An Efficient Method for Counting Large-Scale Plantings of Transplanted Crops in UAV Remote Sensing Images

Abstract

Counting the number of transplanted crops is a crucial link in agricultural production, serving as a key method to promptly obtain information on crop growth conditions and ensure the yield and quality. The existing counting methods primarily rely on manual counting or estimation, which are inefficient, costly, and difficult to evaluate statistically. Additionally, some deep-learning-based algorithms can only crop large-scale remote sensing images obtained by Unmanned Aerial Vehicles (UAVs) into smaller sub-images for counting. However, this fragmentation often leads to incomplete crop contours of some transplanted crops, issues such as over-segmentation, repeated counting, low statistical efficiency, and also requires a significant amount of data annotation and model training work. To address the aforementioned challenges, this paper first proposes an effective framework for farmland segmentation, named MED-Net, based on DeepLabV3+, integrating MobileNetV2 and Efficient Channel Attention Net (ECA-Net), enabling precise plot segmentation. Secondly, color masking for transplanted crops is established in the HSV color space to further remove background information. After filtering and denoising, the contours of transplanted crops are extracted. An efficient contour filtering strategy is then applied to enable accurate counting. This paper conducted experiments on tobacco counting, and the experimental results demonstrated that the proposed MED-Net framework could accurately segment farmland in UAV large-scale remote sensing images with high similarity and complex backgrounds. The contour extraction and filtering strategy can effectively and accurately identify the contours of transplanted crops, meeting the requirements for rapid and accurate survival counting in the early stage of transplantation.

1. Introduction

Crop transplantation is a widely adopted agricultural practice for cultivating vegetables, flowers, tobacco, and other crops. It offers multiple advantages, including seed conservation, higher seedling survival rates, improved land utilization, reduced weed competition, enhanced crop yield and quality, and easier mechanization. These benefits collectively make transplantation a vital means of ensuring the efficiency and sustainability of large-scale agricultural production [1,2].Crop transplantation counting refers to the accurate tallying of the number of surviving crops after transplantation. It is crucial for ensuring crop yield and quality and serves the following purposes: (1) monitoring the growth of crops after transplantation to ensure stable crop yields; (2) enabling the government or agricultural management departments to better formulate relevant policies, such as agricultural subsidy policies; (3) aiding agricultural departments and farmers in land use planning to ensure the efficient use of farmland resources; and (4) more accurately predicting the yields of transplanted crops, enabling agricultural producers to arrange production plans and resource allocation to meet the market demand [3,4,5].Traditional crop counting methods often rely on manual labor or estimation, which are labor-intensive, inefficient, and prone to inaccuracies [6]. Such limitations may lead to imprudent decisions by agricultural managers, ultimately causing economic losses. Precise statistics of early survival rates of transplanted crops is an urgent issue that needs to be addressed for the development of large-scale and refined agriculture.

UAVs, with their flexible deployment and precise positioning capabilities, have been the preferred method for the rapid acquisition and interpretation of high-definition imagery in crop statistics in recent years. Integrating UAV platforms with artificial intelligence technology enables solving complex challenges in precision agriculture [7,8]. Image data obtained from UAVs can provide more detailed and rich vegetation information [9,10], making automatic plant counting from high-resolution UAV remote sensing images feasible [11,12,13]. However, when using UAVs to capture remote sensing images, different crops, as well as grasslands, trees, roads, etc., are mixed together, making it difficult to directly count the number of plants for a single crop. To accurately count the number of transplanted crops in UAV remote sensing images, it is necessary to recognize and analyze fields and crop plants at two different scales, segment the planting area, reduce background interference, and improve the algorithm recognition accuracy. The flying altitude of UAVs not only affects the detail and accuracy of image reconstruction but also affects the processing efficiency and the accuracy of algorithm detection and recognition [14,15,16]. Higher flight altitudes improve data acquisition efficiency but compromise feature resolution and sacrifice detail, limiting target characterization; conversely, lower altitudes enhance detail capture at the cost of increased image volume and reduced processing efficiency. In crop identification, some algorithms reduce the flying altitude of UAVs to obtain richer details for crop detection in UAV remote sensing images or crop small-scale remote sensing images from large-scale ones under limited computing resources [17], leading to incomplete plant contours, repeated counting, and low statistical efficiency. Some crop counting methods based on multi-spectral and hyperspectral data [18] also have limitations due to high collection costs, complex data processing, and difficulties in data integration. It is necessary to optimize the existing methods to achieve precise recognition and counting of fields and plants at different scales, reduce the cost of data discrimination, and improve technical efficiency and accuracy, providing technical support for large-scale and precise production of agriculture. Therefore, the objectives of this paper are as follows: (a) to design and implement an image semantic segmentation framework based on lightweight and attention mechanisms to achieve accurate segmentation of transplanted crop planting areas in complex farmland scenes; (b) to propose a method for removing the background of transplanted crops, effectively eliminating background information such as soil and plastic mulch; and (c) to establish an efficient method for extracting and filtering the contours of transplanted crops, achieving precise counting of transplanted crops without the cost of data annotation.

2. Related Work

The segmentation of farmland boundaries and crops is crucial for the automatic navigation of agricultural machines and precision agriculture [19,20]. In farmland boundary segmentation, image segmentation was utilized to extract navigation routes and deploy them on Edge Computing devices to provide the possibility of automatic navigation for the field operation of agricultural machines [21]. The high-resolution cropland extraction model HRRS-U-Net [22] was introduced, which can maintain a high resolution in the whole network to generate accurate cultivated land boundaries, as well as Residual Learning (RL) and a Channel Attention Mechanism (CAM), to extract deeper discriminative features. Semi-supervised semantic segmentation [23] enables farmland segmentation with limited labeled data. The semantic segmentation network based on deep edge enhancement [24] and a hybrid segmentation using particle swarm optimization and a fully convolutional network [25] can achieve the segmentation of farmland in satellite images. Crop segmentation is widely used in the field of agriculture. Segmenting diseased crop leaves provides the foundation for the automatic detection of early crop diseases [26]. Through the segmentation of plant regions and biomass regions, the number of seedlings can be detected and the biomass of the crop can be estimated [27]. The segmentation of weeds and crops provides a necessary machine vision foundation for agricultural-machinery-based pesticide and fertilizer spraying, weed removal, and crop harvesting operations [28,29,30], which helps to improve the operational efficiency and accuracy of agricultural machinery. Deep-learning-based semantic segmentation provides a fast, accurate, and automated approach for farmland boundary and crop extraction, demonstrating good application effectiveness in complex agricultural scenarios.

