A CNN-Based Micro-UAV System for Real-Time Flower Detection and Target Approach

Abstract

This paper presents the application of a micro unmanned aerial vehicle (UAV) that acts as a pollination agent in a controlled environment simulating greenhouse conditions. The micro-UAV system was integrated with a convolutional neural network (CNN) for autonomous flower detection and navigation. The custom Sequential CNN architecture was used on board to perform real-time binary classification, accurately distinguishing flowers from non-flower objects. The fusion of this deep learning-based detection with precise micro-UAV navigation enables efficient identification and approaches to target flowers within optimal operational distances. Experimental evaluations revealed that the micro-UAV’s onboard camera, combined with CNN processing, outperformed standard webcams in terms of detection speed and accuracy, demonstrating the benefits of specialized hardware. Within the experiment, the micro-UAV was pre-programmed to follow a ‘cross’-shaped flight pattern. Experimental results show that the proposed system successfully detects multiple flowers autonomously between distances of 30.5 cm and 91.5 cm within 149.1 s. Overall, this study validated the integration of neural network capabilities with micro-UAV navigation. These findings are crucial for highlighting the potential of neural network-enabled micro-UAVs as effective pollinators in enclosed agricultural environments and for addressing the challenges faced by natural pollinators in greenhouses.

1. Introduction

Food security, a critical global concern, is intricately linked to the United Nations’ Sustainable Development Goals (SDGs), particularly SDG 2 (Zero Hunger). As the world population continues to grow and is projected to reach 9.7 billion by 2050, ensuring food security and maintaining food quality have become increasingly challenging [1,2]. Future forecasts suggest a potential strain on food systems, with climate change and resource scarcity further complicating these issues. In developing countries, agricultural industries play a pivotal role in addressing food security challenges while simultaneously contributing to national income [3,4]. These industries not only enhance local food production but also create employment opportunities and stimulate economic growth in the region. However, achieving sustainable food security requires a multifaceted approach that integrates innovative farming techniques, such as autonomous robots or agricultural drones, to enhance agricultural production, environmental conservation, and economic development [5,6].

Food security issues are closely intertwined with sustainable and smart agricultural practices. Sustainable agriculture focuses on producing food in an environmentally friendly, economically viable, and socially responsible manner. Smart agriculture, a subset of sustainable agriculture, leverages technological and data-driven approaches to optimize the farming process [7]. The elements of smart agriculture include (a) precision farming techniques, (b) Internet of Things (IoT) sensors and devices, (c) data analytics and artificial intelligence (AI), (d) automated irrigation systems, (e) drone technology for crop monitoring, (f) climate-smart practices, and (g) soil and crop management tools [8,9]. Given the broad challenges of food security and smart agriculture, this study considers the transformative potential of drone technology as a breakthrough in addressing some of these challenges. This study proposes a solution to agricultural problems using micro-unmanned aerial vehicle (UAV) technology with a micro dimension of less than 100 mm × 100 mm. This is to take advantage of navigating tight spaces, such as a vertical farm. Previous drone-based pollination systems utilized larger dimensions [10,11]. The proposed method and solution lay the foundation for research on utilizing micro-UAVs as smart agricultural tools.

The novelty of this research lies in the deployment of a CNN within a real-time, micro-scale robotic pollination framework. While previous studies utilized larger drone platforms with limited maneuverability, this study demonstrates a methodology using a micro-UAV with dimensions less than 100 mm × 100 mm. A key feature of the proposed system is the seamless integration of the drone’s specialized image-processing hardware with an external workstation via a high-speed local Wi-Fi connection. This hybrid configuration allows for the real-time transmission of 5 MP image data to a custom Sequential CNN model, enabling the drone to autonomously identify sunflowers and execute precise ‘cross’-shaped flight patterns in spaces too small for traditional agricultural drones. The main tasks of drone is to find flowers and fly towards them from a specific starting point. Before the drone was used, it was trained to recognize flowers by using a CNN. During operation, the pollinator drone takes off and examines its surroundings. If it spots a flower nearby, it flies close to it. The goal of this study was to improve the navigation and flower detection capabilities of micro drones using CNN technology, potentially aiding pollination tasks in the future.

