Tea Shoot Category Detection Based on UAV Remote Sensing(MDPI Agronomy 2026-08-17)

1. Introduction

As one of the pillar industries of Yunnan’s modern agricultural system with distinctive plateau characteristics, the tea industry plays an important role in regional economic growth, rural revitalization, ecological civilization construction, ethnic tea culture inheritance, and sustainable agricultural development. During the optimization and upgrading of Yunnan’s agricultural industrial structure and the promotion of high-quality modern agricultural development, the tea industry has shown prominent economic, ecological, and social value [1]. The detection efficiency and overall development level of the three tea shoot categories are closely related to the stable operation of the regional tea industry chain and supply chain, while also influencing the continuous improvement of industrial competitiveness. As core indicators for evaluating tea garden productivity and management efficiency, the distribution and abundance of the three tea shoot categories are shaped by multiple physiological and ecological factors, such as photosynthetic efficiency, nutrient absorption and transport capacity, growth and developmental status, and responses to pest and disease stress [2,3]. The distribution and abundance of the three tea shoot categories not only reflect the production potential of tea gardens but also provide an important basis for evaluating tea plant health, growth vigor, and ecological adaptability. Therefore, achieving efficient and accurate UAV remote sensing-based detection of three tea shoot categories in Yunnan has become a key requirement for accelerating the digital transformation of the tea industry, improving refined tea garden management, and supporting the high-quality development of the modern tea industry [4].

In recent years, the tea industry in Yunnan has continued to develop toward large-scale operation, standardized production, and refined management, making the accurate detection of three tea shoot categories an important basis for supporting high-quality industrial development. At present, however, the detection of three tea shoot categories still depends largely on manual sampling surveys and empirical identification methods. These approaches are not only labor-intensive and limited in timeliness but also susceptible to subjective judgment, making it difficult to satisfy the practical needs of modern tea plantation management for real-time, precise, and full-coverage monitoring [5]. With the rapid development of digital agriculture and smart agriculture, UAV remote sensing technology has gradually become an important technical means for agricultural resource investigation and crop growth monitoring. Its advantages, including operational flexibility, high spatial resolution, efficient data acquisition, and relatively controllable cost, provide a feasible technical pathway for improving the detection of three tea shoot categories [6].

To address the limitations of small-object detection in unmanned aerial vehicle (UAV) remote-sensing scenarios caused by rapid viewpoint changes and motion blur, Kıratlı et al. proposed the YOLOv11-EFAC network based on the nano version of YOLOv11n [7]. The network adopts a multistage optimization strategy in which EfficientNet-B0 is used as the backbone, and a Feature Pyramid Network (FPN) and Path Aggregation Network (PANet) are combined for multiscale feature fusion. In addition, a Squeeze-and-Excitation (SE) attention module is introduced to enhance small-object feature representation, while the loss function is optimized to improve small-object localization. Experimental results showed that YOLOv11-EFAC achieved a mean Average Precision at an intersection-over-union threshold of 0.5 ([email protected]) of 83.7% in UAV-based detection, representing a 21.4% improvement in small-object detection performance over its baseline model. These findings indicate the potential of this method for real-time small-object detection and tracking in UAV imagery.

To improve feature extraction, detection accuracy, and deployment efficiency in crop-yield detection, Wang et al. proposed a lightweight YOLOv11-based model named RSWD-YOLO [8]. The model combines a hierarchical-scale feature pyramid with lightweight feature-extraction modules to enhance multiscale feature fusion while reducing computational complexity. It achieved an mAP of 86.1% on the evaluated crop dataset, indicating its potential for efficient crop-yield monitoring.

To improve multiscale small-object detection in UAV imagery with complex backgrounds, Yu et al. proposed OTB-YOLO, which integrates full-dimensional dynamic convolution, triplet attention, and a bidirectional feature pyramid network [9]. On the evaluated UAV dataset, the model achieved a Precision of 95.6%, a Recall of 92.1%, and an [email protected] of 96.6%, outperforming its baseline by 3.2, 2.5, and 3.1 percentage points, respectively.

In our previous study, we developed a YOLOv10-based tea-target detector for low-resolution and heavily occluded UAV imagery [10]. That model combined an improved Shape-IoU loss, wavelet-transform convolution, and a Histogram Transformer to enhance fine-grained feature extraction and multiscale feature fusion. In contrast, the present study focuses primarily on target-scale variation, complex background interference, and inaccurate boundary localization. The proposed HR-YOLOv10-S replaces the wavelet-transform convolution used in our previous work with a Rectangular Self-Calibrated Module for directional spatial-context and boundary modeling. It also introduces Shape Weights and Scale Adjustment Factors into bounding-box regression to achieve shape- and scale-adaptive localization. Although the Histogram Transformer is retained, it is integrated with these newly designed components to form a different joint optimization framework. Therefore, the main novelty of the present work lies in the shape–scale adaptive regression strategy, the rectangular self-calibrated spatial modeling module, and their complementary integration for tea-target detection under complex UAV imaging conditions.