With the application of more deep learning technologies in the agricultural field, there have been an increasing number of crop counting algorithms and datasets based on deep learning [31]. In crop counting based on object detection, Lin et al. [32] proposed an improved YOLOv5s combined with the DeepSort tracking algorithm to count citrus fruits, addressing the issue of low counting accuracy caused by shadows and lighting factors in orchard videos. Ma et al. [33] used an upgraded version of the YOLOv7-tiny algorithm to achieve automatic yield estimation and harvesting in small-scale apple orchards in natural environments. Ba et al. [34] proposed a fast and robust plant counting method for multiple crop types in equidistant and overlapping plant counting. In density-estimation-based crop counting, Lin et al. [35] achieved better results in lychee flower density mapping and counting using a multi-column convolutional neural network compared to the tested object detection counting methods. Ba et al. first proposed the RiceNet [36] model based on an attention mechanism, multi-scale feature fusion, and positive–negative loss to generate plant density maps, which is capable of rice plant counting, localization, and sizing, and a new UAV-based rice counting dataset was proposed to validate the effectiveness of the method. In subsequent research, they designed a novel deep learning network called RPNet [37] based on an attention mechanism to generate high-quality plant density maps and achieve accurate rice plant counting using the multi-supervision loss function RPloss. In crop counting based on segmentation and detection, Kuznichov et al. [38] proposed a data augmentation method for plant segmentation and counting tasks, which preserved the geometric structure of the objects to make the dataset resemble the imaging of plants in real agricultural scenes as closely as possible. Xin et al. [39] implemented the segmentation of wheatear based on deep learning and texture features and wheat grain counting on wheatear segmentation results. Fruit crop counting based on segmentation and detection provides a basis for thinning operations in berry crops [40], Du et al. [41] introduced an adaptive feature fusion method based on the Swin Transformer and proposed the AS-SwinT model, which can achieve grape segmentation and berry counting.

Object-detection-based crop counting performs well in fruit detection scenarios with low crop density and a single crop variety. However, it has lower accuracy in large-scale scenes with high density and complex background information. In such cases, the detection difficulty is reduced and the detection accuracy is improved by cropping the image to reduce the number of crops in a single image [18,42]. Density-estimation-based crop counting methods can predict the quantity of high-density crops in large-scale scenes and demonstrate good performance in predicting the quantity of crops such as wheat and maize. However, density estimation algorithms cannot obtain the specific positions and sizes of targets, which affects further crop phenotype analysis and precision agriculture research. Crop counting methods based on segmentation and detection can perform further object detection based on segmentation results, effectively reducing background interference and improving detection accuracy, but the image segmentation and object detection require a large amount of data annotation work, which incurs significant manual labor costs.

3. Materials and Methods

Transplanted crops are planted in a more orderly manner with uniform row and plant spacing, which relatively lowers the detection difficulty. Counting methods based on object detection have shown good application in the statistical analysis of a small number of transplanted crops [43]. However, due to the vast planting quantities of some transplanted crops, the existing deep-learning-based algorithms have lower processing efficiency. Combining traditional image processing methods can often reduce the amount of data processing and improve processing efficiency, playing a positive role in the counting of transplanted crops. For the counting of single transplanted crops in UAV remote sensing images with complex background information and a mixture of multiple crop planting areas, first, based on DeepLabV3+ [44], a separable convolutional attention network MED-Net is proposed by combining MobileNetV2 and ECA-Net [45]. Utilizing depthwise separable convolution (DSC) can not only achieve the effect of regular convolution but also greatly reduce the number of parameters, and MobileNetV2 is introduced to replace the Xception network to speed up the model inference speed. At the same time, ECA-Net is added to the DeepLabV3+ network coding structure to enhance the representation of important features during feature extraction, which realizes the segmentation of the farmland in the UAV image. Secondly, the HSV color space is used to generate transplanted crop background color masking, and the background, such as soil and mulching film, is removed to obtain the image that only retains the transplanted crops. Finally, the image that only retains the transplanted crops is subjected to Gaussian filtering, morphological denoising, and opening operation denoising to extract the transplanted crop contours. The extracted transplanted crop contours are filtered and counted to finally count the number of plants. This paper uses tobacco plants cultivated through large-scale transplanting as the experimental subject. The algorithm workflow proposed in this paper is shown in Figure 1, which includes stages such as data acquisition, model training, background removal, and transplanted crop counting.

As shown in Figure 1, the specific implementation process of the method is as follows: (1) stitch and crop the UAV image to obtain the experimental data; (2) compress and annotate the image data collected by the UAV, input the annotated images into the MED-Net network for training to obtain a segmentation model, and use the segmentation model to acquire the segmentation image of the transplanted crop planting area; (3) convert the segmentation image from the RGB color space to the HSV color space, and adjust the thresholds of the H, S, and V components through experimental analysis to obtain the color masking image of the transplanted crops; (4) overlay the color masking image with the segmentation image to obtain the background removal image of the transplanted crops; (5) extract and filter the contours from the background removal image to obtain the contour information of the transplanted crops; and (6) automatically count the number of transplanted crops based on the contour information and overlay the contour information onto the segmentation image for display.

To obtain experimental results and perform qualitative and quantitative analyses, the hardware setup used in this study includes an Intel(R) (Intel Corporation, Santa Clara, CA, USA), Xeon(R) Silver 4210R CPU @ 2.40 GHz (Intel Corporation, Santa Clara, CA, USA), NVIDIA GeForce RTX 3060 (NVIDIA Corporation, Santa Clara, CA, USA), and 16 GB RAM (Samsung Electronics Co., Ltd., Seongnam, Gyeonggi-do, Republic of Korea). The operating system is Windows 10 LTSC 2021. The software environment consists of PyCharm 2021, CUDA 10.0.130, and cudnn 7.4.1.5. The model training and testing platform was built using Python 3.6.13, PyTorch 1.2.0, Torchvision 0.4.0 framework, OpenCV-Python 4.1.2.30, and NumPy 1.19.5.

