Non-Destructive Estimation of Fruit Weight of Strawberry Using Machine Learning Models

Abstract

Timely monitoring of fruit weight is a paramount concern for the improvement of productivity and quality in strawberry cultivation. Therefore, the present study was conducted to introduce a simple non-destructive technique with machine learning models in measuring fruit weight of strawberries. Nine hundred samples from three strawberry cultivars, i.e., Seolhyang, Maehyang, and Santa (300 samples in each cultivar), in six different ripening stages were randomly collected for determining length, diameter, and weight of each fruit. Pixel numbers of each captured fruit’s image were calculated using image processing techniques. A simple linear-based regression (LR) and a nonlinear regression, i.e., support vector regression (SVR) models were developed by using pixel numbers as input parameter in modeling fruit weight. Findings of the study showed that the LR model performed slightly better than the SVR model in estimating fruit weight. The LR model could explain the relationship between the pixel numbers and fruit weight with a maximum of 96.3% and 89.6% in the training and the testing stages, respectively. This new method is promising non-destructive, time-saving, and cost-effective for regularly monitoring fruit weight. Hereafter, more strawberry samples from various cultivars might need to be examined for the improvement of model performance in estimating fruit weight.

1. Introduction

Strawberry (Fragaria x ananassa) is one of the most high-value berry fruits across the world, having distinctive flavor, appearance, firmness, and chemical composition (anti-inflammatory properties, vitamin C, antioxidants, flavonoids, sugar, and organic acids) [1,2,3]. In recent times, its consumption has increased widely due to the high level of nutritional content and health benefit [4]. Being non-climateric, faster fruit-bearing, smaller plant size, shorter reproductive stage, and earlier maturation compared to other berries, it has been extensively cultivated all over the world [5]. Due to the value of these quality attributes of strawberries, the global production of it has doubled in the last 20 years to 8.8 million tons [4]. Hundreds of strawberry cultivars have been grown under numerous breeding projects worldwide for increasing the production level as well as for improving the fruit qualities, mainly color, weight, size, and internal compositions of fruits.

In strawberries, among the fruit quality parameters, weight is one of the most important factors in understanding water and nutrient use efficiency, biomass composition, fruit consistency, identification of cultivar, determining crop values, and fruit acceptance by consumers [6]. Moreover, fruit weight is highly sensitive to environmental conditions and management practices; therefore, most of the strawberry cultivars are grown in a greenhouse, where the microclimate and management processes can be precisely maintained [7,8]. In this circumstance, frequently monitoring fruit weight while considering environmental conditions and growth techniques is a paramount concern in strawberry cultivation [9]. Knowing fruit weight in advance helps growers to make more informed decisions regarding irrigation and nutrient management, harvest scheduling, and to optimize profit. Therefore, developing a constructive as well as comprehensible technique for measuring the weight of strawberries regularly using a machine vision technique is essential to improve the productivity as well as the quality of fruits.

The weight of strawberries is commonly estimated by instrumental assessment using a measuring scale. The main advantages of this technique are being user friendly and producing reliable results [10]. This technique can measure a large number of strawberries at the same time [10,11]. However, the traditional procedures are considered destructive, labor-demanding, and time-consuming, and they sometimes give low accuracy due to measurement errors [12,13]. Moreover, frequently monitoring strawberry weight in field conditions may be troublesome when using a measuring scale procedure [14]. Thus, the fruit weight prediction models based on machine learning algorithms and non-destructive measurements are useful. Machine learning algorithms develop models that are used to carry out a particular task based on a provided dataset [15]. In machine learning processes, data are fed to its algorithms, where they learn from the data to build models through a training process for making predictions and to optimize their operations to improve performance by developing intelligence over time [16]. Concerning processing ability to deal with complex data mapping relations, machine learning algorithms have been widely applied in agriculture fields over the past two decades, such as in plant breeding [15], in vitro culture [17], stress phenotyping [18], plant system biology [19], plant identification [20] and pathogen identification [21]. In particular, machine learning algorithms have undergone substantial development and have been applied along with non-destructive methods to estimate fruit weight. They have also received increasing attention as they outweigh the traditional destructive procedures with several advantages, including high throughput measurement, simultaneous multiple assessments and real-time decision making [22].