The primary contributions and prominent features of this study are summarized as follows:

• Micro-Scale Navigation: Demonstration of a micro-UAV platform (<100 mm × 100 mm) capable of navigating the highly confined spaces typical of vertical farms and greenhouses, where larger drones are impractical.
• Real-Time CNN Integration: Implementation of a custom Sequential CNN architecture for real-time binary classification (‘flower’ vs. ‘not flower’) directly integrated with the micro-UAV’s autonomous flight system.
• Hardware Performance Benchmarking: A comparative analysis revealing that specialized onboard image-processing hardware significantly outperforms standard webcams in terms of detection speed and accuracy.
• Operational Optimization: Empirical determination of the optimal operational range ($30.5$ cm to $91.5$ cm) for effective flower detection, providing a foundation for future autonomous pollination strategies.
• Autonomous Flight Validation: Validation of a pre-programmed ‘cross’-shaped flight pattern that enables the drone to systematically identify and approach multiple targets without human intervention.

In Section 2, we provide an overview of drone use in agriculture, the potential of drone pollinators, AI classifiers, and flower recognition systems, and discuss the problems. In Section 3, we present the details of the experimental methods, which consist of a binary classification technique using a custom Sequential CNN to detect flowers. We also elaborate on the integration of AI during real-time drone flights. In Section 4, we discuss the experimental results and effectiveness of drone AI detection versus distance. Finally, we conclude the paper and provide directions for future research in Section 5.

2. Related Works

2.1. Agriculture Micro-UAV

Micro-UAVs, also called drones, offer numerous benefits for smart agriculture. Off-the-shelf agricultural drone solution packages are equipped with onboard monitoring sensors that can provide real-time data [12]. A typical onboard sensor package includes a camera that can be utilized for additional functionalities. Current solutions employing onboard cameras are used to monitor crop conditions, detect pest infestations, and perform other tasks [13]. Another promising area of research is the use of drones as agricultural pollinators. Bees are the primary pollinators of many crops. However, in greenhouses or enclosed agricultural settings, bee penetration is limited, posing challenges to pollination [14]. Given this limitation, pollinator drones are a potential solution. With the help of onboard cameras, pollinator drones can identify the location of flower buds for subsequent pollination [15]. Drones can simultaneously collect data on crop health and pollination efficacy while executing pollination tasks. Automated drone pollination has the potential to reduce labor costs associated with manual pollination methods. However, the use of drones as pollinators is still in the experimental phase and faces challenges such as limited battery life, potential impacts on local ecosystems, and the need for further technological refinement.

2.2. Potential of Autonomous Pollinator Drone

Although bees naturally serve as pollinators, sustainable smart agriculture often employs closed environments, such as greenhouses or indoor spaces, such as vertical farms, which restrict bees’ access to natural pollination [16,17,18]. To address pollination challenges in controlled settings, the use of drones as pollinators has emerged as an innovative solution [15,19,20]. Despite their long and excellent track record as pollinators, traditional pollinators such as bees encounter difficulties reaching enclosed agricultural areas [21,22]. This has led researchers to investigate alternative pollination methods for such crops. In recent years, various solutions have been proposed, including mobile robots with robotic hands and soft actuators with high potential as pollination mechanisms. However, a significant limitation is the initial investment cost, as mobile robots and soft actuators must be developed to meet specific environmental requirements [23,24,25,26]. In contrast, off-the-shelf micro-UAVs are highly programmable, making them well-suited for autonomous pollination tasks [27].

Recent reports have indicated that pollinator drones equipped with specialized attachments can replicate the pollination process. These drones, categorized under the vertical takeoff and landing (VTOL) configuration, offer precise control and the ability to operate in confined spaces such as greenhouses and indoor farming environments [19]. This technology has the potential to sustain crop productivity. Empirical studies have confirmed the efficacy of drone pollination in various crops, including strawberries, peppers, tomatoes, and kiwifruits [15,19,28,29]. The benefits of drone pollinators include their consistent operation regardless of weather conditions, programmability for optimal pollination timing, and potential to reduce labor costs [30,31]. Nonetheless, challenges persist, particularly the need for further refinement of drone sensor payloads, particularly navigation and imaging sensors, to enable autonomous navigation to flowers and match the efficiency of natural pollinators. Additionally, concerns regarding the long-term ecological impact of deploying multiple drones simultaneously in greenhouse agricultural settings remain. Although short-term ecological impacts may enhance agricultural efficiency, long-term ecological consequences require further investigation [32,33]. One notable negative impact is the buzzing noise produced by drones, which may cause biological disturbances to local bird and insect populations [34]. As research in this domain advances, drone pollination technology continues to evolve, potentially serving as a viable complement to natural pollinators to ensure food security and promote sustainable agricultural practices.

