flowers are poorly pollinated, fruit set and quality can drop even if water and nutrients are well managed [1], [2], [3]. This is relevant for expanding urban and peri-urban vegetable systems, including in Central Africa [4].
Tomato is largely self-fertile, but pollen release from the anther cone is strongly helped by vibration. This is why bumblebees are effective pollinators and why manual tapping or vibration of flower trusses is common in greenhouses [1], [5], [6]. Manual techniques can work well, but they require repeated labor and become difficult to scale when flowering is dense or labor availability is limited [7].
Several non-contact alternatives have been explored. Airflow-based approaches can dislodge pollen without touching the flower; early work using pulsating air jets demonstrated that targeted airflow can improve tomato pollination [8]. More recent studies have also shown that acoustic or sonic stimulation can trigger self-pollination mechanisms in tomatoes and influence fruit traits [9], [10], [11].
Autonomous pollination is now an active robotics topic, with prototypes ranging from drones to ground platforms that combine perception, positioning, and a pollination mechanism [12], [13]. Many of these systems rely on computer vision for flower detection and localization, which remains a practical bottleneck under occlusion and variable lighting [14], [15]. At the same time, there is ongoing debate about whether robotic systems should be viewed as replacements for biological pollinators or as complementary tools, especially in protected cultivation where pollinator availability can be constrained [16].
This study focuses on a simple, low-cost option for greenhouse conditions: artificial intelligence-guided targeted airflow delivery. An integrated prototype was developed and tested which (i) localizes tomato flowers with a lightweight You Only Look Once (YOLO)-based detector, (ii) estimates planar coordinates for positioning, and (iii) applies short airflow puffs using a fan mounted on a two-dimensional motion platform. The goal is not to replace biological pollination where it is feasible but to provide a practical alternative where manual labor or managed pollinators are limiting.
The main contributions are (i) a low-cost integrated vision-motion-airflow prototype for targeted ventilated pollination; (ii) a locally collected tomato-flower dataset (710 annotated images) split into training and validation subsets at a 90:10 ratio; and (iii) a performance evaluation covering detection accuracy, real-time inference speed on a personal computer and via the ESP32-based camera module (ESP32-CAM) wireless image pipeline, planar localization error at a 50 cm calibration depth, and operational pollination performance compared with manual tapping.
2. Related Work
Tomato flowers are “buzz-pollinated”: pollen is released more effectively when the anther cone is vibrated. This underpins both bee-based pollination and common manual techniques [1], [5]. In greenhouses, managed bumblebee pollination can improve fruit set and fruit characteristics [6]. However, recurring costs and management constraints can limit its use in some production contexts [7].
Mechanical methods remain relevant, including non-contact options such as airflow and acoustic stimulation [8], [9]. Robotic pollination research typically couples three elements: perception (to locate flowers), motion/positioning (to reach them), and a pollination mechanism. Recent surveys outline progress and open challenges, particularly the difficulty of robust perception in cluttered canopies and the cost and complexity of reliable positioning [12], [13].
The bio-inspired and agricultural robotics literature also emphasizes that field-ready systems must balance robustness, maintainability, and affordability [17] [18]. A greenhouse tomato case study using a drone-based pollination platform illustrates the feasibility of artificial intelligence-triggered pollination actions [19].
On the perception side, deep learning object detectors are increasingly used in agriculture, and recent reviews report rapid growth of YOLO-style approaches for detection tasks under deployment constraints [14-15]. For flower-specific applications, precision pollination studies have reported flower detection and localization pipelines using both one-stage detectors and instance segmentation. For example, Mask Region-based Convolutional Neural Network (Mask R-CNN) has been used for flower detection and identification for precision pollination in orchards [20].
Tomato-focused studies have combined deep-learning-based flower detection with visual servoing and three-dimensional flower localization to support robotic pollination [21], [22]. The present work aligns with this line of research but targets a simpler mechanism and a low-cost build. Instead of contact-based vibration tools, it uses short airflow puffs delivered after centering on the flower, and it evaluates the end-to-end system with operational success criteria and time-per-flower comparisons. This emphasis reflects the practical constraints that often determine whether a prototype can be adopted in small and medium greenhouse operations.