Recently, few papers have focused on machine learning techniques combined with digital image processing to estimate fruit weight [23,24,25]. Zhang et al. [26] estimated the weight and volume of apples using a 3D reconstruction technique which includes noncontact measuring methods involving computer vision, and the results showed that the least squares support vector machine model was found to have a higher correlation between the projected and actual weight of apples with a determination coefficient (R²) equal to 0.860. Teoh and Syaifudin [27] also described an image processing technique for predicting the weight of mango where the procedure can explain a maximum of 0.934 of the changes in measured and projected data in the training stage. In the case of strawberry, a number of studies have been conducted to develop models based on machine learning algorithms and non-destructive techniques for predicting volume, firmness, and internal quality parameters (total soluble solids content, acidity, sugar, and moisture content) [3,28,29] and the classification of ripeness level of strawberries, which is related to grading and sorting of fruits [30,31,32,33]. However, estimation of fruit weight of strawberries using non-destructive techniques is limited in the available literature. Hence, in this study, we have introduced a non-destructive, time-saving, and cost-effective method where the pixel numbers of strawberry images and machine learning algorithms have been used to develop models for measuring fruit weight.

Pixel numbers acquired by image processing techniques have been widely applied for the analysis of plants and animals research [34,35,36,37]. Images are first captured using a digital camera; subsequently, the image processing technique is applied for pixel numbers calculation. A number of research studies have been carried out to employ pixel number and linear and nonlinear-based machine learning models in measuring the volume of oranges [38], apples [9], tomatoes [39], and strawberries [40]. The effectiveness of these models in estimating volume of fruit is two-fold. Initially, the models can be employed to find the changes in volume in various ripening periods of fruit, which may give rise to enhanced knowledge. Subsequently, they are useful for monitoring fruit quality to explain the relations between the volume and pixel numbers. The main purpose of these models is to establish a relationship between the explanatory and response variables [41]. The most popular type of regression-based model is the simple linear form, which is widely applied in diverse research areas [41,42,43]. However, for nonlinear data, a linear-based regression model (LR) is difficult to perform well. Hence, machine learning-based nonlinear models such as support vector regression (SVR) are commonly used [44,45,46]. Therefore, in this study, we have developed both linear and nonlinear-based machine learning models to predict fruit weight and to compare their performance. The main objectives of this present study are three-fold: first, to determine the biometrical characteristics, i.e., length, diameter, and weight at the six ripening periods of three strawberry cultivars; second, to acquire pixel numbers of strawberry images using image processing techniques; and finally, to develop LR and SVR-based models using the pixel numbers for estimating the fruit weight of strawberries.

2. Materials and Methods

2.1. Sample Collection of Strawberries

Three strawberry cultivars, i.e., Seolhyang, Maehyang, and Santa, the most popular and commonly grown in South Korea, were selected for this experiment. Each variety of strawberry was separately grown in three greenhouses at Smart Farm Systems Laboratory of Gyeongsang National University during the period of September 2021 to January 2022. Five hundred strawberry plants of each cultivar were cultivated in individual greenhouses under the combination of Hoagland solution and bio plus compost soil (Figure 1). The bio plus compost soil includes zeolite (9.00%), peat moss (11.00%), perlite (11.00%) and cocopeat (68.86),which is a well-known medium for strawberry cultivation of different varieties inside greenhouses in Korea [46]. Daily application of irrigation water ranged from 20–30 mL for each plant during the early stages of strawberry growth to 30–50 mL for each plant for overall ripening stage [47]. During the whole growing stages of strawberry, the inside air temperature of all three greenhouses was maintained using Farm-link control system (Farmlink™ v 3.0, Jinju, Gyeongnam, Republic of Korea), while the illuminance, humidity and CO₂ were monitored through MCH 383SD (Lutron Co. Ltd., Taipei, Taiwan) sensor unit.