Determining the appropriate size and suitability of pollinator drones for robotic pollination in greenhouse environments is crucial for optimizing their effectiveness and efficiency [15]. The drone must be micro-sized to navigate confined spaces while carrying essential pollination equipment. Micro drones can easily maneuver between plant rows and surrounding structures [35]. However, accommodating a sufficient pollen payload along with necessary sensors or cameras presents significant challenges. The ideal size balances the agility and functionality, which is the primary reason why this study selected a micro-UAV configuration as the main drone platform.

2.3. Simultaneous Navigation and Flower Recognition

Flower detection and navigation are crucial components for the development of autonomous pollination systems using drones. Recent studies have focused on improving the accuracy and efficiency of these processes to enhance drone performance in agricultural settings [36]. AI techniques for computer vision, particularly deep-learning algorithms, have shown promising results in flower detection. CNNs have been successfully applied to identify and locate flowers in complex greenhouse environments, achieving high accuracy rates. These models can distinguish among different flower species and their growth stages, thereby enabling targeted pollination [37,38].

The enclosed nature of greenhouses limits the use of traditional global positioning systems (GPS), necessitating alternative navigation methods [39,40]. Researchers have explored various approaches to overcoming the limitations of GPS in enclosed spaces. Visual Simultaneous Localization and Mapping (VSLAM) techniques have been implemented to allow drones to create and update maps of their environments in real time [41]. Among these techniques, cameras and computer vision algorithms can help drones maintain their position and navigate through greenhouses [39,42]. Sensor fusion, which combines data from cameras, infrared sensors, and inertial measurement units, has improved the precision of drone movement around plants [43]. Some studies have investigated the use of artificial landmarks or QR codes in greenhouses to aid drone localization [44,45].

The importance of navigation in greenhouse settings cannot be overlooked. Accurate navigation ensures that drones systematically cover all plants that require pollination without missing areas or over-pollinating others. They also prevent collisions with greenhouse structures or plants, which can damage both drones and crops [46]. Infrared or ultrasonic sensors can be employed for obstacle avoidance and precise maneuvering around plants [47]. Efficient navigation contributes to energy conservation, allowing drones to operate for longer periods and cover larger areas. Additionally, precise navigation enables the collection of valuable data on plant health and growth patterns, contributing to overall greenhouse management and crop optimization [48]. The integration of flower detection and navigation systems has led to the development of more efficient pollination strategies, optimized flight paths, and reduced energy consumption.

2.4. Drone-Based Pollination

Recent advancements in autonomous pollination have diversified across various platforms and detection architectures. While earlier studies focused on feasibility, contemporary research from 2024 to 2026 has shifted to enhancing precision under complex environmental conditions. For example, one study demonstrated an intelligent integrated approach to fruit detection by leveraging YOLOv5-v1 as a DCNN architecture for detecting fruits with a high mean average precision rate of 86.8% in fruit detection [49]. Another research effort enhanced an object detection model, RAFS-YOLO, built upon the YOLOv11 framework, specifically targeting strawberry maturity detection, combining a lightweight deep learning model with an RGB-D camera [50]. This modification improved detection accuracy and robustness in challenging illumination scenarios. In contrast to these large-scale orchard applications, our study focuses on the unique constraints of confined greenhouse environments, utilizing a micro-UAV platform (<100 mm) and the custom lightweight Sequential CNN architecture to balance agility with detection accuracy.

3. Experimental Setup

This section provides an overview of the proposed drone pollination system, which comprises two main parts. The first part involved the use of the custom sequential machine learning CNN algorithm for binary classification of flowers, as illustrated in Figure 1. This algorithm, commonly used for image analysis, helps us categorize objects into two classes: ‘flower’ or ‘not flower’. We developed an AI-based flower identification system that enables drones to hover toward detected flowers without human intervention. Building on this flower classification using the custom sequential algorithm, the second part of our study focused on autonomous navigation for real-time field test flower recognition. A drone platform with an onboard camera is the most crucial device. An indoor laboratory was chosen for the real-time field test drone flight and flower recognition evaluation.