Recent robotic pollination systems differ considerably in their perception methods, positioning mechanisms and approaches to flower stimulation. Some systems prioritize three-dimensional mobility and precise visual servoing, while others emphasize mechanical simplicity or increased operating speed. Table 1 compares the proposed prototype with selected recent pollination systems in terms of crop application, detection and localization method, positioning hardware, pollination mechanism and reported hardware cost.
| Study | Crop and Evaluation Setting | Detection and Localization Approach | Positioning Hardware | Pollination Mechanism | Reported Hardware Cost |
|---|---|---|---|---|---|
| Hiraguri et al. [19] | Tomato; greenhouse field experiment | Convolutional neural network-based classifier for identifying pollinable flower stages; motion-capture positioning and robot operating system-based flight control | Small autonomous drone | Contact-based vibrator mounted on the drone | Not reported; the implementation cost was described as higher than conventional methods |
| Singh et al. [21] | Tomato; laboratory experiments using three-dimensional printed flower plants | You Only Look Once version 8 (YOLOv8)-based flower orientation detection, Intel RealSense D435 red--green--blue and depth (RGB-D) camera and visual servoing | Clearpath Husky mobile base and Universal Robots UR5 six-degree-of-freedom arm | Motor-driven vibratory brush | Not reported |
| Ahmad et al. [23] | Watermelon; open-field experiments | Deep learning-based estimation of flower size and orientation, followed by size-based depth estimation and visual servoing | Six-degree-of-freedom robotic arm, three-dimensional camera and linear actuator | Linear-actuator-based pollination end-effector | Not reported |
| Masuda et al. [24] | Tomato and other self-pollinated plants; greenhouse field tests | Perspective and close-up cameras; flowers were manually designated during the reported field tests | Suspended rail robot with a flexible multi-degree-of-freedom manipulator | Non-contact airflow supplied by a blower | Not reported |
| Present study | Tomato; greenhouse experiments | You Only Look Once version 8 nano (YOLOv8n) flower detection and single-camera planar coordinate conversion at a 50 cm calibration depth | Low-cost two-dimensional belt-driven motion stage | Non-contact targeted airflow from a 12 V fan | USD 167 |
Table 1 shows that the proposed system occupies a different design position from recent mobile and manipulator-based pollination systems. Drone-based and robotic-arm systems provide greater three-dimensional mobility and can approach flowers at different orientations and depths. However, they also depend on comparatively complex hardware, including motion-capture systems, mobile bases, six-degree-of-freedom arms, depth cameras or flexible manipulators. None of the selected studies reported a complete prototype cost, which limits direct monetary comparison.
In contrast, the present prototype had a documented build cost of USD 167 and was constructed using a two-dimensional belt-driven stage, an ESP32-CAM and a small fan. Its main advantages are therefore affordability, mechanical simplicity and non-contact operation. These advantages are accompanied by a smaller working envelope and dependence on calibration at a fixed camera-to-flower distance.
3. Materials and Methods
An experimental design was adopted combining engineering development and controlled performance evaluation. Tomato plants (Solanum lycopersicum L., ‘Radia’ variety) were cultivated for data collection and comparative testing. In total, 120 plants were used: 70 plants for acquiring images for model training, 10 plants for system testing and calibration, and 40 plants for comparative trials between automated and manual pollination. The 40 evaluation plants were divided equally between the two treatments, with 20 plants assigned to manual pollination and 20 plants assigned to automated pollination.
The system comprises four main subsystems: an ESP32-CAM for image capture, a YOLOv8n-based detector to identify flowers and estimate their pixel coordinates, a two-dimensional motion platform driven by National Electrical Manufacturers Association (NEMA) 17 stepper motors (42 mm frame size) and GT2 timing belt-pulley transmissions to position an end-effector carriage, and a 12 V direct-current fan mounted on the carriage to deliver localized airflow puffs for pollination. Figure 1 presents the computer-aided design model of the pollination system and highlights the main components used in the prototype. The prototype was designed for portability and constructed with locally available materials, resulting in a total build cost of approximately USD 167.

Figure 2 summarizes the end-to-end workflow of the proposed artificial intelligence-guided ventilated pollination platform, from image capture and flower detection to positioning and airflow delivery.