For collecting the fruit samples of each strawberry cultivar, harvesting was performed at six distinct ripening stages, i.e., Dark red, Bright red, Three-quarter red, Half red, Turning and Whiting. The ripening stages were identified and divided according to the development of color on the skin, which ranged from dark red (final stage) to white (initial stage) [6]. For the measurement of biometrical parameters, i.e., length, diameter, and fresh weight, 50 fruit samples were collected in each ripening stage of strawberry cultivars.

2.2. Measurement of Biometrical Characteristics

The fruit sample was cleaned with distilled water, drained and followed by cleaning with the use of paper towels to remove extra water from their surface before the measurement of length, diameter, and fresh weight. A total of 300 samples of each cultivar were used to determine their length, diameter and fresh weight. The fresh weight of each strawberry (in g) was measured using a digital balance (FX-300iWP, A&D Company Ltd., Tokyo, Japan), whereas the diameter and stem length (in mm) were measured by using a digital Vernier caliper (Model: E03-150 122–521, Datac Co., Ltd., Seoul, Korea).

2.3. Image Acquisition of Strawberry Fruits

Each strawberry sample’s image was acquired using a laboratory-based imaging system (Figure 2). A rectangular light chamber of dimension (80 × 80 × 80 cm) was used as studio to capture an image of each fruit sample using a camera setup (SONY DSC-RX100 vii, Sony, Seoul, Korea). The acquired image was in RGB format with resolution of 5472 × 3648 pixels. The light chamber consists of dual light-emitting diodes (LED) strips, where each strip of lamps had a 10 W total light emitting capacity [3]. In order to prevent shadows on the strawberry and to obtain a high-quality image, the interior of the light chamber (with the exception of the bottom surface) was made of an aluminum surface. To create a uniform background, the bottom surface of the light chamber was made of a black surface. The sample was placed at a fixed distance of 80 cm from the camera lens. Estimation of pixel numbers from spherical or quasi-spherical images is quite easy because they have a strong correlation with some dimensional parameters of the 2D projected region [48]; however, it is very hard to calculate for strawberries due to its natural irregularities in shape. In order to minimize this limitation, we captured images of each strawberry on both sides of the horizontal plane. The same procedure was followed during the image collection period for all 900 strawberry samples.

2.4. Pixel Numbers Calculation from Strawberry Image

Open-source Python libraries (Python 3.7.0) were used to process each image for calculating their number of pixels. Since our study focused only on the fruit area, both the opposite sides of each strawberry image were segmented from their background by employing remove.bg platform, and it was subsequently resized by 500 × 500 pixels. The background removal images were converted to Red, Green, Blue (RGB) and binary images consequently for estimating pixel numbers by using image processing algorithms. Figure 3 shows the typical strawberry sample with background removal (BR), RGB, and binary image in each cultivar.

2.5. Data Pre-Processing for Prediction Models Development

In this study, overall datasets, i.e., pixel numbers of each image and fruit weight, were obtained in all three cultivars of 900 strawberries (

Table 1. Cultivar-wise compositions datasets and number of data fruit weight of strawberries during the testing and training period.

Cultivar Name	Dataset	Number of Data
Maehyang	D1	300
Seolhyang	D2	300
Santa	D3	300
All three cultivars	D4	900

). Data preprocessing techniques commonly included four steps, i.e., missing data analysis, feature extraction, data normalization, and partitioning training and testing data. Since there were no missing values in the measured data; therefore, no method for imputing the missing data was taken into consideration in the current study. Throughout the model development stage, only one variable (pixel numbers) was considered as an input parameter; hence, feature extraction and data normalization methods were not applied during the data preprocessing time [44]. Additionally, the performance of machine learning models also depends on the size of data partition during the training and testing periods [3]. Moreover, sources of uncertainty in machine learning models arise when the testing and training data are mismatched due to the presence of noise in the dataset [49]. A number of studies used different portions of their data in training and testing stages such as 70% and 30% (training and testing), 80% and 20%, and 90% and 10% to reduce the uncertainty [50,51,52]. After evaluating the model’s performance with the three data subsets (70:30, 80:20, and 90:10), we used 80% of the collected data for training purpose and the remaining 20% of data for testing purposes during the model development.