The quantitative evaluation of segmentation performance was conducted through a multi-metric framework, including Mean Intersection over Union (mIoU), to assess spatial overlap accuracy between predictions and ground truth, accuracy for pixel-wise classification correctness, Mean Average Precision (mAP) to evaluate precision–recall balance across probability thresholds, and Mean Precision (mPrecision), which quantifies the true positive detection rates in positively predicted regions. This synergistic combination of complementary metrics enables rigorous cross-algorithm comparison, ensuring comprehensive characterization of model performance in agricultural scene segmentation tasks.

3.1. Study Area and Data

3.1.1. Study Area Introduction

To validate the effectiveness of the proposed method, combined with the existing experimental conditions, this paper takes tobacco, a large-scale transplanted crop, as the experimental object. The experimental areas are located in the core tobacco cultivation demonstration areas of Kunming City, Yunnan Province. The specific locations and topography of the research areas are shown in Figure 2, where (a) represents a map of China’s administrative divisions, with the blue highlighted area indicating the geographic location of Yunnan Province within China; (b) represents a map of Yunnan Province, where pink polygons delineate the administrative boundaries of Shilin County and Xundian County, respectively; (c) and (d) represent topographic maps of Xundian and Shilin County, corresponding to the pink region in (b). All sub-figures include a scale bar and elevation legends, and the background color gradient represents the elevation distribution. Among them, Experimental Area 1 is situated in Tangdian Town, Xundian County, Kunming City (between longitude 102°50′ and 102°59′ and latitude 25°40′ and 25°54′). It has an average elevation of 2056 m, an annual average temperature of 14.4 °C, a low-latitude plateau mountainous monsoon climate, and an average annual precipitation of 950 mm. The main crops grown in this area include tobacco, maize, and rice, with a tobacco planting area of 8000 hectares. Experimental Area 2 is located in Xijiekou Town, Shilin County, Kunming City (between longitude 103°29′ and 103°28′ and latitude 24°21′ and 24°51′). It has an average elevation of 2038 m, an annual average temperature of 15.6 °C, a low-latitude plateau mountainous monsoon climate, and an average annual precipitation of 954 mm. The main crops grown in this area include tobacco, maize, rice, and ginseng fruit.

3.1.2. Dataset Construction

The data acquisition equipment is the DJI Matrice M300 RTK UAV, and the data were acquired from the tobacco planting areas in Xundian and Shilin County, Kunming City, Yunnan Province, China, respectively, corresponding to about 40 days after the transplanting of tobacco seedlings. As shown in Figure 3, the image data contain non-tobacco planting areas such as maize fields, rice fields, unplanted farmland, grasslands, woods, roads, buildings, etc., which together constitute the complex background of the UAV remote sensing images of tobacco fields.

According to the different tasks of UAV monitoring, flight plans for different growth stages of tobacco are formulated. In this paper, data collection was conducted during the early growth stage of tobacco, and data collection work was performed at different flight altitudes. The flight data were stitched using DJI Terra, and the parameters for data acquisition and stitching are shown in

Table 1. Data acquisition, stitching parameters.

Parameters	Options	Values
Acquisition	Flight altitude	40 m, 50 m, 80 m, 100 m
	Camera lens	ZENMUSE P1, 3:2 (8192 × 5460)
	Ground Sampling Distance (GSD)	0.5 cm/pixel, 0.625 cm/pixel, 1 cm/pixel, 1.25 cm/pixel
	Flight parameters	Flight speed: 15 m/s, overlap: 70% (Along-track), 60% (Across-track)
Stitching	Software	DJI Terra
	Central Meridian Longitude (CML)	Kunming (02:42E 25:03N)
	Geodetic Coordinate System	CGCS2000

.

Meanwhile, to reduce the data processing load, the data were preprocessed using Global Mapper software; the stitched data were cropped according to the standard of 5600 × 3800 pixels. After cropping, the data were downsampled by a factor of four to obtain data with dimensions of 1400 × 950 pixels. The composition of the UAV tobacco field segmentation dataset is summarized in

Table 2. Composition of the UAV Tobacco Field Segmentation Dataset.

Options		Values
Original data	Flight altitude	Region	Number
	40 m	Shilin, Xundian	194
	50 m	Shilin, Xundian	303
	80 m	Shilin, Xundian	180
	100 m	Xundian	87
Filter data			208
Enhance data	Method	Operation
	Augmentor	rotate	20
		flip	40
		skew	40
		scale	40
		Random distortion	20
		shear	20
		crop	40
	Mosaic		172
Dataset	Training set		540
	Test set		60

.

Through the analysis of the original image data, it was found that the UAV images of tobacco fields mainly include the tobacco field area, as well as background areas such as roads, villages, grasslands, forests, maizefields, and fallow land. Some data exhibit high similarity, which is not conducive to the extraction of effective features during the model training process. Therefore, 208 images with diversity were selected from the original data in Table 1, and the selected data were annotated using the image annotation tool Labelme. To enhance the diversity and increase the quantity and balance of the tobacco field segmentation dataset, data augmentation was performed using the Mosaic and Augmentor tools. The Mosaic data augmentation involved randomly selecting four original images, adjusting their size and position, and then stitching them together to create new data. The Augmentor tool was used to perform data augmentation through operations such as rotation, scaling, cropping, and deformation. After the data augmentation was completed, the dataset was divided into training and testing sets in a 9:1 ratio. The training set consisted of 540 images, while the testing set contained 60 images. Partial tobacco field images and labels are shown in Figure 4.

As shown in Figure 4, during the dataset construction process, only the transplanted crop planting areas were annotated. This methodology was adopted because minimal cropping was performed and image compression was avoided during the experiment to preserve the integrity of transplanted crop detail features and contours in UAV large-scale remote sensing images as much as possible.

