Leaf Area Estimation in High-Wire Tomato Cultivation Using Plant Body Scanning

Abstract

Accurate estimation of the leaf area index (LAI), a key indicator of canopy development and light interception, is essential for improving productivity in greenhouse tomato cultivation. This study presents a non-destructive LAI estimation method using side-view images captured by a vertical scanning system. The system recorded the full vertical profile of tomato plants grown under two deleafing strategies: modifying leaf height (LH) and altering leaf density (LD). Vegetative and leaf areas were extracted using color-based masking and semantic segmentation with the Segment Anything Model (SAM), a general-purpose deep learning tool. Regression models based on leaf or all vegetative pixel counts showed strong correlations with destructively measured LAI, particularly under LH conditions (R² > 0.85; mean absolute percentage error ≈ 16%). Under LD conditions, accuracy was slightly lower due to occlusion and leaf orientation. Compared with prior 3D-based methods, the proposed 2D approach achieved comparable accuracy while maintaining low cost and a labor-efficient design. However, the system has not been tested in real production, and its generalizability across cultivars, environments, and growth stages remains unverified. This proof-of-concept study highlights the potential of side-view imaging for LAI monitoring and calls for further validation and integration of leaf count estimation.

1. Introduction

In the cultivation of fruiting vegetables such as tomatoes in controlled environments (e.g., greenhouses), increasing the yield per unit area is crucial for recovering initial investments and ensuring stable farm management. Crop yield depends heavily on the production of assimilates generated through photosynthesis, that is, the overall biomass production. Therefore, several cultivation techniques have been developed to improve light interception and enhance photosynthetic activity [1].

Leaf area is a key parameter that directly affects the light interception efficiency in crops, as it determines the amount of incident light energy available for photosynthesis per unit ground area. In tomato cultivation, normalized leaf area index (LAI) is commonly used as a management indicator of canopy development. LAI is a dimensionless value representing the ratio of the total leaf surface area (m²) to one square meter of ground area within a crop canopy, and it serves as a widely accepted metric for quantifying a crop’s total light interception capacity. For instance, major tomato production models, such as TOMGRO [2] and TOMSIM [3] use LAI as a parameter to estimate light interception within the crop canopy.

In greenhouse tomato cultivation, changes in the LAI have a direct impact on both yield and fruit quality, making it a critically important indicator for monitoring crop growth and managing cultivation practices. A cultivar comparison study conducted in South Korea revealed that different LAI conditions influenced growth characteristics, such as plant height, fruit count, and fruit firmness, suggesting that proper LAI management contributes to improved productivity [4]. Furthermore, Ref. [5] analyzed yield across multiple cultivars along with canopy structure, demonstrating that LAI is a useful metric for explaining differences in yield strategies. In addition, studies related to water stress have attempted to quantitatively evaluate the impact of irrigation volume on the LAI, thereby assessing its influence on crop growth [6]. More recently, Japanese researchers have proposed a leaf area management approach aimed at maximizing dry matter production, reporting that appropriate adjustment of LAI can lead to near-potential maximum yields [7].

As described above, the LAI is a central indicator for growth management in tomato cultivation, and its proper control can guide plant development toward optimal conditions, thereby contributing to enhanced productivity. However, despite its importance, numerous technical challenges still exist in accurately and continuously monitoring the LAI in practical settings.

Various methods have been proposed for estimating the LAI. The most fundamental approach is destructive measurement, which involves harvesting leaves from a plant and measuring their area using a leaf area meter. Although this method provides high accuracy and is widely used in academic research, it is labor-intensive and time-consuming, making it unsuitable for routine or large-scale monitoring.

To address this issue, several non-destructive indirect estimation methods have been proposed. One simple approach estimates individual leaf area using models based on leaf length and width as primary variables [8,9,10,11]. Based on this estimation of individual leaf area, some studies have attempted to estimate the LAI [7,12]. Although intuitive and straightforward, this method requires manual measurement, which can be burdensome in practice.

Other low-effort techniques include optical techniques and image processing technologies. Optical methods exploit the fact that plants have distinct light absorption characteristics depending on the wavelength, allowing the estimation of leaf area. For instance, techniques based on the Lambert–Beer law have been evaluated as simple approaches for estimating the canopy-level LAI. Commercial LAI sensors (e.g., MIJ-15 LAI, EMJ, Fukuoka, Japan) that utilize the difference in light absorbance between photosynthetically active radiation (PAR) and near-infrared (NIR) wavelengths are available [13]. However, in crops such as tomatoes, which have a spatially uneven leaf distribution, light can be distributed non-uniformly within the canopy, making the measurement susceptible to errors depending on sensor placement.

To quantify such heterogeneous light environments within the canopy, bar-type line quantum sensors are commercially available and widely used in cultivation research. These sensors offer the advantage of integrating light measurements over a defined area; however, their high cost and need for regular maintenance pose economic and operational constraints for their continuous use in production environments.

In recent years, advancements in image processing and sensor technologies have garnered increasing attention as non-invasive measurement techniques. Several methods have been proposed for estimating the LAI in tomatoes using three-dimensional (3D) scanners to acquire point cloud data of the plant structure [14,15,16,17]. These approaches enable detailed structural analyses and high-precision measurements using three-dimensional data. However, in addition to computational demands, 3D measurement and analysis typically require the use of dedicated scanning devices, which increases the cost and complexity of deployment in greenhouse productions. Furthermore, since 3D reconstruction generally requires capturing images from multiple directions around the plant [14,15,16], such operations can become a major bottleneck in actual greenhouse environments, where space for human workers and equipment is often restricted. Therefore, while 3D imaging remains a powerful tool for plant structure analysis, several challenges must be addressed before it can be widely implemented in practical agricultural operations.