Images were captured during the flowering phase in both greenhouse settings and an open tomato farm to introduce variability in lighting and background. An ESP32-CAM was used at multiple resolutions (1280 × 720, 800 × 600 and 640 × 480 pixels). Viable images with clearly visible flowers were selected, and all flowers were annotated as a single class (“Flower”) using Label Studio. In total, 710 images were labeled and exported in YOLO format. Each image contained between one and four labeled open flowers, with most images containing two or three annotated flower instances. Consequently, the total number of labeled flower instances was greater than the number of images in the dataset. The annotated dataset was split programmatically at the image level into training and validation subsets at a 90:10 ratio using a Python script. Plant identity was not used to constrain the split, and therefore the training and validation subsets were not explicitly made plant-disjoint. Of the 710 annotated images, 639 images were used for model fitting and 71 images were used for validation, early stopping, model selection and quantitative performance reporting. After training, the selected model was additionally evaluated on real-time tomato foliage images and video to examine its operational behavior under natural canopy motion, occlusion, variable illumination and complex backgrounds. Table 2 summarizes the dataset composition and the train/validation split used for model development and evaluation.
Subset | Images | Purpose |
|---|---|---|
Training | 90% of 710 (639) | Model fitting |
Validation | 10% of 710 (71) | Validation, early stopping, model selection and quantitative metric reporting |
Training was performed in Google Colab using a Python 3 runtime and an NVIDIA Tesla T4 graphics processing unit with 15,360 MiB of memory. The Ultralytics 8.3.159 implementation of You Only Look Once version 8 (YOLOv8) was used to fine-tune pretrained YOLOv8n weights (yolov8n.pt). The model was trained for a maximum of 200 epochs using an input image size of 640 × 640 pixels, a batch size of 16 and early-stopping patience of 50 epochs. Training stopped after 127 epochs. The optimizer was set to automatic selection and was resolved by Ultralytics to AdamW, with an initial learning rate of 0.002, $\beta_1=0.9$, $\beta_2=0.999$ and weight decay of 0.0005. A linear learning-rate schedule with a final learning-rate factor of 0.01 and a three-epoch warm-up period was used. Automatic mixed-precision training was enabled, and reproducibility settings included a random seed of 0 and deterministic execution. Training augmentation included hue, saturation and brightness variations of 0.015, 0.7 and 0.4, respectively; translation of 0.1; scaling of 0.5; horizontal flipping with a probability of 0.5; and mosaic augmentation with a probability of 1.0. Rotation, vertical flipping, shear, perspective transformation, MixUp, CutMix and copy-paste augmentation were disabled. The best-performing weights (best.pt) based on validation metrics were saved and selected for inference. Table 3 lists the YOLOv8n training configuration reported in this study.
| Setting | Value |
|---|---|
| Training environment | Google Colab, Python 3 |
| Hardware | NVIDIA Tesla T4 graphics processing unit, 15,360 MiB |
| Software | Ultralytics 8.3.159 |
| Model | YOLOv8n |
| Initialization | Pretrained yolov8n.pt weights |
| Task/classes | Object detection; one class, Flower |
| Input image size | 640 $\times$ 640 pixels |
| Maximum epochs | 200 |
| Completed epochs | 127, following early stopping |
| Batch size | 16 |
| Early-stopping patience | 50 epochs |
| Optimizer | AdamW, automatically selected |
| Initial learning rate | 0.002 |
| Final learning-rate factor | 0.01 |
| Learning-rate schedule | Linear |
| AdamW parameters | $\beta_1 = 0.9$; $\beta_2 = 0.999$ |
| Weight decay | 0.0005 |
| Warm-up | 3 epochs; momentum = 0.8 |
| Random seed | 0 |
| Deterministic mode | Enabled |
| Mixed precision | Enabled |
| Active augmentations | Hue, saturation, and value: 0.015/0.7/0.4; translation: 0.1; scale: 0.5; horizontal flip: 0.5; mosaic: 1.0 |
| Disabled augmentations | Rotation, vertical flip, shear, perspective, MixUp, CutMix, and copy-paste |
| Workers/cache | 8 workers; cache disabled |
| Validation intersection over union/maximum detections | 0.7/300 |
To convert detections from pixel coordinates to real-world distances, a planar calibration approach was implemented. The camera was held at a fixed position, and a reference plane was established at 50 cm from the camera, perpendicular to the optical axis. Objects with known dimensions were placed on this plane, allowing derivation of a pixel-to-millimeter scale factor. During inference, the scale factor was applied to estimate horizontal and vertical offsets for positioning the fan relative to detected flowers. Figure 3 illustrates the planar calibration concept used to convert pixel measurements into millimeter offsets at a fixed camera-to-plane distance of 50 cm. The resulting pixel-to-millimeter scale factor was specific to the camera configuration used during calibration. Recalibration is therefore required whenever the camera-to-plane distance, image resolution, camera orientation, lens setting, or mounting geometry is changed. A change in working distance alters the apparent size of objects in pixels and consequently changes the millimeter-per-pixel scale. Similarly, tilting or rotating the camera relative to the working plane introduces perspective effects that cannot be represented accurately by a single uniform scale factor. If the camera is translated laterally while its distance and orientation remain unchanged, the scale factor may remain approximately valid, but the coordinate origin relative to the motion platform must be re-established. During practical operation, the camera should therefore be rigidly mounted and maintained at the calibrated distance of 50 cm.