To achieve autonomous flower detection in a confined environment, a cohesive hardware and software architecture was developed. Figure 2 illustrates the overall system architecture, demonstrating the real-time data flow between the micro-UAV and the computational workstation. Due to the limited onboard processing capabilities of the micro-UAV, the system utilizes a distributed approach. The drone captures and transmits HD video via a local Wi-Fi network to the external workstation. The workstation processes these frames using the custom lightweight Sequential CNN to identify pollination targets. Upon successful positive classification, the workstation immediately transmits the corresponding flight control commands back to the UAV, enabling closed-loop, autonomous navigation.

3.1. AI Binary Classification for Flower Recognition

Before we conduct the real-time field test, there is an important step called AI binary classification for flower recognition using a custom sequential machine learning CNN algorithm. As shown in Figure 1, this classification is a key part of the process for handling the raw visual data of the flowers. To ensure the robustness of the custom sequential model, a balanced dataset was curated specifically for binary classification. The dataset characteristics are detailed as follows:

• Data Sources: The training set consists entirely of a proprietary database created for this research. All images were captured directly within the experimental laboratory environment to ensure strict environmental consistency and accurate representation of the micro-UAV’s operational conditions.
• Dataset Size: A total of 555 image samples were collected and utilized for the model’s training and evaluation phases.
• Class Distribution: The 555 samples were distributed across two primary categories: the ‘flower’ class (featuring sunflowers) and the ‘non-flower’ class.
• Data Partitioning: The dataset was partitioned into two subsets: 80% for training (444 images) to train the custom sequential model and update its internal weights over the 64 epochs, and the remaining 20% (111 images) was reserved for validation and testing to tune hyperparameters and evaluate final performance.
• Preprocessing: All images were resized to a standard $224\times 224$ pixel resolution and normalized to match the input requirements of the custom sequential CNN framework.
• Data Augmentation: In this foundational phase of the research, no data augmentation techniques were applied. The custom sequential model was trained exclusively on the original raw dataset to establish a baseline for detection performance in the controlled laboratory environment.

The process began with the collection of flower images, as shown in Figure 3. Subsequently, we pre-trained the model and performed classification. We used a pre-trained model, a custom sequential model, which can be fine-tuned specifically for our dataset. These categorized data subsets were fed into the custom sequential framework, which learns to recognize specific patterns and features of the flowers. Finally, the custom sequential framework performs a binary classification task to determine whether the input image shows a ‘flower’ or ‘not flower’. We compared the classification performance of a drone’s onboard camera and a standard webcam, focusing on the time it takes to recognize a flower, measured in seconds. To ensure a rigorous and fair benchmark between the two imaging systems, a highly controlled experimental setup was established. The comparative baseline was a standard commercial webcam connected directly to the workstation via USB 3.0, operating at a resolution of 720p at 30 fps to match the HD720P30 video format of the DJI Tello drone. During the comparative test, both the drone and the webcam were statically mounted side-by-side at a fixed optimal distance of 60.5 cm from a single, stationary sunflower target. The laboratory lighting was maintained at a constant fluorescent baseline to eliminate photometric variables. Both video streams were processed sequentially by the workstation using the identical custom sequential weights and inference scripts. The detection time was recorded as the latency required for the system to register the specified number of consecutive positive classifications (10, 20, 30, 40, 50, and 60 detections).

When selecting the AI architecture for the micro-UAV’s real-time navigation system, standard deep learning models such as custom sequential were evaluated. However, due to the strict latency requirements of autonomous flight and the hardware constraints of the system, a massive architecture with millions of parameters was deemed computationally inefficient for this specific application. Instead, a custom, lightweight Sequential CNN was explicitly chosen and designed for this study. Because the system only requires binary classification (‘flower’ vs. ‘non-flower’) in a controlled environment, this custom architecture achieves high detection accuracy while drastically reducing computational payload, thereby enabling rapid real-time inference.