On the other hand, 2D image-based analysis can also serve as a practical alternative in applied settings. For example, an estimation method utilizing side-view images acquired by large-scale facility-based measurement robots has been reported [18]. In this method, the plant body is captured using a wide-angle camera fixed at a constant height. However, it is difficult to observe the vertical arrangement of leaves accurately, which limits their applicability to cultivation management practices that rely on the positional distribution of leaves. Furthermore, this method employs a semantic segmentation network that requires background removal training, resulting in high training and computational costs.

These observations highlight that while 3D imaging techniques offer highly accurate structural measurements of plant canopies, their practical implementation remains limited due to system complexity, high computational demands, and the need for multi-angle imaging, which is often constrained by space availability in greenhouse environments. A 2D imaging approach designed for production-scale settings has also been proposed; however, methods relying on fixed-height side-view cameras are unable to capture the full vertical profile of the plant, making them insufficient for analyzing canopy architecture in high-wire cultivation systems.

This study hypothesizes that a vertical scanning 2D imaging system, combined with a general-purpose segmentation model, can accurately estimate leaf area index (LAI) in high-wire tomato cultivation. To test this hypothesis, this study proposes a novel approach that combines a vertical scanning imaging system capable of capturing the entire plant profile with a simplified image analysis method.

The objective of this study is to develop and evaluate a novel 2D image-based pipeline that enables whole-plant LAI estimation. By scanning the entire plant from top to bottom using an RGB camera, the system captures a complete vertical profile of the plant. It is designed to function within the narrow space of cultivation pathways, making it suitable for use in real greenhouse environments. The captured images are then processed using the Segment Anything Model (SAM) for zero-shot leaf segmentation. This approach aims to overcome the spatial and technical limitations of prior methods, including the need for wide-angle multi-view imaging or fixed-position cameras. A key feature of the proposed method is its ability to capture the entire leaf structure of plants, which fixed-viewpoint methods cannot accomplish. This makes it distinctly different from existing observation techniques.

The LAI estimation was performed in two plots with different deleafing treatments. The estimated values were compared with destructively measured ground-truth data to evaluate the accuracy of the proposed method. In addition, a comparison was made with conventional LAI estimation techniques used by farm workers to assess the effectiveness and general applicability of the proposed method. Finally, the potential of this approach for future applications in crop growth diagnostic technologies is discussed.

2. Materials and Methods

2.1. Overview of the Study Workflow

The workflow of this study is illustrated in Figure 1. The entire process is divided into the following three stages:

First, the target tomato plants were prepared, and the scanning procedure was executed (Section 2.2).

Second, images were generated from the acquired scanning video, followed by image analysis using masking and segmentation techniques (Section 2.3).

Third, three separate simple linear regression models were constructed to estimate the LAI using different pixel-based variables—vegetation pixels (VPs), leaf pixels (LPs), and stem pixels (SPs)—as explanatory variables, respectively. These models are described and compared in Section 2.4, where the differences in their estimation accuracy are also evaluated.

For the accuracy evaluation, the destructively measured LAI was used as the ground-truth data to assess the prediction accuracy of the regression models. In addition, the accuracy of the image-based LAI estimates was compared with that of conventional visual and experience-based estimation methods performed by skilled workers to evaluate the practicality of the proposed approach.

The following sections provide detailed explanations of each step in the order of the workflow.

2.2. Target Plants and Imaging Method

2.2.1. Target Plants

This study focused on tomatoes (Solanum lycopersicum L.) cultivated using a high-wire training system. The target plants for imaging were healthy individuals grown in a high-eave greenhouse at the National Agriculture and Food Research Organization (NARO), Institute of Vegetable and Floriculture Science, Tsukuba, Ibaraki, Japan. The cultivation management followed standard high-wire cultivation practices. The planting density was set to 3.3 plants/m², and fertilization and irrigation were managed using an automatic watering system.

The cultivar used was ‘Momotaro York,’ a variety widely cultivated in Japan. Plants were trained on a single main stem, and all axillary buds were removed. The target plants for imaging were randomly selected from those grown in a greenhouse. The apical point of each plant was cut from the base of the stem to a height of 3000 mm. To achieve an LAI of approximately 6.0 m²/m², the position of the lowest leaf was adjusted on the basis of the expected number of attached leaves. In addition, to clearly observe the structural differences resulting from each deleafing condition, all lateral shoots, flowers, and fruits were removed, resulting in a plant structure composed only of the main stem and leaves.

2.2.2. Deleafing Treatments

To compare and examine the effects of different leaf area management strategies, two types of deleafing methods (hereinafter referred to as conditions) and three levels of deleafing for each (hereinafter referred to as levels) were defined. An overview of these conditions and their levels is shown in Figure 2.

The two deleafing conditions were as follows:

• The leaf height change condition (LH), in which deleafing was performed by altering the vertical position of leaves.
• The leaf density change condition (LD), in which deleafing was performed by reducing the overall leaf density uniformly across the plant.