Deep learning-based object detectors have been increasingly applied to agricultural and UAV remote-sensing tasks, including crop recognition, tea-plant monitoring, small-object detection, and yield-related assessment [4,7]. Recent studies have improved detection performance through multiscale feature pyramids, attention mechanisms, lightweight convolutional structures, bounding-box regression optimization, and local–global feature fusion [11,12]. Nevertheless, UAV imagery acquired in complex agricultural environments remains challenging because of substantial target-scale variation, dense occlusion, uneven illumination, background clutter, and high visual similarity between crops and surrounding vegetation [7,8,9,10]. These difficulties are particularly evident in mountainous tea plantations, where variations in canopy morphology, overlapping branches and leaves, and changing viewpoints can weaken boundary features and reduce small-target detection accuracy. Although existing methods have improved either multiscale representation, contextual modeling, or localization accuracy, their ability to simultaneously model target geometry, directional spatial context, and global background statistics remains limited. Therefore, further research is needed to develop a tea-target detector that combines scale-adaptive regression, enhanced spatial-structure modeling, and robust multiscale feature fusion for complex UAV imaging conditions.

To address the above limitations, this study proposes an improved object detection network, HR-YOLOv10-S, for UAV remote sensing-based detection of three tea shoot categories in complex tea plantation environments [13]. The network is based on the YOLOv10 architecture and is designed to improve detection under conditions involving substantial target-scale variation, complex background interference, and inaccurate boundary localization. During bounding-box regression, shape weights and scale adjustment factors are introduced to improve the matching between predicted and ground-truth boxes. These factors assign adaptive weights and scale adjustments to candidate boxes according to their aspect ratios and scale characteristics, thereby improving localization accuracy. A rectangular self-calibrated module is incorporated into the backbone network. This module combines horizontal and vertical separable convolutions with a lightweight self-calibration mechanism to enhance the modeling of spatial structures and target boundaries. In addition, a histogram transformer is introduced to exploit global statistical information and suppress interference from non-target regions, including soil, weeds, and shadows. This design facilitates multiscale feature fusion and semantic enhancement. The study aims to develop an accurate and robust model for detecting and extracting information on three tea shoot categories from UAV remote-sensing imagery and to provide technical support for refined tea plantation management and the digital development of the tea industry [14]. It also provides a new theoretical basis and practical paradigm for the deep integration of deep learning and agricultural remote sensing technologies, as well as for the design and optimization of intelligent perception models in complex agricultural scenarios.

2. Materials and Methods

2.1. Dataset Construction

Menghai County in Xishuangbanna Dai Autonomous Prefecture and Fengqing County in Lincang City, Yunnan Province, were selected as the data-acquisition areas. These two plateau mountainous tea-producing regions contain substantial variation in terrain, canopy structure, planting density, background composition, and illumination. Such variation is particularly relevant to the present detection task because it produces changes in target size, appearance, occlusion, and the visual similarity between tea targets and surrounding vegetation. In addition, the dataset includes multitemporal and multiview UAV images acquired at different tea growth stages and times of day. Consequently, the dataset contains pronounced variations in target scale, morphology, viewing angle, and lighting conditions, which directly correspond to the detection challenges addressed by HR-YOLOv10-S [15].

The data acquisition strategy covered the key growth stages of tea trees, including the young bud stage, the one bud and one leaf stage, and the one bud and two leaves stage [16]. Several rounds of aerial imaging were conducted at different times and from different viewing angles to increase sample diversity under low-altitude UAV remote-sensing conditions. Images were acquired at vertical distances of 1.0, 2.0, and 3.0 m above the tea canopy. It should be emphasized that these values represent the relative vertical distance between the UAV and the upper surface of the tea canopy, rather than the absolute flight height above the local ground surface. The objects investigated in this study were mature Yunnan large-leaf tea plants, which generally have plant heights of approximately 2.0–3.5 m, while some old tea trees can exceed 5 m. Therefore, the corresponding absolute UAV height above the local ground surface was approximately 3.0–6.5 m in conventional mature tea plantations and generally approximately 6.0–8.0 m or higher in old-tea-tree plantations, depending on the actual tree height and local terrain. Data collection was carried out in three time periods. These periods were 7:00 to 9:00 in the morning, 12:00 to 14:00 at noon, and 16:00 to 18:00 in the evening. This strategy enabled the collection of tea plantation image features under different sun heights and lighting conditions. It also made the dataset more adaptable to complex lighting environments.