3.2. Farmland Segmentation Model

The DeepLabV3+ model can be used to achieve farmland segmentation. However, it may have difficulty in accurately identifying and segmenting complex field boundaries, as well as distinguishing similar transplanted crop planting areas. This limitation hinders the effectiveness in meeting the requirements of transplanted crop planting area segmentation. Therefore, improvements are needed to address the existing issues of segmentation models in transplanted crop planting area applications.

3.2.1. DeepLabV3+

The DeepLabV3+ model combines the Atrous Spatial Pyramid Pooling (ASPP) and the Encoder–Decoder structure and introduces dilated convolutions to expand the receptive field while preserving resolution, enabling the adjustment of feature resolution extracted by the Encoder to balance accuracy and computational cost. Additionally, DeepLabV3+ utilizes the Xception model for segmentation tasks and applies DSC in the ASPP and Decoder module, resulting in a faster and stronger Encoder–Decoder network. The network structure is illustrated in Figure 5.

As shown in Figure 5, in the Encoder, the initial features of the input image are extracted using the Xception backbone network. The shallow features are then transferred to the Decoder, while the deep features are passed to the ASPP feature enhancement network. In the ASPP module, different dilated convolutions with varying rates are applied to the deep features for feature enhancement, and the results are concatenated. The enhanced features are further compressed and adjusted in channel dimensions using 1 × 1 convolutions before being transferred back to the Decoder. In the Decoder, the channel dimensions of the shallow features are adjusted using 1 × 1 convolutions. Then, the adjusted shallow features are concatenated with the upsampled output of the enhanced features from the Encoder. The concatenated features undergo two DSC operations to obtain the final effective feature layer. Finally, a 3 × 3 convolution is applied to adjust the channel dimensions of the final effective feature layer, and upsampling is used to resize the output prediction to the same size as the input image.

3.2.2. MED-Net Construction

Due to the inherent structure of the DeepLabV3+ model, it has some limitations: (1) pooling operations during the Encoder feature extraction process compress the spatial dimensions of the input data, resulting in the loss of feature information. (2) Dilated convolutions blur the boundaries of small objects, leading to the loss of boundary and fine detail features, which results in poor segmentation performance for small-sized object boundaries. (3) The Xception feature extraction network utilizes depthwise convolution (DW) followed by pointwise convolution (PW) operations, combined with ResNet’s skip connections, which increases the complexity of the Xception network. Moreover, the scattered computation process and lower computational efficiency of the DeepLabV3+ model increase the training difficulty and require higher hardware requirements.

To achieve segmentation of the planting area of a single transplanted crop in UAV remote sensing images, this paper proposes a separable convolution attention semantic segmentation network called MED-Net, which combines MobileNetV2 and ECA-Net, based on the DeepLabV3+ model; it enables the segmentation of farmland in UAV remote sensing images. The  MED-Net structure is shown in Figure 6.

As shown in Figure 6, the Encoder part of MED-Net consists of MobileNetV2, ECA-Net, and ASPP. First, MobileNetV2 replaces the Xception network in DeepLabV3+, effectively simplifying the feature extraction network by reducing the number of layers, minimizing parameter size, lowering model complexity, and improving computational efficiency. Secondly, the high-level semantic features obtained from MobileNetV2 are passed to the ECA-Net, which applies adaptive attention weighting to the channel dimension of the feature maps; this helps the model to better understand and utilize information from different channels, enhancing the model’s focus on different channels. Then, different-rate dilated convolutions in ASPP are used to extract enhanced features from deep layers, which are then concatenated. Finally, the compressed and adjusted feature maps obtained through 1 × 1 convolution are transferred to the Decoder. In the Decoder part, the low-level semantic features obtained from MobileNetV2 in the encoding stage are first passed through the ECA-Net to calculate attention weights for each feature channel. This suppresses less important feature channels and increases the weights of important feature channels. Then, the weighted features are adjusted by 1 × 1 convolution to match the feature channel dimensions and are concatenated with the enhanced features from the encoding stage, which have undergone upsampling. After that, 3 × 3 convolution is applied to adjust the feature channels, followed by bilinear interpolation upsampling to obtain output prediction results with the same size as the input image.

The MobileNet series of lightweight neural networks was first introduced by the Google team in 2017. Among them, MobileNetV2, proposed in 2018, achieved better accuracy while reducing the model parameter size by 20% compared to MobileNetV1. The main improvements in MobileNetV2 compared to MobileNetV1 can be summarized in the following two points:

(1) Inverted Residuals: Compared to the residual structure, the inverted residual structure first uses a 1 × 1 convolution to increase the number of channels, then extracts features through a 3 × 3 DW, and finally reduces the number of channels using a 1 × 1 convolution. The entire process can be described as “expand–convolve–compress” and has an hourglass shape. The 3 × 3 standard convolution is replaced by a 3 × 3 DW. The first two ReLU activation functions are replaced with ReLU6, which is defined by the Equation (1). The specific structural differences can be observed in Figure 7.

$$y=\mathrm{Re}LU6\left(x\right)=min\left(max\right(x,0),6)$$

(1)

The inverted residual structure used in MobileNetV2 deepens the network by first increasing and then reducing the number of channels, which enhances gradient propagation and feature representation capabilities. This helps the network to better capture details and features in images. Additionally, ReLU6 provides better non-linear expression capability compared to ReLU. It truncates the input values larger than 6 to 6 directly, reducing the computational complexity of the network and improving performance. By replacing the first two ReLU activation functions with ReLU6, MobileNetV2 can improve computational efficiency while maintaining a certain level of accuracy.

(2) Linear Bottlenecks: The linear bottleneck structure refers to mapping the high-dimensional space to a low-dimensional space and using linear activation in the final convolutional layer to reduce the loss of low-dimensional feature information. In MobileNetV2, the linear bottleneck structure consists of two convolutional layers. The first 1 × 1 convolution is used for dimensionality reduction, reducing the number of input channels to 1/6 of the original size. The second 3 × 3 DW is used for feature extraction and expansion, expanding the number of channels back to the original size. Additionally, a linear activation function is applied in the final layer to reduce the loss of low-dimensional feature information. This reduces computational complexity and the number of model parameters, improving the computational efficiency of the network while maintaining a certain level of accuracy.