Under the LH condition, deleafing was performed from the bottom leaves upward, considering leaf aging and reduced photosynthetic capacity. This method reflects common practice in actual cultivation and simulates deleafing operations aimed at maintaining an appropriate LAI throughout the growing period. In contrast, the LD condition involved selective thinning of leaves at uniform intervals throughout the entire plant. Leaves were removed systematically, starting from the top of the plant, based on their order. For example, in the LAI ≈ 4 treatment, every third leaf was removed while the other two were retained, following a consistent positional pattern. This ensured a uniform reduction in density while preserving the natural vertical structure of the canopy. This method was designed to accommodate variations in leaf density caused by cultivar or environmental differences, with the goal of developing a robust LAI estimation method applicable to tomato plants with diverse growth characteristics.

The following three deleafing levels were established under each condition. Under both conditions, the leaf attachment height before deleafing was adjusted within the range of 1200–3000 mm, based on grower experience, such that the initial LAI would be approximately 6.0 m²/m².

Under the LH condition:

• Removal of the bottom one-third of the leaves: LAI ≈ 4.0.
• Removal of the bottom two-thirds of leaves: LAI ≈ 2.0.

Under the LD condition:

• Removal of two out of every three leaves: LAI ≈ 4.0.
• Removal of two out of every three leaves: LAI ≈ 2.0.

Plants with no deleafing treatment were treated as the control (untreated group), with an LAI of approximately 6.0, and served as the baseline for comparison. Deleafing was performed on the same day as image acquisition, between 12:00 and 18:00.

Under each of the above deleafing conditions (2) and levels (3), four replicates were prepared, resulting in a total of 2 × 3 × 4 = 24 experimental plots. Each plot contained three tomato plants that were randomly selected from the cultivation area. This design ensured consistent evaluation across replicated units, allowing for statistical analysis.

2.2.3. Scanning Method

A scanning system developed in a previous study [19] was used for plant scanning. This system comprises a camera, light-shading panel, and light source. By adjusting the angle of the panel, the system directs light specifically to the target plant during scanning, creating a clear contrast with the background. This enables the stable acquisition of plant images with a darkened background through differential lighting.

In this study, the original system, which was initially designed to move horizontally and simultaneously capture multiple plants, was modified to perform the vertical scanning of individual plants. This adjustment enabled the full imaging of the plant body from the base to the apex. As shown in Figure 3a, the scan unit used by [19] was rotated 90°and mounted vertically (500 mm in height and 1000 mm in width). It was thereafter configured to move up and down using a portable electric winch (2955AT, Master Lock Company, Oak Creek, WI, USA) installed at the base of the system, with the movement guided via a pulley mechanism. The winch was operated at a no-load lifting speed of 1.95 m/min. The winch speed of 1.95 m/min was chosen based on preliminary testing and prior application of the system [19], where this speed provided a good balance between image sharpness and operational efficiency. At this speed, motion blur was minimized while allowing the entire plant to be scanned within a practical timeframe (approximately 55 s per plot). Each end of the scan unit was equipped with two flange rollers that ran along the guide pipes to ensure stable vertical motion. The guide pipes and flange rollers prevented lateral deviation of the scanning unit during vertical movement, thereby contributing to stable imaging. During initial testing, we observed slight irregularities in downward motion, likely caused by roller friction. To ensure smooth and consistent movement, scanning was performed only during the upward motion of the unit. After acquisition, the resulting images were visually inspected; although minor motion blur was observed in some frames, it was within an acceptable range and did not affect the accuracy of subsequent image analysis.

The maximum scanning height was set to 3500 mm, allowing coverage of the entire vertical leaf distribution of the tomato plant. As mentioned above, the system uses light contrast to prevent the illumination of background plants during imaging, resulting in clear and isolated images of the target plant (Figure 3b). This background exclusion was achieved mechanically through the system’s structural and optical design, as described in [19]. Specifically, a shading panel was positioned between the camera and surrounding plants, while directional lighting illuminated only the target plant from a fixed angle. The shading panel blocked stray light from reaching surrounding plants, and the controlled lighting minimized background visibility. As a result, only the target plant appeared bright in the image while surrounding plants remained unlit and visually suppressed without requiring post-processing.

The scanning system was mounted on rails placed over hot-water piping between the greenhouse rows, allowing easy movement between plots. Imaging was performed at night, after the lights were turned off, on 31 July 2019. Scanning was performed per plot to capture three tomato plants in a single image. The imaging process was based on prerecorded video, and recordings were made using a dedicated recorder (F44 Dual Audio Input Main Unit, Axis Communications AB, Lund, Sweden). In the scanning system, one modular camera (F1005-E, Axis Communications AB, Lund, Sweden) with a fixed 2.8 mm focal length and f/2.0 aperture was used. The camera had a horizontal angle of view of 113° and a vertical angle of 62°. Illumination was provided by four LED bar lights (LEDSC980-W, MISUMI Group Inc., Tokyo, Japan), with two lights placed on each side of the plant to ensure uniform brightness on the target plants. Videos were recorded at a resolution of 1920 × 1080 pixels and 30 frames per second (fps). No background screens were used during scanning, and the intensity of the lighting unit was maintained constant throughout the process.

2.3. Analysis of Scanning Images

2.3.1. Generation of Panorama Images

To visualize the entire plant body as a single image, panoramic images were generated from video data obtained using the scanning system. Images were generated following the approach described in a previous study [19]. The central pixel columns were continuously extracted from the video frames and arranged vertically to create a single elongated image representing the full profile of the plant. A custom automated script written in Python (version 3.11) was used for pixel extraction. Consequently, a continuous panoramic image of the entire length of the target plant was obtained.