The two-dimensional displacement platform was built on a rigid wooden frame with linear bearings on round iron guide rails. Each axis was actuated with NEMA 17 stepper motors; power transmission used GT2 belts and pulleys to obtain fine linear resolution. Motion commands were generated from the estimated flower coordinates and executed via an ESP32 controller interfaced with A4988 (stepper motor driver module) drivers. Based on motor and pulley specifications (200 steps per revolution; 40 mm travel per revolution), the nominal mechanical calibration constant was 5 steps per millimeter. The fan was switched by the controller and activated once the end-effector reached the target coordinate.
Model performance was assessed using standard detection metrics (precision, recall, F1-score, and mean average precision). Practical inference speeds were measured with images from a personal computer webcam and from the ESP32-CAM transferred via Wi-Fi. In both cases, YOLOv8n inference was executed on the personal computer; the ESP32-CAM was used only for image acquisition and wireless image transmission. In addition, the detector was exercised on real-time foliage footage to verify stability under live greenhouse conditions. Coordinate conversion was evaluated by comparing estimated distances to measured distances across camera depths. The motion platform was evaluated by commanding known displacements along the vertical and horizontal axes and comparing the commanded distances with the measured distances travelled by the end-effector carriage. Each operating condition was tested over five runs, and the mean travel time was calculated from these repetitions. Tests were conducted under two acceleration and maximum-speed settings to assess positioning accuracy and movement consistency and to verify the absence of observable step loss. Finally, pollination trials compared automated pollination to manual pollination with respect to (i) operational success and (ii) time per flower. For each plant in the automated treatment, operational success was calculated as the number of flowers for which the end-effector successfully centered on the target and applied an airflow puff, divided by the total number of flowers assessed on that plant. The overall automated success rate was calculated as the mean of the 20 plant-level percentages. For manual pollination, success was defined as gently tapping each target flower; under this definition, the manual baseline for the tested block was 100\% because no flowers were missed. For the time per flower, for each treatment the time required to pollinate each plant was recorded. A mean time per flower was then calculated for each plant, and the resulting 20 plant-level mean times per treatment were used for the statistical comparison. A one-sample t-test was used to compare the automated plant-level success percentages with the 100\% reference value, while an independent-samples $t$-test was used to compare plant-level mean time per flower between the manual and automated treatments.
4. Results
Training stopped after 127 epochs following the early-stopping criterion. On the validation subset, the YOLOv8n detector achieved a precision of 0.841, a recall of 0.847, an F1-score of 0.84 and a mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) of 0.906, indicating reliable identification of tomato flowers in cluttered canopies. Inference operated at approximately 6.5 frames per second on a personal computer webcam feed and 3.5 frames per second on the ESP32-CAM wireless feed. Figure 4 shows the precision and recall curves for the YOLOv8n tomato-flower detector on the validation subset during training.

Figure 5 shows how the F1-score varies with the confidence threshold, which helps justify the chosen operating point for inference.

Figure 6 reports the mAP@0.5 and mean average precision averaged over intersection-over-union thresholds from 0.5 to 0.95 (mAP@0.5–0.95) curves for the tomato-flower detector on the validation subset, indicating convergence of detection performance.