ECA-Net enhances the representation ability of different channel features by learning a weight vector to weigh them, which has tremendous potential for improving model performance. Figure 8 shows the network structure of ECA-Net. First, the input features are processed using global average pooling (GAP) to obtain all the uncompressed features without dimensionality reduction. Then, a fast one-dimensional convolution with shared weights and a size of k is applied to learn features from all the uncompressed features. The weights for each channel are generated using the $Sigmoid$ function to determine their proportion. Finally, combine the original input features with the learned channel weights to generate feature representations enhanced by channel attention. ECA-Net avoids dimensionality reduction and proposes a local cross-channel interaction strategy and an adaptive selection of one-dimensional convolution kernel size. It effectively captures cross-channel interaction information while significantly reducing model complexity.

In ECA-Net, only the information interaction between the feature channel $y_i$ and its k neighboring channels is calculated. The channel attention weight values $w_i$ are generated for each channel using the function $\sigma$, and the formula for calculating the channel attention weights is as in Equation (2):

$$w_i=\sigma (\sum _{j=1}^k{\omega }^j{y}_i^j),y_i^j\in {\mathsf{\Omega }}_i^k$$

(2)

where ${\mathsf{\Omega }}_i^k$ denotes the set of k neighboring channels of $y_i$.

In order to further enhance the feature extraction performance of ECA-Net, the information interaction between channels is achieved by using a $1D$ convolution $C1D$ with a kernel size of k. By sharing the same set of learnable parameters, Equation (2) can be rewritten as shown in Equation (3).

$$w=\sigma \left(C1D_k\left(y\right)\right)$$

(3)

Since ECA-Net aims to appropriately capture local cross-channel interactions, it is necessary to determine the coverage of local cross-channel interactions, which is the kernel size k of the $1D$ convolution, where k is proportional to the channel dimension C. Given the size of the channel dimension C, the calculation of the kernel size k can be expressed as an adaptive function as in Equation (4):

$$k=\psi \left(C\right)={\left|{\displaystyle \frac{{log}_2\left(C\right)}{r}}+{\displaystyle \frac{b}{r}}\right|}_{odd}$$

(4)

where ${\left|t\right|}_{odd}$ denotes the nearest odd number to t, $r=2,b=1$. Through mapping $\psi \left(C\right)$, high-dimensional channels have longer interactions, while low-dimensional channels undergo shorter interactions by utilizing non-linear mapping.

3.2.3. Establishing Background Color Masking Image for Transplanted Crops

After completing the farmland image segmentation based on MED-Net, most of the farmland areas unrelated to the counting task could be removed, resulting in the segmentation map of the transplanted crop planting area as shown in Figure 9a. However, the existence of small-sized background areas such as mulching film, soil, and weeds between transplanted crops can interfere with the subsequent counting of transplanted crops, and MED-Net struggles to separate these areas from the transplanted crops. Therefore, this paper constructs a method for removing small-sized background areas to achieve fine segmentation of the transplanted crops and background. The steps of this method are as follows:

First, as shown in Figure 9b, convert the farmland segmentation image in RGB format to the HSV color space. Then, adjust the three components of hue (H), saturation (S), and value (V), observe the effect of removing transplanted crop backgrounds, and statistically determine the corresponding H, S, and V values to establish the threshold ranges for H, S, and V. This process obtained the transplanted crop color masking image as shown in Figure 9c.

Secondly, overlay the color masking image onto the original image, utilizing the masking effect to mask the background areas of transplanted crops in the images. Define two input images (arrays) $src1$ and $src2$ with the same size, respectively, and $dst$ denotes the output array with the same size as the input arrays, where all elements default to 0, i.e., a black color masking image; then, there is

$$\left\{\begin{array}{c}dst\left(I\right)=src1\left(I\right)\wedge src2\left(I\right),mask\left(I\right)\ne 0\\ dst\left(I\right)=0,mask\left(I\right)=0\end{array}\right.$$

(5)

Finally, perform an “AND” operation between the color masking image and the farmland segmentation image using Equation (5). In the white regions of color masking image $mask\left(I\right)\ne 0$, display the corresponding content from the farmland segmentation image. In the black regions of color masking image $mask\left(I\right)=0$, mask the corresponding regions in the farmland segmentation image. This process enables the separation of transplanted crops from the background, resulting in the transplanted crop image with the background removed, as shown in Figure 9d.

By further utilizing the HSV color masking image to remove the background of transplanted crops on the basis of the MED-Net farmland segmentation image results, this approach addresses the limitations of MED-Net in removing small-sized background regions such as soil and mulching film between transplanted crops. Additionally, it also addresses the problem of HSV being unable to remove grasslands, woods, and other non-transplanted crop planting areas that have similar colors to transplanted crops.

3.3. Methods for Transplanted Crops Counting

The counting of transplanted crops is mainly achieved through the extraction and statistics of transplanted crop contours. The following are the specific implementation steps:

(1) Image grayscale conversion: By converting the image to grayscale, the original data size is reduced, which improves data processing speed and reduces interference while avoiding banding distortion. The grayscale value Y for each component is obtained using Equation (6).

$$Y=0.229R+0.587G+0.114B$$

(6)

where R, G and B denote the corresponding three channel components in the RGB color space.

(2) Gaussian filtering for image smoothing: This process is performed to smooth the image by reducing pixel variations. It is used to remove noise from the image before performing edge detection, thus improving image quality. Let $S(u,v)$ be the input image and $G(u,v)$ be the result after applying a two-dimensional Gaussian function. The value of the central pixel in the template is replaced by the weighted average grayscale value of the pixels in its neighborhood. The two-dimensional Gaussian function is defined as shown in Equation (7).

$$G(u,v)={\displaystyle \frac{1}{2\pi {\sigma }^2}}{e}^{-(u^2+v^2)/\left(2{\sigma }^2\right)}$$

(7)

(3) Image binarization: A grayscale threshold value $T_1$ is set. When $src(x,y)>T_1$, $dst(x,y)=255$. When $src(x,y)\le T_1$, $dst(x,y)=0$.

By reducing the dimensions of the image through grayscale conversion, Gaussian filtering, and binarization, the variation between pixels in the image is reduced. This process helps to separate the transplanted crops from the background, highlighting the contours of the plants. The resulting image after processing is shown in Figure 10a.