2.3.2. Extraction of Green Vegetation Area via Masking

A masking process was applied to extract the plant area from the panoramic images. First, the acquired RGB images were converted to the HSV color space, and thresholding was applied to the three components: hue, saturation, and value. In OpenCV, the hue is defined in the range of 0–179. Based on this, the hue range corresponding to green was set to 35–85, which approximately corresponds to 70–170° in traditional hue angle terms. The thresholds for saturation and the value were set to 0–230 and 10–200, respectively. Pixels meeting these criteria were extracted as green vegetation areas (i.e., plant regions).

The threshold values were manually tuned with reference to the acquired images. However, because the scanning in this study was performed at night under controlled lighting conditions, the influence of ambient light was eliminated. Therefore, it was considered feasible to reliably extract plant regions using fixed threshold values. However, in other nighttime imaging scenarios where the lighting or camera configuration differs from that used in this study, or under conditions where leaf color varies significantly (e.g., due to cultivar differences or nutrient deficiencies), the same HSV thresholds may not yield reliable segmentation. In such cases, threshold values would need to be recalibrated to maintain accuracy.

The extracted regions were treated as binary masks. After removing the background, the number of pixels corresponding to the green portions of the plant, referred to as vegetation pixels (VPs), was calculated. The resulting VP values served as fundamental data for the subsequent leaf area estimation.

2.3.3. Separation of Leaf and Stem Areas via Segmentation

Following the masking process, which extracted the plant region from the image, a segmentation process based on deep learning (DL) was applied to separate the leaf and stem areas. The Segment Anything Model (SAM) was developed by Meta [20]. The SAM has been pre-trained on a large-scale segmentation dataset and can perform high-accuracy segmentation in a zero-shot manner, without requiring fine-tuning for specific domains. Recent studies have explored its application in agriculture, addressing limitations in crop disease and pest segmentation [21,22]. In poultry science, the SAM has demonstrated superior performance in chicken segmentation and tracking [23]. It has also been used to enhance the USDA’s Cropland Data Layer (CDL) by applying it to satellite imagery to improve crop-specific land cover mapping through better field boundary detection [24].

In this study, the SAM was implemented using a tool that allowed interaction via a graphical user interface. The model variant used was ViT-h, a large-scale Vision Transformer architecture. Region selection was performed using the prompt mode in which the main stem region was specified by the user, prompting the model to automatically generate the corresponding segmentation mask.

As a result of this process, the main stem area was clearly separated from the remainder of the plant, allowing high-accuracy identification of leaf and stem regions. From the generated segmentation image, the number of pixels corresponding to the leaf area (leaf pixels, LPs) and stem area (stem pixels, SPs) was calculated. These pixel counts were thereafter used as input variables for the subsequent LAI estimation.

Among them, LP (leaf pixels) corresponds directly to the projected area of leaves and is expected to serve as the most reliable predictor of LAI, which is the primary target variable in this study. Under conditions where plant canopies overlap, making it difficult to isolate individual plant areas, or where mechanical constraints prevent full imaging of the plant from the base to the apical point, accurate computation of LPs becomes challenging. In such cases, SPs may serve as a practical alternative, as stem regions are generally more consistently visible and less affected by occlusion or structural complexity. Although SPs do not directly represent leaf area, their use as a secondary explanatory variable may enhance the robustness of LAI estimation in environments where reliable leaf segmentation is difficult to achieve.

2.4. Estimation and Accuracy Evaluation of LAI

2.4.1. Acquisition of Ground-Truth LAI via Destructive Measurement

To evaluate the accuracy of the LAI estimates obtained through image analysis, ground-truth data (hereinafter referred to as LAI_gt) were acquired using destructive measurements. The target plants were cut at the base and separated into stem and leaf components, after which all leaves were removed for measurement.

Leaf area was measured using an automatic leaf area meter (LI-3100C, LI-COR, Lincoln, NE, USA). Destructive measurements were performed the following day and two days after the scanning procedure. The measured leaf area for each plot was recorded as the reference LAI value (LAI_gt) and used as a benchmark for constructing and evaluating the regression models described in this section.

2.4.2. Regression Analysis for LAI Estimation

A simple linear regression model was constructed to estimate the LAI using pixel-based information obtained from image analysis, with destructively measured LAI_gt as the dependent variable. The following three types of pixel information were used as explanatory variables:

• VPs (Vegetation Pixels): Number of pixels corresponding to the vegetation area extracted Via masking (Section 2.3.2)
• LPs (Leaf Pixels): Number of pixels corresponding to the leaf area (Section 2.3.3)
• SPs (Stem Pixels): Number of pixels corresponding to the stem area (Section 2.3.3).

Each of these features was used individually as an explanatory variable, and a separate simple linear regression analysis was performed for each. The prediction model is represented by the following linear equation:

$${\widehat{\mathit{LAI}}}_k=\mathrm{a}{·x}_k+\mathrm{b},$$

(1)

where ${\widehat{\mathit{LAI}}}_k$ and $x_k$ represent the estimated LAI and the number of pixels for either VPs, LPs, or SPs. The variable k denotes the identifier for the survey plot. The values $\mathrm{a}$ and $\mathrm{b}$ denote the slope and intercept of the regression line, respectively. These coefficients were estimated by substituting LAI_gt for $\widehat{\mathrm{LAI}}$ and using the least squares method. The goodness of fit of each model was evaluated using the coefficient of determination (R²).