To complement the summary metrics above, Figure 7 and Figure 8 present the raw and normalized confusion matrices obtained from the saved validation results. The raw confusion matrix contained 162 true-positive detections, 27 false-negative detections and 28 false-positive detections. The 71-image validation subset therefore contained 189 labeled flower instances, corresponding to the sum of the 162 correctly detected flowers and 27 missed flowers. This number is greater than the number of validation images because each image contained between one and four labeled open flowers, with most images containing two or three. The 28 false-positive detections represent additional predictions made on background regions and are not included in the number of labeled flower instances. At the operating threshold used to generate the matrix, this corresponds to a precision of approximately 85.3% and a recall of approximately 85.7%. The normalized matrix indicates that approximately 86% of the annotated flower instances were correctly identified, whereas approximately 14% were assigned to the background class.


A qualitative examination of the saved validation predictions was conducted to identify cases in which the detector did not perform as expected. Figure 9 compares the ground-truth annotations with the corresponding model predictions for three representative examples.
In the first example, two flowers were positioned close together and partially overlapped in the image. The model detected only one of the annotated flowers, indicating that close spacing and overlap can cause adjacent flower instances to be merged or missed. In the second example, one of two neighboring flowers was not detected, while the remaining flower was assigned a relatively low confidence score. The image contained uneven illumination and a visually complex background, which reduced the visual distinction between the flower and surrounding plant structures. In the third example, the model correctly detected the annotated flower but also classified a background region as a flower, producing a false-positive prediction. These examples show that the principal detection errors were associated with closely spaced flowers, partial occlusion, uneven illumination and visually complex canopy backgrounds. They also demonstrate that the aggregate performance metrics do not fully describe the errors that may occur during practical greenhouse operation.

Under the calibration settings of a 50 cm camera-to-plane distance and a resolution of 640 × 480 pixels, coordinate estimation errors were approximately ±2 mm in the horizontal axis and ±5 mm in the vertical axis. The error increased when flowers were located closer to or farther from the reference plane, confirming that the planar conversion was depth-dependent. Consequently, the reported localization accuracy applies primarily to targets positioned near the 50 cm working plane and should not be generalized to flowers distributed across a wide range of canopy depths without recalibration or depth compensation. An example of the model output on a tomato canopy image, including detected flowers and the computed positioning coordinates used by the system, is shown in Figure 10 below.

The two-dimensional motion platform accurately reproduced the commanded travel distances under the tested operating conditions. Each condition was evaluated over five runs, and the reported travel times represent the mean of these repetitions. For the vertical axis, a commanded displacement of 160 mm resulted in a measured displacement of 160 mm under both motion settings. Mean travel times were 2.31 s at a set acceleration of 100 mm/s$^2$ and maximum speed of 200 mm/s, and 2.21 s at 200 mm/s$^2$ and 400 mm/s, respectively. For the horizontal axis, the commanded and measured displacement were both 640 mm, with mean travel times of 4.96 s and 3.22 s under the corresponding low- and high-speed settings. No observable step loss was recorded during the tests. These results indicate satisfactory positioning accuracy and consistent platform operation over the tested displacement and speed ranges, supporting reliable centering of the end-effector on detected flowers. The constructed two-dimensional motion stage used to position the fan-based pollination end-effector is shown in Figure 11 below.

In comparative trials ($n=20$ plants per treatment), the automated system achieved a mean operational success rate of $83.68 \pm 6.07 \%$, where success was defined as centering the pollination end-effector on a flower and delivering an airflow puff. Manual pollination achieved 100\% operational success for the tested block because each target flowe was tapped and no flowers were missed under the study definition. A one-sample t-test indicated that the automated success rate differed significantly from the 100% baseline, $t(19)=-12.03$, $p<0.001$. Mean time per flower was comparable between treatments: manual pollination averaged $4.14 \pm 0.59$ s per flower and the automated system averaged $4.46 \pm 0.62$ s per flower, with no statistically significant difference between treatments, $t(38)=-1.70$, $p= 0.097$. A comparison of the manual and automated pollination trials, including operational success definitions and time per flower, is provided in Table 4.
| Metric | Manual | Automated |
|---|---|---|
| Plants per treatment | 20 | 20 |
| Operational success definition | Flower gently tapped | Centered on flower + airflow puff applied |
| Success rate | 100\% | $83.68 \pm 6.07\%$ |
| Time per flower | $4.14 \pm 0.59$ s | $4.46 \pm 0.62$ s |
| Time comparison ( | -- |
Figure 12 summarizes the comparative outcomes for manual versus automated pollination, including operational success rate and time per flower.