2.6. Development of Linear Regression (LR) Model

Linear regression is the initial form of regression analysis where the relationship between the dependent and independent variables is considered in the linear approach. It is a modeling technique where the relationship is modeled using linear predictor functions [53]. The LR model has a wide range of applications in three major areas, i.e., for formulating numerical predictions, for examining the relationship between the variables and for time series modeling. It is commonly applied in various sectors such as crop yield prediction, weather forecasting, electricity demand estimation, and business projection, among others [43,54]. The LR model is represented by the following Equation (1).

$$Y_i={\beta }_o+{\beta }_1{X}_i+{\epsilon }_i$$

(1)

where the dependent variable $Y_i$ is represented of a linear function ${\beta }_o+{\beta }_1{X}_i$ of the explanatory variable $X_i$ together with an error term ${\epsilon }_i$. ${\beta }_o$ and ${\beta }_1$ are the intercept parameter (bias term) and slope parameter (coefficient), respectively. The basic structure of LR is presented in Figure 4a.

2.7. Development of Support Vector Regression (SVR)

Support vector machine is a powerful non-linear technique that makes a hyperplane or set of hyperplanes in a high-dimensional space for classification called as support vector classification (SVC), and the support vector regression (SVR) is developed on the same principle of SVC [55,56]. The performance of the SVR model is influenced by the basic kernel functions, since selection of the kernel functions is important for handling non-linear relationships more efficiently [57]. After testing on a number of SVR structures applying the three kernel functions, i.e., polynomial, sigmoid functions and radial basis function (RBF) with gamma (γ), penalty parameter of the error term (C) and epsilon (ε), in this study, as a kernel type, we decided to use a radial basis function (RBF) with γ = 0.5, C = 50 and ε = 0.1. The predicted fruit weight of strawberries was followed by the SVR operated Equation (2).

$${\widehat{\mathrm{Y}}}_0={\displaystyle \sum }_{i=1}^{n}K\left(X_i,X_0\right)\left({\alpha }_i-{\alpha }_i^{*}\right);K\left(X_i,X_0\right)=\frac{\exp {\left(-\left|X_i-X_j\right|\right)}^2}{Y}$$

(2)

In the equation, ${\alpha }_i$ and ${\alpha }_i^{*}$ represent support vectors; $n$ represents the numbers of datas; $K\left(X_i,X_0\right)$ denotes the radial basis function. The fundamental structure of the one-dimensional SVR is shown in Figure 4b.

2.8. Application Methodology and Model Performance Metrics

Python open source libraries (Python 3.7.0) have been employed to develop LR and SVR models in this study. Python, a high-level programming language, is widely applied in diverse research areas. In the Python environment, NumPy, Pandas, and Matplotlib were used for data management, processing, and presentation [58,59]. Two performance metrics, i.e., root mean square error (RMSE) (Equation (3)), and coefficient of determination (R²) (Equation (4)) have been used to evaluate the models’ performance. Statistical analysis was conducted using Statistical Package for the Social Sciences (IBM, SPSS Statistics 22.0.0.0, New York, NY, USA), and the data are represented as figures using Origin Pro 9.5.5. (OriginLab, Northampton, MA, USA). Figure 5 demonstrates the summative workflow of the current study methodology.

$$RMSE=\sqrt{\frac{1}{n}{{\displaystyle \sum }}_{t=1}^n {\left(y_{t actual-} y_{t predicted}\right)}^2}$$

(3)

$$R^2=1-\frac{{{\displaystyle \sum }}_{t=1}^n{(y_{t actual-} y_{t predicted})}^2}{{{\displaystyle \sum }}_{t=1}^n{(y_{t actual-} y_{t mean})}^2}$$

(4)

3. Results and Discussion

3.1. Changes in Biometrical Characteristics and Pixel Numbers

For analyzing biometrical characteristics (length, diameter, and weight) and pixel numbers of fruit images, a total number of 900 strawberries in three cultivars, i.e., Seolhyang, Maehyang, and Santa were collected from experimental greenhouses. The summary statistics of each measurement of length, diameter, weight, and pixel numbers of fruit images are shown in

Table 2. Biometrical characteristics and pixel numbers of images in three strawberry cultivars at ripening stages.