2.4.3. Accuracy Evaluation of Leaf Area Estimation

Leave-one-out cross-validation was performed to evaluate the predictive performance of each constructed regression model, in which each sample was used once as the test data while the remaining samples were used for training. This ensured that model evaluation was conducted on data not seen during training, thereby providing an unbiased estimate of generalization performance. The mean absolute error (MAE) and mean absolute percentage error (MAPE) were used as evaluation metrics for the prediction accuracy. These metrics are defined as follows:

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{{\sum }_{k=1}^n\left|y_k-{\hat{y}}_k\right|}{n}},$$

(2)

$$\mathrm{M}\mathrm{A}\mathrm{P}\mathrm{E}={\displaystyle \frac{100}{n}}{\sum }_{k=1}^n\left|{\displaystyle \frac{y_k-{\hat{y}}_k}{y_k}}\right|,$$

(3)

where $y_k$ denotes the ground-truth LAI value (LAI_gt) obtained through destructive measurement, and ${\hat{y}}_k$ represents the LAI value estimated by the regression model using image-based variables (VPs, LPs, or SPs). These correspond to the “actual” and “predicted” values, respectively, in conventional statistical terminology used for MAE and MAPE. The variable $n$ refers to the number of plots per condition (12 in this study), and $k$ denotes the index of the plot.

To further assess the practical applicability of image-based LAI estimation, it was compared with conventional estimation methods used by farm workers. These worker-based estimates were referred to as LAI_worker, and the estimation method followed the approach described by [7]. Specifically, the worker visually measured the width and length of individual leaves, estimated the area of each leaf, and multiplied this value by the total number of leaves to calculate the overall LAI.

3. Results

3.1. Example Scanning Images

Figure 4 shows the representative scanning images under different deleafing conditions and levels. The top row corresponds to the leaf height change (LH) condition, whereas the bottom row corresponds to the leaf density change (LD) condition. From left to right, the columns represent the three deleafing levels: LAI ≈ 2, 4, and 6, respectively.

Under the LH condition, deleafing was performed from the lower leaves upward. As the LAI decreased, the distribution of attached leaves gradually shifted upward, which was visually confirmed. However, the remaining leaves appeared densely packed, indicating a uniformly compact canopy structure across the plant.

In contrast, the LD conditions employed a deleafing method that removed leaves at even intervals throughout the plant. Consequently, even at the same LAI values, noticeable gaps appeared in the middle section of the canopy, and leaf overlap (occlusion) was relatively reduced. This resulted in a clearer visibility of individual leaves, which is expected to contribute to the improved accuracy of leaf area estimation.

3.2. Example Results of Image Analysis

Figure 5 illustrates an example of the image analysis process applied to a scanned image, allowing visual confirmation of the changes at each processing step.

The first image (original image) was captured using the customized scanning method developed in this study. At the time of acquisition, the background had already darkened and was effectively removed owing to differences in lighting, resulting in a clear image of the target plant.

The second image (masked vegetation area) shows the result of thresholding in the HSV color space, used to extract the vegetation area. In most cases, non-plant elements, such as support clips or wires, were successfully removed, yielding an image in which the projected area of the plant was accurately quantified in terms of the pixel count.

The third and fourth images (segmented leaf and stem area) present the outcomes of semantic segmentation using the SAM. In these images, the main stem and leaf regions are clearly separated. Consequently, it is possible to individually extract LPs and SPs, which are subsequently used as features for LAI estimation.

3.3. Results of Regression Analysis for Estimating LAI from Pixel Counts

Figure 6 presents the results of the simple linear regression models constructed using pixel counts (VPs, LPs, and SPs) as explanatory variables and the measured LAI_gt as the dependent variable. Across all models, a significant linear relationship was observed between the pixel counts and LAI_gt, indicating that the pixel-based features extracted through image analysis can serve as effective indicators for quantitatively estimating the LAI of tomato plants.

In particular, the models using VPs and LPs exhibited high coefficients of determination (R² = 0.72–0.86), suggesting that both the vegetation area obtained through masking (VPs) and the more refined leaf-specific area obtained through segmentation (LPs) are useful for LAI estimation. However, when compared under the same deleafing conditions, the difference in R² values between the VP- and LP-based models was minimal, and no clear improvement in prediction accuracy due to leaf-specific segmentation was observed. This implies that although segmentation might enhance precision in extracting leaf areas, the vegetation area alone (VPs) already provides a sufficiently strong correlation with the LAI.

In contrast, the model based on SP count exhibited a negative correlation with LAI, but with lower R² values, indicating a weaker relationship than the other models.

When comparing the deleafing conditions, the LH condition yielded particularly high predictive performance, with R² values generally in the 0.80 range. In contrast, under the LD condition, R² values remained in the 0.70 range—approximately 0.1 lower than those under LH. Additionally, as shown in the figure, the LD condition exhibited a tendency for increased prediction error, particularly at moderate deleafing levels (LAI ≈ 4). On the other hand, under the LH condition, an increase in prediction error was observed at LAI ≈ 6.

3.4. Comparison of Estimation Accuracy Among Models

Table 1. Comparison of estimation errors (MAE and MAPE) for LAI under two deleafing conditions (LH and LD) using different explanatory variables: VPs, LPs, SPs, and worker-based visual estimation (LAI_worker).