(4) Image morphology and opening operation: First, a 3 × 3 cross-shaped structuring element M is defined as shown in Equation (8). The structuring element M is then used for morphological operations on the image, returning a structuring element with the specified shape and size. Next, the opening operation is performed, which involves erosion followed by dilation. This operation separates the transplanted crops from the surrounding background, removing irrelevant white pixel interference areas around the plants while preserving the rest. The result is shown in Figure 10b.

$$M=\left[\begin{array}{ccc}0& 1& 0\\ 1& 1& 1\\ 0& 1& 0\end{array}\right]$$

(8)

(5) Contour extraction: Output the outer contour information of the transplanted crops by compressing the elements in the horizontal, vertical, and diagonal directions and retaining only the endpoint coordinates in those directions; the outline of the transplanted crops is drawn by the contour information, as shown in Figure 10c.

(6) Contour filtering: Contour filtering consists of two operational steps, namely filtering based on the Euclidean distance between contour centroids and filtering based on the pixel area within the contour. The centroid of a contour refers to the geometric center of a closed contour. The coordinates of the centroid are determined by calculating the moments of the contour’s bounding rectangle, as shown in Equation (9):

$$x_i={\displaystyle \frac{M_{10}}{M_{00}}},y_i={\displaystyle \frac{M_{01}}{M_{00}}}$$

(9)

where $x_i$ and $y_i$ denote the centroid’s horizontal and vertical coordinates, respectively. $M_{00}$ denotes the contour’s 0th-order moment, and $M_{01}$ and $M_{10}$ denote the contour’s 1st-order moments.

First, contour filtering is performed based on the Euclidean distance between contour centroids. Set the Euclidean distance between the centroids of contours $C_i$ and $C_j$ as $d_{i,j}$ and the centroid distance threshold as $T_d$; the pixel areas of the contours are denoted as $A_i$ and $A_j$, respectively. When the Euclidean distance between the centroids of two contours is greater than or equal to the distance threshold $T_d$, the contour information of both contours is retained. When it is less than the distance threshold $T_d$, the pixel areas of the two contours are compared and the contour with the larger pixel area is retained. The contour results after filtering based on the centroid distance are shown in Figure 10d. Then, set the threshold values $T_{min}$ and $T_{max}$ for the pixel areas of transplanted crop contours to eliminate contours that are too small or too large. When the pixel area of a transplanted crop contour satisfies $T_{min}<A_i\le T_{max}$, retain the information of contour $C_i$; otherwise, delete the information of contour $C_i$, and the contour results after filtering based on contour pixel area are shown in Figure 10e. The specific filtering strategy is shown in Algorithm 1.

(7) Counting transplanted crops: By analyzing the filtered contour information, the coordinates of the minimum bounding rectangle around each transplanted crop contour are determined. These coordinates are then used to draw rectangular bounding boxes around the plant locations in the image of the farmland. The counting of transplanted crop number results are shown in Figure 10f.

Algorithm 1: Transplanted Crop Contour Filtering Algorithm

4. Results and Discussion

A DJI ZENMUSE P1 35 mm lens is used for data acquisition, and the corresponding Ground Sample Distance (GSD) is calculated via the formula $GSD$ = H/80, where H denotes the UAV flight altitude. The experimental part of the experiment is conducted with the data acquired at H = 100 m as the test sample, corresponding to $GSD$ = 1.25 cm/pixel.

4.1. Background Removal of Tobacco Plants

4.1.1. MED-Net Based Tobacco Field Segmentation Experiments

To achieve tobacco field segmentation at the plot scale by removing non-tobacco planting areas such as roads, woods, and non-tobacco crops from the UAV remote sensing images, we built MED-Net using the aforementioned experimental software and hardware environment. The training and testing sets were allocated in a 9:1 ratio for model training and testing. The data allocation and some parameters for model training are shown in

Table 3. Model training parameters.

Parameters	Values
Total data	600
Training Set (90%)	540
Testing Set (10%)	60
batch size	4
learning_rate	$7\times {10}^{-3}$
epochs	112
momentum	0.9
Weight_decay	$1\times {10}^{-4}$
optimizer	SDG

.

To compare and analyze the performance indicators of MED-Net, the PSPnet, Unet, DeeplabV3+_Xception, and DeeplabV3+_MobileNetV2 models were introduced. The learning rate was set to $7\times {10}^{-3}$, batch size to 4, momentum parameter to 0.9, and weight decay to $1\times {10}^{-4}$ for all the models. The variations regarding the loss function values for the five models with respect to epoch were obtained as shown in Figure 11.

When the loss function performance indicators of the five models no longer improved after approximately 112 training iterations, a qualitative comparative test analysis was conducted to evaluate the tobacco field segmentation performance of these models. The comparative results of the tobacco field segmentation are shown in Figure 12.

As shown in Figure 12, Figure 12a denotes the original image, Figure 12b denotes the corresponding label file for the image, and Figure 12c–f show the segmentation results of different algorithms denoted by green regions. Field (i) mainly consists of large unplanted farmland, greenhouses, and maize planting areas, with tobacco fields occupying only a small fraction of the UAV image. MED-Net is able to accurately segment the tobacco fields, while DeeplabV3+_MobileNetV2 can segment larger tobacco field areas but fails to capture smaller regions. The remaining three algorithms failed to achieve tobacco field segmentation. In field (ii), all the compared algorithms can achieve complete segmentation of tobacco fields, but MED-Net shows superior segmentation results for the ridges in the tobacco fields. In field (iii), where the tobacco field areas are scattered and relatively small, interspersed with different crop planting areas, grasslands, and woods, MED-Net can achieve relatively complete segmentation of the tobacco field areas compared to the other algorithms. In field (iv), PSPnet, DeeplabV3+_Xception, and DeeplabV3+_MobileNetV2 can segment the tobacco field areas, but the segmentation of the ridges between tobacco fields is poor, with incomplete boundary segmentation in some areas. Unet shows incorrect segmentation, and MED-Net achieves effective segmentation of the ridges between tobacco fields while ensuring complete segmentation boundaries. In field (v), there are a large number of maize planting areas, which have similar mulching film coverings and planting methods as the tobacco field areas; the differences between these two types of fields in UAV remote sensing images are relatively small, and all the tested algorithms can correctly segment the tobacco field areas, but MED-Net demonstrates superior segmentation results.