To evaluate the operational feasibility of the proposed system, the architectural footprint and computational complexity of the classification model were quantified. The custom Sequential CNN was designed to be exceptionally lightweight, processing input images at a 32 × 32 × 3 resolution. The model architecture is built with the following sequential layers:

• A Conv2D input layer with 32 filters and a 3 × 3 kernel size.
• A MaxPooling2D layer with a 2 × 2 pool size.
• A second Conv2D layer with 32 filters and a 3 × 3 kernel size.
• A second MaxPooling2D layer with a 2 × 2 pool size.
• A Flatten layer, preceded by a 0.25 Dropout layer to mitigate overfitting.
• A fully connected Dense layer with 128 units.
• A final output Dense layer with 2 units, preceded by a 0.5 Dropout layer.

To ensure full transparency and reproducibility, the computational environment utilized for model training and real-time execution is detailed as follows.

• Hardware Environment:

  –

  The database loading, network training, and image processing tasks were executed on a local laptop computer, which was subsequently connected to the DJI Tello Robomaster TT drone for real-time inference. Because the OpenCV processes are computationally intensive, the system leveraged a Graphics Processing Unit (GPU) equipped with Compute Unified Device Architecture (CUDA) capabilities. This provided the essential computational power and dedicated CUDA cores required to accelerate the model training and maintain real-time object detection performance.

• Software Environment:

  –

  Programming Language & IDE: Python version 3 served as the primary programming language for developing the machine learning algorithms, processing images, and training the CNN. Visual Studio Code (VS Code) version 1.70 was utilized as the integrated development environment (IDE) to write, debug, and execute the training routines.

  –

  Computer Vision & Deep Learning Libraries: OpenCV, an open-source computer vision library, was employed for intensive image and video processing, feature extraction, and object recognition tasks. The deep learning implementation and actual model training were conducted utilizing the Keras and TensorFlow frameworks, supported by NumPy for numerical data manipulation.

3.2. Micro-UAV Specification

The system for the micro-UAV pollinator navigation platform involves carefully selecting a lightweight, compact drone capable of stable flight within the confined and controlled environment of a greenhouse. The DJI Tello Robomaster TT was selected with key specifications including payload capacity to accommodate integrated sensors and cameras, extended battery life for prolonged operation, and precise maneuverability to navigate narrow aisles and avoid obstacles. The hardware setup prioritizes modularity and reliability, ensuring that the UAV can be easily maintained and adapted to different greenhouse layouts and crop types.

Table 1. Drone Technical Specifications [51].

Item	Parameter	Specification
Drone	Take-off weight	87 g (including propeller blades, propeller blade protector, and batteries)
	Dimensions	$98\times 92.5\times 41$ mm
	Propeller blade	3”
	Built-in functions	Infrared height determination, barometer, LED indicator, downward vision sensor, Wi-Fi, HD 750P image transmission
	Interface	Micro USB charging port
	Removable battery	1.1 Ah/3.8 V
Flight Performance	Maximum flight distance	100 m
	Maximum flight speed	8 m/s
	Maximum flight time	8 min
	Maximum flight height	30 m
Camera	Image	5 MP
	Field of View (FoV)	82.6°
	Video	HD720P30
	Format	JPG (images), MP4 (videos)
	Electronic image stabilization	Supported

summarizes the specifications of the DJI Tello Robomaster TT used in this study. The onboard camera was central to the functionality of the drone for flower detection. A high-resolution 5 MP camera enabled the accurate identification and monitoring of flower positions. To facilitate real-time tracking and continuous CNN inference, the system was configured to handle the drone’s video stream at a rate of 15 frames per second (fps) at HD720P resolution. Sunflowers were used in this project because they have unique visual characteristics, such as bright yellow petals and distinct shapes, which make them distinguishable from the surrounding environment. An indoor laboratory was selected as the experimental field to mimic an indoor greenhouse scenario, as shown in Figure 4.