5. Discussion
The detection results and real-time frame rates indicate that a lightweight YOLO-based model can support practical flower localization for targeted actions in a greenhouse. This is consistent with a broader trend in precision agriculture where compact detectors are preferred when deployment hardware is limited [14], [15]. In the proposed prototype, the perception module was paired with a simple positioning mechanism and a non-contact pollination method. The mean operational success rate of 83.68% suggests that targeted airflow can be effective, which matches earlier evidence that air jets can trigger pollen release without touching the flower [8]. Although the overall detection metrics were encouraging, the confusion matrix and qualitative analysis show that the detector was not equally reliable under all canopy conditions. The confusion matrix recorded 27 missed flower instances and 28 false-positive detections. The representative validation examples indicate that missed detections occurred when flowers were closely positioned, partially overlapping or difficult to distinguish from surrounding plant structures. A false-positive prediction was also observed where a background region presented visual features that resembled those of a flower. Similar difficulties with occlusion, scale variation and complex agricultural backgrounds have been identified as important limitations of object detection systems used in precision farming [14], [15].
Some false-positive predictions occurred at relatively low confidence levels, suggesting that increasing the confidence threshold could suppress a proportion of these errors. However, a stricter threshold could also increase the number of missed flowers. Future model development should therefore combine confidence-threshold optimization with additional training images containing overlapping flowers, difficult illumination conditions and visually similar background structures. Hard-negative mining, targeted photometric augmentation and evaluation at higher image resolutions could also improve robustness. The time-per-flower comparison is also encouraging. Under the study procedure, manual tapping and the automated airflow-based system produced similar average times, and the difference was not statistically significant. This finding is practically relevant because reducing labor requirements is an important consideration in the adoption of managed pollinators or mechanical pollination methods [6], [7].
Compared with recent robotic pollination platforms, the proposed system was developed around a deliberately simplified hardware architecture. Drone-based systems offer flexible movement within three-dimensional greenhouse space, while mobile manipulators and suspended robots provide greater reach and control of flower orientation [19], [21], [23], [24]. These capabilities are supported by additional hardware such as motion-capture infrastructure, depth cameras, six-degree-of-freedom robotic arms and flexible manipulators. The present system instead uses a fixed two-dimensional stage, a single low-cost camera and a fan-based end-effector. This reduced the prototype cost and simplified construction and maintenance. However, the resulting platform cannot compensate fully for large variations in flower depth or orientation. The system should therefore be considered a low-cost solution for constrained greenhouse layouts rather than a direct replacement for more flexible three-dimensional robotic platforms. The displacement tests also showed consistent agreement between commanded and measured travel distances across the repeated runs under each operating condition. However, a more extensive repeatability assessment involving a larger number of positioning cycles and statistical measures of endpoint-position dispersion would be required to characterize long-term positioning precision.
Several limitations should be noted. First, occasional degradation of the vision pipeline was observed during peak sunny periods in the greenhouse. During the periods when these slowdowns or stalls occurred, average greenhouse temperature readings were generally above approximately 35 °C and illumination was intense. The same behavior was not observed during experiments conducted after the peak sunny hours, when greenhouse temperatures were generally below this range. No controlled thermal experiment was conducted and temperature and illumination effects were not independently isolated; therefore, 35 °C cannot be interpreted as a formal failure threshold. Instead, it represents an empirical operating condition at which additional thermal protection was considered advisable. In subsequent operation, the camera was installed inside a ventilated protective enclosure and cooled using a small 12 V fan, which resolved the observed instability. Controlled testing across defined temperature and illumination ranges would nevertheless be required to establish the thermal operating limits of the system.
Second, the coordinate-localization method relied on planar scaling calibrated at a fixed reference depth of 50 cm. This approach reduced computational and hardware requirements, but its accuracy depended on maintaining the same camera position, orientation, image resolution and working distance used during calibration. Moving the camera closer to or farther from the canopy changes the pixel-to-millimeter relationship, while camera tilt introduces non-uniform perspective distortion across the image. The current system is therefore most suitable for greenhouse arrangements in which the camera can be rigidly mounted and the flowers remain near a controlled working plane. Recalibration would be required after changing the camera mounting geometry or operating distance. For canopies with substantial depth variation, practical deployment would require either calibration at several reference depths, interpolation between depth-specific scale factors, or direct depth measurement using stereo vision, a depth camera or another three-dimensional localization method.