To validate the effectiveness of the proposed modules, an ablation study was conducted comparing DeeplabV3+_Xception, DeeplabV3+_MobileNetV2, and MED-Net, with the experimental results presented in

Table 4. Ablation Study on Lightweight Backbones and Attention Mechanisms.

Model	DeeplabV3+_Xception	DeeplabV3+_MobilenetV2	MED-Net
Model Size (MB)	209	22.4	22.3
Number of Test Images	60	60	60
Test Execution Time (s)	10.910	9.186	9.849
Inference Speed (img/s)	5.50	6.53	6.09
mIoU (%)	91.45	94.79	96.49
mPA (%)	95.52	97.32	98.2
Accuracy (%)	95.56	97.34	98.36

.

As show in Table 4, MobileNetV2 reduces the baseline model size by 89.3% to 22.4 MB and accelerates the inference speed by 18.7% to 6.53 images per second while improving the segmentation accuracy by 3.34% in mIoU compared to the Xception-based architecture. The integration of ECA-Net enhances MobileNetV2 through channel attention, achieving improvements of 1.70% in mIoU, 0.88% in mAP, and 1.02% in accuracy. MED-Net combines MobileNetV2’s architectural efficiency with ECA-Net’s feature refinement, attaining 96.49% mIoU with a compact 22.3 MB model, proving that hybrid lightweight–attention architectures can achieve high-precision segmentation.

The mIoU test results in Figure 13 indicate that MED-Net achieved an mIoU of 96.76% for the tobacco field region and 96.21% for the background region. Compared to DeeplabV3+_MobileNetV2, MED-Net achieved a 2.05% higher mIoU for the tobacco field region and a 2.42% improvement for the background region compared to DeeplabV3+_MobileNetV2. It outperformed the other tested models and had the highest level of overlap with the ground truth in terms of segmentation results.

To further investigate the accuracy and stability of the model segmentation, accuracy tests were conducted on different quantities of test sets. The results are shown in Figure 14, and it can be observed that MED-Net consistently achieved the highest accuracy across the tested quantities of test sets. The accuracy of MED-Net showed minimal fluctuations, indicating a relatively stable performance. This consistency reflects the model’s ability to learn more general features and patterns, enabling it to have good generalization capabilities even with variations in sample distribution.

The comparison in

Table 5. Performance comparison results of different segmentation models.

Model	mIoU/%				mAP/%				mPrecision/%
	$T\text{\_}F$	$B\text{\_}G$	$Total$		$T\text{\_}F$	$B\text{\_}G$	$Total$		$T\text{\_}F$	$B\text{\_}G$	$Total$
PSPnet	86.46	84.64	85.55		91.51	93.05	92.28		93.31	91.01	92.16
U-net	89.49	88.27	88.88		92.5	96.0	94.25		96.06	92.04	94.05
DeeplabV3+_Xception	91.62	90.21	90.91		96.0	94.01	95.21		95.19	95.35	95.27
DeeplabV3+_MobileNetV2	94.71	93.79	94.25		96.0	98.02	97.0		97.09	97.07	97.08
MED-Net	96.76	96.21	96.46		98.4	98.0	98.2		98.14	98.32	98.23

Where $T\text{\_}F$ denotes the result of the tobacco field area segmentation, $B\text{\_}G$ denotes the result of the background area segmentation, and $Total$ denotes the overall test result.

shows that, when replacing the backbone network Xception with MobileNetV2 in DeeplabV3+, the overall mIoU of the model is improved by 3.34%, mAP is improved by 1.79%, and mPrecision is improved by 1.81%, which verifies the improvement in the model performance by MobileNetV2 when the ECA-Net is introduced to construct MED-Net on top of the MobileNetV2-based DeeplabV3+; the overall mIoU of MED-Net is improved by 2.21%, mPA by 1.2%, and mPrecision by 1.15% compared to the MobileNetV2-based DeeplabV3+, which verifies that ECA-Net enhances DeeplabV3+’s ability to capture global contextual information across channels and suppress irrelevant information, thereby strengthening the segmentation performance of the model.

4.1.2. HSV-Based Background Removal Experiments

For the removal of small-sized background areas such as mulching film and soil between tobacco plants, on the basis of MED-Net-based tobacco field segmentation images, the optimal HSV threshold for removing the background of tobacco plants was obtained through experiments. Here, $27\le H\le 55$, $39\le S\le 172$, $113\le V\le 194$, and this threshold can effectively remove background pixels while preserving the pixel information of the tobacco plant parts, achieving the removal of small-sized backgrounds. Some processing results are shown in Figure 15, where (i) denotes the tobacco field segmentation image, (ii) denotes the result of removing the background of the tobacco plant using the HSV threshold, (iii) denotes a localized zoomed-in display of the result of removing the background of the tobacco plant, and Figure 15a–e denote the different tobacco fields.

4.2. Tobacco Plant Counting

Set the binary threshold value $T_b=85$ and apply operations such as binarization, image morphology, and opening operation to the tobacco plant image after background removal to obtain the contour image of the tobacco plant. By manually labeling the contour images of tobacco plants at a flight altitude of 100 m, the pixel Euclidean distance between tobacco plant centroids, as shown in Figure 16a, and the pixel area data of tobacco plants, as shown in Figure 16b, can be obtained.

Calculate the pixel Euclidean distances between the centroids of 1500 tobacco plants, as well as the pixel areas of these 1500 tobacco plants. Then, create a pixel distance boxplot histogram as shown in Figure 17a and a pixel area boxplot histogram as shown in Figure 17b.

By analyzing the statistical results of Figure 17, when the UAV’s flight altitude is 100 m, the pixel Euclidean distance distribution range between the tobacco plant centroids is (40, 72) pixels and the pixel area distribution range of the tobacco plant contours is (121, 978) pixels. Based on these two distribution ranges, set the centroid distance threshold value to $T_d=40$ and the contour pixel area threshold value to $121<A_i\le 978$. By processing the filtered contour information, the results of the tobacco plant counting were obtained as shown in Figure 18; the specific statistical results can be found in

Table 6. Statistical results of the number of tobacco plants.