3.3. Integration of AI Detection Capability on the Micro-UAV

In the next part of our research, we incorporated a drone into the real-time field experimental setup, as depicted in Figure 4. In our laboratory tests, we arranged four identical flowers in a ‘cross’ configuration, maintaining specific distances between the flowers and drone, as illustrated in Figure 5. The drone was initially positioned at its center. We systematically adjusted the distance between the drone and the flowers for each subexperiment, conducting tests at intervals of 15.5 cm, 30.5 cm, 60.5 cm, 91.5 cm, and 116.5 cm, as shown in Figure 5. At each distance, we documented the efficacy of drone in flower detection. Upon takeoff, the micro-UAV establishes a stable connection with the workstation using a local Wi-Fi network. The workstation serves as the primary computational hub, processing the real-time video stream using the custom sequential algorithm to classify objects as ‘flower’ or ‘not flower’. This approach ensures that the drone maintains its 8 min flight time and 87 g take-off weight while benefiting from robust AI-driven navigation and detection capabilities. Following a vertical ascent to approximately 1.5 m above ground level, the drone commenced its search for flowers.

The flight path was pre-programmed as shown in Figure 6. Initially, the drone ascended vertically after it was successfully connected to the local Wi-Fi. Once it reached approximately 1.5 m above ground level, the drone was incrementally rotated 10° until a flower was detected in its video frame. The process of flower recognition during flight is illustrated in Figure 7. The drone advances forward when a flower is detected. In the event that no flower is identified in the video stream, the drone continues to rotate in 10° increments until the flower is located. If a flower is detected during rotation, the drone moves towards it. The drone lands when its battery level falls below 20%.

4. Results and Discussion

To ensure the reproducibility of the AI binary classification model, the specific hyperparameters used to train the custom sequential architecture are summarized in

Table 2. Training hyperparameters for the custom sequential CNN model.

Parameter	Value
Epochs	64
Batch Size	32
Optimizer	Adam
Learning Rate	Default (Adam)

. The training process was conducted over 64 epochs, with 14 iterations per epoch. The Adam optimizer was selected, utilizing its default learning rate, to effectively minimize the loss function during the transfer learning phase.

The custom dataset, partitioned into ‘flower’ and ‘non-flower’ classes, was utilized to train and evaluate the model. The training process was conducted over 64 epochs, with 14 iterations per epoch. As illustrated in Figure 8, the model achieved a training accuracy of 1.0, indicating that it successfully learned the visual features of the training data. Crucially, alongside this perfect training score, the testing accuracy stabilized at approximately 0.97 (97%). This demonstrates that the custom CNN effectively generalized the learned features to unseen data without succumbing to severe overfitting, ensuring it can accurately predict the correct labels for novel inputs during real-time deployment.

The preliminary experiment aimed to assess the classification algorithm’s ability to recognize a flower, with the performance measured in seconds. A comparison was made between the drone’s onboard camera and the standard webcam.

Table 3. Processing time comparison between webcam and drone camera.

Number of Detections	Webcam (Seconds)	Drone Camera (Seconds)
10	28.01	11.34
20	44.20	12.73
30	61.65	13.04
40	90.07	13.43
50	102.55	14.44
60	129.66	14.86

presents the experimental data for detection time. Detection commenced at intervals of 10, 20, 30, 40, 50, and 60. Both the webcam and drone camera operated in real-time mode. As the number of detections increased, the time required for both the webcam and the drone camera also increased. Notably, the webcam required more time for detection than the camera of the drone. This discrepancy can be attributed to several factors, primarily the hardware specifications. The DJI Tello Robomaster TT drone is equipped with specialized image-processing hardware and dedicated processors, including GPUs, optimized for real-time image-analysis tasks. These components facilitate parallel computations and expedite the detection process. In contrast, webcams typically rely on a computer’s CPU, which may have fewer computational resources and slower processing speeds compared to specialized hardware.

To evaluate the system’s real-time inference speed and latency, a comparative experiment was conducted, measuring the total time required to complete a set number of continuous detections. Both the webcam and the custom CNN integrated with the drone operated continuously in “real-time mode,” rapidly generating detection outputs until manually terminated. As detailed in Table 3, as the required number of detections increased, the time required for both systems naturally increased. However, the drone’s camera demonstrated vastly superior inference speeds, requiring only 14.86 s to complete 60 continuous detections—averaging an effective inference speed of approximately 0.25 s per detection. In stark contrast, the webcam experienced significant system latency, taking 129.66 s to complete the same 60 detections (averaging over 2 s per detection). This drastic reduction in latency can be attributed primarily to the hardware architecture. The DJI Tello Robomaster TT drone is equipped with specialized image-processing hardware and dedicated processors, including GPUs, that are explicitly optimized for real-time image analysis. These components facilitate parallel computations, effectively accelerating the inference speed. Conversely, the webcam setup relies entirely on the host computer’s CPU, which possesses fewer computational resources and slower sequential processing speeds, resulting in the observed detection bottleneck.