Condition	X_var	Y_var	MAE ± S.D. (m2/m2)	MAPE ± S.D. (%)
LH	VP	LAIgt	0.59 ± 0.37	15.94 ± 9.76
	LP	LAIgt	0.60 ± 0.39	16.20 ± 9.80
	SP	LAIgt	1.03 ± 0.61	26.31 ± 14.82
	LAIworker	LAIgt	0.38 ± 0.24	10.76 ± 8.41
LD	VP	LAIgt	0.68 ± 0.36	22.59 ± 12.45
	LP	LAIgt	0.66 ± 0.37	22.01 ± 12.04
	SP	LAIgt	1.09 ± 0.85	43.09 ± 38.34
	LAIworker	LAIgt	0.49 ± 0.32	15.65 ± 8.49

summarizes the estimation accuracy of the LAI based on four types of explanatory variables—VPs, LPs, SPs, and LAI_worker—under two different deleafing conditions (LH and LD). Here, VPs, LPs, and SPs refer to the pixel counts derived from image analysis, whereas LAI_worker represents the values estimated by skilled workers using conventional manual methods. To evaluate the model accuracy, LOOCV was applied, and regression models were developed for each plot. The accuracy was assessed using the MAE and MAPE as evaluation metrics.

When comparing deleafing conditions, estimation errors were consistently lower under the LH condition across all explanatory variables, indicating higher prediction accuracy than under LD.

Focusing on the differences between the explanatory variables, the models based on VPs and LPs demonstrated similar levels of accuracy. Under the LH condition, the MAE for VPs was approximately 0.59 m² m⁻², and the MAPE was 15.94%. Under the LD condition, errors increased, with a VP MAE of 0.68 m² m⁻² and a MAPE of 22.59%. The values for the LPs were nearly equivalent to those for the VPs under both LH and LD conditions, with only slight differences.

In contrast, models using SP count exhibited the lowest accuracy among all variables. This was particularly evident under LD conditions, in which both the magnitude of the error and its variability (standard deviation) were significantly higher. These results suggest that LAI estimation based on stem area is more susceptible to variations caused by differences in plant structure owing to the deleafing treatment.

The conventional estimation method using workers (LAI_worker) exhibited the highest prediction accuracy under both conditions. For LH, the MAE was 0.38 ± 0.24 m² m⁻² and the MAPE was 10.76 ± 8.41%. For LD, the MAE was 0.49 ± 0.32 m² m⁻² and the MAPE was 15.65 ± 8.49%. Overall, LAI_worker proved to be the most reliable estimation method, although it required considerable time and manual effort, which remains a limitation in practical applications.

4. Discussion

4.1. Impact of Deleafing Conditions (Comparison Between LH and LD)

The results of this study revealed that across all explanatory variables (pixel counts), the LH condition (leaf height change) consistently yielded higher estimation accuracy for LAI compared with the LD condition (leaf density change). Notably, in the LH plots, the regression models using VPs and LPs—where leaves are the primary structural components—achieved high coefficients of determination (R² ≥ 0.85), along with the lowest MAE and MAPE values. The difference in performance between VPs and LPs was relatively small. This suggests that separating stems from vegetation pixels through additional segmentation may not be essential in all cases. Depending on the measurement conditions and required accuracy, the use of VPs alone may offer a sufficient balance between precision and processing efficiency. Although VPs and LPs showed comparable predictive performance overall, VPs may serve as a practical substitute for LPs in many applications due to their simplicity and lower computational cost. However, in cases where the proportion of leaves within the total plant coverage is relatively small, such as under low-LAI conditions or during early growth stages, using VPs alone may limit sensitivity in capturing variations in leaf area. In such scenarios, LP-based estimation can provide improved accuracy by focusing specifically on leaf structures and excluding stem or other non-leaf components.

Although SPs (stem pixels) do not directly represent leaf area and their performance was lower than that of VPs and LPs, it still showed a moderate correlation with LAI under the LH condition (R² = 0.57). This suggests that SPs may be considered as a supplementary indicator, particularly under measurement conditions where VPs or LPs cannot be reliably obtained due to occlusion, leaf overlap, or restricted imaging situations.

The difference in the estimation performance under the LH and LD conditions can be attributed to the effect of the deleafing method on the relationship between the projected leaf area and the actual leaf count. Under the LH condition, deleafing is performed from the lower part of the plant upward, resulting in a near-linear relationship between the number of leaves and the projected area captured in scanning images. This structural clarity facilitates the reliable estimation of leaf count and area based on pixel data, thereby contributing to higher model accuracy. However, in one of the LH-condition plants used in this study, the leaves were noticeably oriented toward the camera, leading to an overestimation of VPs and LPs relative to the actual LAI. This suggests that leaf orientation can have a non-negligible impact on estimation accuracy and should be considered when applying the model to diverse plant structures.