Field	Contours			Predicted Value	False Detection	Predicted True Value	Missed Detection	Actual Value	Missed Detection Rate	Accuracy
	$U\text{\_}C$	$D\text{\_}T$	$A\text{\_}T$
a	10,691	4420	119	6152	18	6134	189	6323	2.99%	97.01%
b	9893	5823	337	3733	86	3647	168	3815	4.40%	95.60%
c	19,564	14,691	68	4805	19	4786	202	4988	4.05%	95.95%
d	7258	3774	44	3440	0	3440	57	3497	1.63%	98.37%
e	11,360	5384	78	5898	26	5872	114	5986	1.90%	98.10%

Where $U\text{\_}C$ denotes the number of contours before filtering, $D\text{\_}T$ denotes the number of contours removed through the distance threshold filter between tobacco plant centroids, and $D\text{\_}T$ denotes the number of contours removed through the area threshold filter.

.

As shown in Figure 18a, the background is relatively simple, with little background interference. However, there are areas where tobacco plants have vigorous growth, and the leaves of the tobacco plants overlap and occlude each other, making it difficult for the algorithm to separate regions with excessive overlap, resulting in missed detections. In Figure 18b, there is excessive weed growth in some areas, which interferes with the extraction of tobacco plant contours. Additionally, some tobacco plants grow slowly, and even manual observation of the images requires repeated confirmation for smaller plants, resulting in lower prediction accuracy. In Figure 18c, there are cases where tobacco plants are shaded by woods, and, when removing the background based on the HSV color space, the tobacco plants occluded by shadows are easily mistaken as background information and removed. This leads to a failure to extract the contours of the tobacco plants in the shaded areas, resulting in missed detections in the counting of tobacco plants. In Figure 18d,e, the uniform background and consistent growth patterns of tobacco plants result in significantly higher counting accuracy.

From Table 6, it can be observed that the method employed in this paper effectively filters tobacco plant contours, resulting in a relatively high accuracy in the counting of high-density tobacco plants in large-scale images captured by UAVs.

The performance of MED-Net in tobacco field segmentation, achieving an mIoU of 96.49%, aligns with advancements in lightweight network architectures integrated with attention mechanisms. For instance, Bai et al. [37] demonstrated that channel attention modules significantly improve feature discriminability in agricultural segmentation tasks, which corroborates our findings that ECA-Net enhances MobileNetV2’s ability to suppress irrelevant background features. The observed 18.7% improvement in inference speed over DeepLabV3+_Xception is consistent with the trends in UAV-based crop monitoring, where lightweight models like MobileNetV2 are prioritized for real-time processing [17,42]. Furthermore, the HSV-based background removal strategy, through adaptively adjusting color space ranges and applying morphological operations, can effectively improve the image segmentation efficiency and enhance the clarity of crop contours. This approach demonstrates promising applications in visual navigation path extraction for agricultural robots [46], as well as crop recognition and localization tasks [47,48]. The contour filtering strategy outperforms traditional thresholding methods [36] and aligns with density map approaches [37] in balancing efficiency and precision. Future integration of multi-spectral data, as suggested by Valente et al. [12], could further mitigate shadow-related false negatives, while adaptive thresholding for varying growth stages may enhance robustness across phenological phases [36]. These results collectively advance UAV-based precision agriculture, offering a scalable framework for large-scale crop management.

This study develops a counting method for transplanted crops in complex agricultural environments, systematically achieving the following objectives: (a) a transplanted crop planting area segmentation framework, MED-Net, which integrates a lightweight backbone network with an attention mechanism, has been proposed, and a large-scale UAV-based tobacco field segmentation image dataset has been constructed to validate the effectiveness of MED-Net. (b) A method based on color masking for removing the background of transplanted crops has been proposed, achieving fine segmentation of transplanted crops and small-sized background areas, preventing incomplete contours of some tobacco plants due to over-segmentation, and improving the accuracy and efficiency of transplanted crop statistics. (c) A pixel-level method for extracting and filtering the contours of transplanted crops has been developed using Gaussian filtering, morphological operations, and opening operations for denoising to extract crop contours. By employing a dual-constraint mechanism of contour area thresholds and Euclidean distance between centroids, the counting of transplanted crops has been achieved under zero-annotation conditions.

5. Conclusions

This study proposes a UAV-based framework for automated transplanted crop monitoring, with two primary objectives: (1) achieving accurate field-scale farmland segmentation and (2) enabling efficient transplanted crop counting. A lightweight MED-Net model integrating DeepLabV3+, MobileNetV2, and ECA-Net is developed to achieve accurate farmland segmentation at the field scale. The HSV-based secondary segmentation approach effectively eliminates complex background interference from soil and weeds. Combined with contour analysis, the method realizes transplanted crop counting without requiring large-scale detection model training. The experimental results demonstrate its effectiveness in reducing over-segmentation caused by device limitations while maintaining operational efficiency. Compared with object-detection-based approaches, this framework significantly reduces data annotation demands and demonstrates practical value for large-scale production management, providing a scientific foundation for breeding, cultivation, and production decisions in transplanted crop management.

6. Future Work

Firstly, a robust method should be designed to accurately crop images of transplanted crop planting areas after the background has been removed, which will reduce the computational load of the algorithm and address the issue of over-segmentation that leads to inaccuracies in transplant crop counting. Secondly, it is essential to establish pixel thresholds corresponding to the plant area and inter-plant spacing of transplanted crops based on different UAV flight altitudes and the various growth stages of the crops as this is crucial for enhancing the precision of plant counting under diverse conditions. Thirdly, a dedicated framework should be developed for calculating crop coverage rates through pixel-level analysis, which can serve as an important indicator for evaluating planting density uniformity and crop growth status. Additionally, pixel-level image processing can be utilized to advance the phenotypic analysis of transplanted crops, providing a scientific basis for crop breeding, planting management, and production decision-making. Finally, the developed method should be broadly applied to the image analysis of other crops, including the counting and planting area estimation of transplanted crops in large-scale, orderly, and uniform planting scenarios.