Images were acquired using a DJI Mavic 2 Pro UAV equipped with a Hasselblad L1D-20c camera featuring a 1-inch sensor and a resolution of 20 megapixels. Autofocus was disabled during image acquisition. To reduce variability caused by occasional focusing errors, all images were visually inspected during quality screening, and images showing noticeable defocus or motion blur were excluded from the dataset. Before model training and evaluation and before being fed into the networks, all images were resized and padded using letterbox preprocessing to an input size of 640 × 640 pixels. The same preprocessing procedure and input resolution were applied to all compared models.

This study collected a total of 1364 UAV images of tea plantations in Yunnan. Three tea-shoot categories were annotated: tea bud, one bud and one leaf, and one bud and two leaves. The data preprocessing procedure included image denoising, lens-distortion correction, redundancy screening, and image-quality screening. Images with blurred focus, abnormal exposure, severe occlusion, or excessive duplicated spatial content were removed, resulting in 1247 high-quality images for construction of the original dataset. The resulting dataset contained approximately 4380 annotated instances, including about 1100 tea buds, about 1700 one-bud-and-one-leaf targets, and about 1600 one-bud-and-two-leaves targets. As shown in Figure 1A, the three categories exhibited a moderate class imbalance, with one-bud-and-one-leaf targets being the most frequent. After image acquisition, a manual redundancy-screening procedure was performed before dataset partitioning. Images containing more than 5% duplicated spatial content relative to another retained image were regarded as redundant samples, and the corresponding duplicated images were manually removed. This procedure was conducted on the original image pool prior to the division of the dataset in order to reduce the presence of highly similar samples caused by repeated UAV observations of the same tea-canopy regions. Therefore, the 5% criterion refers to the threshold used for manual redundancy screening rather than to an overlap ratio calculated among the training, validation, and test subsets. The study accurately annotated three target categories. These categories were tea buds, one bud and one leaf, and mature one bud and two leaves. The data preprocessing stage mainly included image denoising, lens distortion correction, and quality screening. Low-quality samples with blurred focus, abnormal exposure, or serious occlusion were removed, and 1247 high-quality images were finally obtained to build the original dataset. As shown in Figure 1, Figure 1A shows the quantity distribution histogram of tea buds, one bud and one leaf, and mature one bud and two leaves. Figure 1B shows the size distribution of labels in the width and height directions, with the center points of all labels normalized to the image center. Figure 1C shows the spatial distribution features of labels in the image coordinate system. Figure 1D shows the aspect ratio distribution of the marked bounding boxes. Figure 1E further shows the detailed features of the labels in the images [17]. After redundancy screening and image-quality control, 1247 high-quality images were retained. These images were subsequently divided into training, validation, and test sets at a ratio of 6:2:2, resulting in 748, 249, and 250 images, respectively.

Figure 1. Visualization of dataset labels.

2.2. Data Augmentation

In UAV-based tea plantation remote sensing images, the original data are often affected by several factors. These factors include changes in natural illumination, fluctuations in UAV flight attitude, and differences in imaging conditions. These factors can cause problems such as imbalanced sample distribution, large target scale variation, and limited scene diversity. These problems not only increase the difficulty of target recognition and feature extraction, but also tend to cause overfitting during the training of deep learning models, thereby limiting their generalization ability and robustness in complex environments and unknown scenarios. To improve the adaptability and robustness of the model in complex environments and enhance the sufficiency of feature learning and the diversity of data distribution, this study applied multiple data augmentation strategies to the training set based on the original dataset.

Mean blur was therefore applied by replacing each pixel value with the average value of its local neighborhood, thereby producing uniformly smoothed images and simulating the loss of local texture details. Gaussian blur was implemented by assigning Gaussian-distributed weights to neighboring pixels, which generated gradually smoothed image regions and simulated more natural optical or motion-related blur. Median blur replaced As shown in Figure 2, multiple data augmentation techniques were applied to the training set, including HSV color-space perturbation, mean blur, Gaussian blur, median blur, random cropping, and random scaling along the x- and y-axes. To simulate the imaging differences in UAV remote-sensing images acquired at different times and under different weather and illumination conditions, random perturbations were applied to the original samples in the HSV color space, namely hue, saturation, and value. By randomly adjusting the hue, saturation, and value parameters, training samples with diverse color and illumination characteristics were generated [18]. Considering that UAV image acquisition can be affected by platform vibration, changes in flight motion, and focusing variability, the collected images may exhibit different degrees and forms of blur degradation [19]. each pixel value with the median value within its local neighborhood, there