In the subsequent phase of this experiment, the drone was programmed to navigate in a ‘cross’-shaped pattern, as illustrated in Figure 9. The drone’s rotational movement was detected, and the graph indicated that the drone’s camera identified four flowers during its rotation to locate them. The first detection occurred between 32.2 s and 38.6 s, at angles ranging from 49° to 113.6°. The second detection took place from 46.5 s to 49.9 s, at an angle of −147.2° to −95°. The third detection was recorded between 103.7 s and 108.8 s, at an angle of 69.9° to 133.1°. The final detection occurred from 143.4 s to 149.1 s, at an angle of 82.5° to 145.2°. At certain points, the angle on the line graph dips in the negative direction, indicating a counterclockwise rotation and a descent or lowering of the drone.

According to

Table 4. Boolean detection results at different distances.

Distance (cm)	Detection (Boolean)
15.5	0
30.5	1
60.5	1
91.5	1
116.5	0

, if the distance between the drone and the flower was too small (i.e., 15.5 cm), the drone failed to detect it. Conversely, if the distance is too large (i.e., 116.5 cm), the drone also struggles to locate the flower. This is because when the drone is too close, its sensors and cameras may not have a sufficiently wide field of view to capture the entire object or provide accurate measurements, making it difficult to effectively detect and track the flower. However, if the drone is too far, the flower’s size in the camera’s field of view diminishes, causing the drone’s computer vision algorithms to have difficulty accurately identifying the flower, especially if there are other objects or cluttered backgrounds. Additionally, lighting conditions can affect object detection; if the flower is poorly lit or the lighting is too harsh, the drone’s sensors and cameras may struggle to distinguish the flower from its surroundings.

To assess the system’s practical viability, the micro-UAV’s flight duration and energy consumption were evaluated in terms of overall system efficiency. The micro-UAV is equipped with a 1.1 Ah/3.8 V removable battery, yielding a theoretical maximum flight time of 8 min. A major advantage of the proposed system architecture is the offloading of the CNN computational load to an external workstation. By eliminating the need for heavy onboard AI processing, the drone’s power draw is optimized, dedicating its limited energy reserves almost entirely to rotor propulsion, hover stability, and HD video transmission.

During the ‘cross’-shaped flight experiment, the system successfully navigated and identified all four target flowers within 149.1 s. This continuous flight mission utilized approximately 31% of the total theoretical flight capacity. Factoring in the system’s safety protocol, which initiates an automatic landing sequence when the battery level falls below 20%, the effective operational window is approximately 6.4 min per charge. The ability to complete a multi-target search-and-approach task within 2.5 min demonstrates that the system is highly energy-efficient, capable of executing multiple targeted pollination cycles in a confined greenhouse setting before requiring a battery swap.

5. Conclusions

This study effectively demonstrates the implementation of a micro-UAV equipped with a CNN for autonomous flower detection and navigation within a controlled greenhouse environment. The drone successfully identified sunflower targets at optimal distances, but detection accuracy was only achieved when the drone was positioned excessively close to or far from the flowers, underscoring the need to maintain an effective operational range. The real-time detection capabilities of the drone’s specialized camera surpassed those of conventional webcams, thereby underscoring the benefits of dedicated onboard hardware. The ‘cross’-shaped flight patterns confirmed the drone’s ability to navigate and accurately identify multiple flowers. These findings highlight the potential of micro-UAVs as efficient pollination agents in enclosed agricultural settings and offer a promising solution to the challenges encountered by natural pollinators in greenhouses. Future research should focus on enhancing sensor capabilities, extending operational ranges, and implementing robust data augmentation techniques during model training. Furthermore, subsequent studies will employ a more comprehensive suite of evaluation metrics—specifically including Precision, Recall, and F1-score—to rigorously quantify the model’s reliability and false-positive rates under the unpredictable lighting and dynamic flight conditions of real-world greenhouses.