In contrast to LH, the LD condition involves uniform deleafing across the entire plant, indicating that a similar reduction in leaf count does not consistently translate into a proportional change in the projected area. Variability in leaf orientation and overlap (occlusion) causes inconsistencies in image representation. In some cases, physically removed leaves may still be present owing to occlusion, leading to ambiguous pixel-to-area correspondence and reduced estimation accuracy. Further analysis of the original images provided insights into the sources of increased prediction errors at LAI ≈ 4 (LD) and LAI ≈ 6 (LH). For LAI ≈ 4 under the LD condition, although leaf count was reduced due to deleafing, the removed leaves were originally in occluded positions, often obscured by overlapping foliage. As a result, their removal did not lead to a substantial change in visible pixel count, weakening the correlation with LAI and increasing estimation error. In contrast, under the LH condition with LAI ≈ 6, the overall number of leaves was inherently high, leading to greater variability in initial pixel detection across replicates. This variation likely contributed to increased prediction errors due to inconsistencies in segmentation accuracy under high-density conditions.

Moreover, structural features associated with LD conditions, such as uneven leaf density, are also influenced by management factors, including cultivar differences (e.g., internode length variability) and planting density. Therefore, the pixel-based estimation method employed in this study may have limited applicability under conditions in which the leaf distribution is highly non-uniform.

Therefore, to ensure the stable performance of LAI estimation based on pixel counts from scanning images, it is desirable that the vertical distribution of leaves be clearly defined and relatively uniform in density. Particularly in practical applications, it is essential to recognize that the estimation accuracy can be significantly affected by differences in the deleafing strategy and cultivation system.

4.2. Comparison with Conventional Estimation Methods

The comparative results confirmed that the conventional estimation method performed by skilled workers (LAI_worker) yielded the highest prediction accuracy. Both the MAE and MAPE values were superior to those of the image-based models. However, this method requires a sequence of manual operations in which the worker visually counts individual leaves, measures their length and width, and calculates the leaf area per leaf and total LAI using a predefined estimation formula. Particularly in large-scale greenhouses, where the number of evaluation samples increases, such labor-intensive tasks, while not directly contributing to revenue, can become a significant burden in daily operations. To address this, the image-based estimation method using scanned images demonstrated the potential to achieve a level of accuracy comparable to that of conventional methods while significantly reducing labor. Notably, under the LH condition, regression models using VPs or LPs achieved a practically acceptable level of precision, although they did not surpass LAI_worker in performance. Furthermore, the proposed image-based method significantly reduces labor by eliminating manual leaf handling and measurement. It also enables quick, non-destructive assessments with fixed scanning time, making it easier to plan and manage survey schedules—an important advantage in large-scale greenhouse operations.

A key factor contributing to the high accuracy of the conventional method is the explicit inclusion of leaf count as an explanatory variable. The results of destructive sampling in this study revealed a strong correlation between the number of leaves and LAI (R² = 0.87), indicating that leaf count could serve as a powerful predictor of the LAI. In contrast, image-based models did not directly include leaf count as a variable, which may have contributed to the gap in estimation accuracy.

Incorporating individual leaf segmentation techniques (e.g., [25]) into the image analysis pipeline could enable the estimation of leaf count from images and subsequently include it as an additional explanatory variable in LAI prediction models. Commercial applications capable of estimating individual leaf area may offer further detailed leaf-level information [26]. Alternatively, from a physiological modeling perspective, studies have shown that leaf number in tomato plants is strongly influenced by temperature [27]. For instance, the TOMGRO model predicts leaf appearance rate based on the cumulative thermal time derived under daily temperature conditions [2]. Leveraging such findings, it may be feasible to dynamically estimate leaf number using environmental data such as accumulated temperature.

By integrating such physiological models with image-based estimation methods such as scanning, further improvements in prediction accuracy may be realized. However, it is important to note that leaf count alone does not uniquely determine total leaf area, as individual leaf areas can vary depending on factors such as genotype and assimilate partitioning [28] or cultivar, environment, and nutrient status [29]. Therefore, relying solely on the leaf count may introduce errors.

In summary, the development of a robust LAI estimation method that balances both accuracy and operational efficiency will likely require a multivariate approach that includes leaf count and dynamic modeling of environmental influences.

4.3. Comparison with Existing Estimation Systems

Although relatively few studies have reported the estimation accuracy of the LAI in tomato cultivation, a comparison with several related studies highlights the effectiveness and unique features of the proposed method.

First, we compared our results with those of a previous study using two-dimensional (2D) image analysis [18]. In that study, tomatoes grown under high-wire cultivation in high-eave greenhouses were imaged using a rail-mounted wide-angle camera. The LAI was estimated via semantic segmentation based on a UNet architecture. They reported that the estimation error compared with the growers’ visual assessments was less than 10%. However, it is important to note that their reference LAI values were not destructively measured but were instead calculated based on the estimated leaf area and leaf count provided by growers, unlike in this study, which used ground-truth data from destructive sampling. In addition, in their system, the wide-angle camera was fixed at a constant height, making it difficult to capture the entire plant structure without distortion. In contrast, the scanning system used in the present study captured the full vertical profile of the plant from the base to the apex, enabling the visualization and quantification of changes in leaf position and height, which makes it more suitable for integration with cultivation management strategies.

Masuda [17] analyzed the tomato plants used in this study. A low-cost 3D scanner (Intel RealSense) was used to obtain the point cloud data, and the leaf area was estimated by separating the plant organs using a segmentation network. The authors reported a relative error of 21%. Compared with the MAPE in this study (approximately 16% under LH and 22% under LD), our 2D-based method showed comparable or slightly better accuracy. Because both studies used the same samples and base data, a direct comparison of accuracy was feasible. This suggests that accurate and practical LAI estimation can be achieved using simple 2D side-view images without complex 3D analysis.