Third, the evaluation focused on operational success criteria: for the automated system, success meant centering on the flower and applying the airflow puff; for manual pollination, success meant that each flower was gently tapped and none were missed during the trial block. These criteria are useful for assessing end-to-end targeting behavior, but they do not directly measure downstream outcomes such as fruit set, seed count, or fruit quality. Future studies should connect the automation outcomes to agronomic indicators used in pollination research and tomato production studies [1], [25]. A further limitation concerns the evaluation of the flower-detection model. The dataset was divided into training and validation subsets, and no independent held-out test set was used. The same validation subset was used for early stopping, model selection and quantitative performance reporting. Consequently, the reported precision, recall, F1-score and mAP values should be interpreted as validation performance and not an independent estimate of generalization to unseen data. The additional real-time tests on tomato foliage provided an operational assessment under greenhouse conditions, but they did not replace an independently labeled test set. Future work should therefore include a separate held-out test dataset, preferably collected under different greenhouse conditions or during a different production cycle. Additionally, because the training-validation partition was performed at the image level, plant-disjointness between the two subsets was not enforced. This should be considered when interpreting the reported validation performance, and future datasets should preferably use plant-disjoint training, validation and independent test partitions. Finally, although the dataset included varied viewpoints and backgrounds, it remains modest in size (710 images). Broader data collection across seasons, cultivars, and greenhouse lighting conditions would help reduce failure modes and improve generalization, as emphasized in recent reviews of autonomous pollination and agricultural object detection [12], [13].
6. Conclusion
This study developed and experimentally evaluated a low-cost artificial intelligence-guided ventilated pollination system for greenhouse tomato flowers. The system combined YOLOv8n-based flower detection, planar coordinate estimation, a two-dimensional motion platform and a fan-based pollination end-effector. The detector achieved a precision of 0.841, a recall of 0.847, an F1-score of 0.84 and an mAP@0.5 of 0.906. At the 50 cm calibration distance, coordinate estimation errors were approximately ±2 mm horizontally and ±5 mm vertically. The integrated prototype achieved a mean operational success rate of 83.68\%, while the difference in mean time per flower between automated and manual pollination was not statistically significant. With an estimated construction cost of USD 167, the prototype demonstrates the feasibility of a mechanically simple and non-contact pollination system for resource-constrained greenhouse production. The next stage of the work should extend the present two-dimensional positioning system to three-dimensional flower localization so that flowers located at different canopy depths can be targeted more accurately. This could be achieved through direct depth sensing or a multi-depth calibration approach. A multi-target planning strategy is also needed to determine the order in which several detected flowers should be treated and to prevent the same flower from being pollinated more than once. Longer greenhouse trials should be conducted over complete flowering and fruit-development periods to evaluate fruit set, seed formation, fruit quality, energy consumption, system reliability and maintenance requirements. Further work should also expand the image dataset with more difficult examples involving flower overlap, occlusion and intense illumination, while improving thermal management and embedded inference stability under midday greenhouse conditions.
Conceptualization, F.D.F.C., and A.R.T.; methodology, A.R.T., and E.F.M.D.; software, A.R.T., and F.D.F.C.; validation, J.K.T., and A.R.T.; formal analysis, A.R.T., E.F.M.D., F.D.F.C., and J.K.T.; investigation, A.R.T., E.F.M.D., F.D.F.C., and J.K.T.; resources, A.R.T., E.F.M.D., F.D.F.C., and J.K.T.; data curation, A.R.T., E.F.M.D., F.D.F.C., and J.K.T.; writing—original draft preparation, F.D.F.C. and A.R.T.; writing—review and editing, F.D.F.C. and A.R.T.; visualization, J.K.T.; supervision, J.K.T. All authors have read and agreed to the published version of the manuscript.
The data used to support the research findings are available from the corresponding author upon request.
The authors thank the Department of Agricultural Engineering at the University of Dschang for providing the experimental greenhouse and for access to the facilities and equipment.
The authors declare no conflicts of interest.