Other studies utilizing 3D measurements include that conducted by [14], who employed a portable scanning LIDAR system to capture LAI in tomato canopies with LAI ≈ 3.5 m². They constructed detailed 3D point cloud models and achieved extremely high accuracy, reporting a MAPE of 4.6%, and analyzed complex plant structures such as leaf inclination angles. Ohashi et al. [15] used point cloud-derived surface models to estimate LAI in tomato plants with LAI ≤ 3. Their reported accuracy was R² = 0.600 and RMSE = 0.27, which is somewhat lower than the MAE ≈ 0.6 observed in this study. However, a direct comparison is complicated by differences in evaluation metrics. In cases where LAI exceeds 4, the results of this study indicate that a linear relationship between pixel count and LAI can be maintained up to approximately LAI ≈ 6, suggesting that the method retains a certain sensitivity even under relatively high foliage density. However, the applicability of the method beyond this threshold remains unverified. Further investigation is required to clarify the limitations of the proposed 2D imaging approach under extremely high-LAI conditions.

Taken together, these comparisons demonstrate that the LAI estimation method proposed in this study, which is based on side-view scanning images and simple image processing, offers a favorable balance among cost, labor efficiency, and accuracy. Future work should validate the proposed 2D method under a wider range of conditions, including different genotypes, environmental settings, and growth stages, to clarify its limitations and generalizability.

4.4. Limitations of This Study

While the results of this study demonstrate the potential of image-based methods for estimating LAI in high-wire tomato cultivation systems, a limitation should be acknowledged. To clearly observe the structural differences resulting from each deleafing condition, all flowers and fruits were removed prior to imaging. This preprocessing step, while beneficial for model accuracy, may limit the generalizability of the method to real-world greenhouse conditions where fruit presence is inevitable. Although the presence of a small number of fruits may have minimal impact under high-LAI conditions, their influence can become more significant when the LAI is low. Therefore, integrating an image analysis algorithm that detects and masks fruit areas will be an important extension for practical applications.

Although background conditions and environmental variation within the imaging facility were partially controlled and widely discussed in the previous study [19], this study was conducted under a specific set of imaging conditions. Any mechanical changes to the relative positioning of the imaging system and plant, or variations in lighting or background elements, may affect segmentation accuracy and should be considered when applying this system in different environments.

Before introducing the system into commercial greenhouses, its scalability and potential for cost reduction must be carefully evaluated. Although a detailed cost analysis has not yet been performed due to the prototype nature of the system, its application is expected to reduce labor costs associated with plant monitoring. Since the system requires the installation of sensors and imaging devices, it is envisioned for use in production facilities of a certain scale, where such technological investment is more viable.

This study was limited to tomato plants, and applicability to other crops remains to be explored. Potential targets include cucumbers, paprika, bell peppers, and eggplants, which are also important in greenhouse production. Some of these crops often have wide leaves and greater occlusion, which may affect measurement accuracy. Additionally, this study was conducted using a single tomato cultivar, which may limit the model’s generalizability to other genotypes. Differences in leaf morphology and canopy architecture among cultivars could affect the accuracy of segmentation and leaf area estimation. Further studies using multiple genotypes are necessary to evaluate the model’s robustness across genetic variation.

Another important limitation is the current lack of automation in the segmentation pipeline, which remains a significant technical barrier to real-time analysis using the system. To enhance the robustness and practical utility of the proposed method, future work should focus on expanding the system to accommodate different growth stages and validating its performance across diverse genotypes. Such developments will help ensure the stability and generalizability of the approach under a wider range of conditions.

5. Conclusions

This study aimed to establish a non-destructive and labor-efficient method for estimating the LAI in high-wire tomato cultivation by proposing a quantitative LAI estimation model based on side-view scanning images and image analysis. Using a custom-developed vertical scanning system, vertical plant profiles were captured and processed through HSV-based masking and DL-based semantic segmentation using the Segment Anything Model (SAM), enabling the extraction of pixel counts corresponding to leaves and stems.

The regression models constructed using the extracted pixel counts demonstrated strong correlations with destructively measured LAI values. In particular, models using VP and LP as explanatory variables achieved high accuracy, with coefficients of determination (R²) exceeding 0.85. Evaluation of prediction accuracy further showed that under the LH condition (deleafing based on leaf height), the models achieved low MAE and MAPE, second only to the conventional estimation method. This indicates that image-based estimation is effective under cultivation conditions where the leaf distribution corresponds linearly to the projected area in the images.

In contrast, under LD conditions (deleafing by altering leaf density), the estimation accuracy tended to decrease slightly owing to occlusion and leaf orientation effects. These results suggest that under conditions of uneven leaf density, the accuracy of simple linear regression models based on pixel counts may be limited.

A comparative analysis with previous studies confirmed the practicality of the proposed method. Although this approach is a preliminary proof of concept, the proposed approach achieved an estimation accuracy comparable to that of 3D-based methods, despite relying only on simple 2D scanning and basic image processing. However, it is essential to note that the system has not yet been tested under real-world production conditions, and its generalizability across different genotypes, environments, and growth stages remains to be verified.

Future studies should aim to validate the method under diverse conditions and further improve robustness by incorporating leaf count estimation, individual leaf area variability, and environmental data. Leaf count can be estimated by combining segmentation techniques or thermal time-based modeling. Integrating such information is expected to further enhance accuracy and support the practical deployment of LAI estimation systems in greenhouse crop management.