Cultivar Name	Length (mm)	Diameter (mm)	Weight (g)	Pixel Numbers
Maehyang	41.78 ± 3.28 a	31.93 ± 2.41 a	19.00 ± 3.61 a	51,460 ± 7425 a
Seolhyang	41.25 ± 3.69 a	32.51 ± 2.17 ab	20.41 ± 3.34 ab	57,579 ± 6776 ab
Santa	45.27 ± 2.95 b	34.45 ± 2.10 b	23.28 ± 3.65 b	59,966 ± 8502 b
All three cultivars	42.77 ± 3.77	33.97 ± 2.48	20.90 ± 3.96	56,335 ± 8396

Values represent the mean ± the standard deviation of 300 distinct strawberries measurement (N = 300) of each strawberry cultivar. Values mentioned by the same letter within the same column are not significantly different (p < 0.05).

. Length, diameter, weight, and pixel numbers of strawberry images changed significantly in Maehyang and Santa cultivars (p < 0.05); however, the variations were not statistically significant in Seolhyang with the other two cultivars (Maehyang and Santa) for diameter, weight and pixel numbers.

Biometrical data of strawberries showed that shape of Santa cultivar was larger compared to Seolhyang, and Maehyang. Length, diameter, and weight in tested Maehyang (Length = 35.36–50.37 mm; diameter = 26.10–38.66 mm; weight = 11.21–36.67 g), Seolhyang (Length = 29.42–51.42 mm; diameter = 25.47–40.34 mm; weight = 10.88–38.45 g) and Santa (Length = 34.51–53.58 mm; diameter = 27.52–42.72 mm; weight = 12.10–40.24 g) was into the range mentioned in literature for the three strawberry cultivars [60]. ‘Pajaro,’, ‘Chandler,’, and ‘Selva’ strawberry cultivars showed similar biometrical characteristics [61,62,63]. Similar to biometrical parameters, pixel numbers also obtained the highest for the Santa cultivar, followed by Seolhyang and Maehyang. This result indicated that the pixel numbers of strawberry images increased with the increase in the shape and weight of fruit. Similar findings were also reported for ‘Oso Grande’ and ‘Sweet Charly’ strawberries [64]. In this study, fruit length and diameter of each strawberry are scarcely described. Moreover, in general, fruits collected in three strawberry cultivars showed good biometrical characteristics.

3.2. LR Model Performance for Fruit Weight Prediction

The performance of the linear regression model mainly depends on the existence of a linear relationship between the explanatory and response variables [42]. A simple linear model is widely applied in diverse research fields due to its simplicity of nature along with an easy interpretation of its outcome [41,46,65]. In the current study, we applied a linear regression model, i.e., simple linear regression (LR) in predicting the fruit weight of strawberries. The pixel numbers of images in three strawberry cultivars have been used as an input parameter for developing LR model. The performances of LR in predicting fruit weight are shown in

Table 3. Performance metrics (RMSE and R²) of both models for predicting the fruit weight of strawberries during the training and testing period. Italicized values indicate the best performance result for each model.

Model Name	Dataset	Training		Testing
		R2	RMSE	R2	RMSE
LR	D1	0.963	0.712	0.896	1.054
	D2	0.950	0.785	0.871	1.136
	D3	0.947	0.856	0.860	1.207
	D4	0.954	0.758	0.880	1.101
SVR	D1	0.942	0.891	0.856	1.239
	D2	0.934	0.953	0.838	1.362
	D3	0.930	0.984	0.830	1.402
	D4	0.936	0.946	0.840	1.280

.

The study results showed that the highest R² (0.963 and 0.896) and lowest RMSE (0.712 and 1.054) were observed for the D1 dataset, indicating that the LR model can explain a maximum of 96.3% and 89.6% of the variations of actual and predicted data during the training and testing stages, respectively. Contrary to that, the lowest performance was observed for the D3 dataset, where the lowest R² (0.947 and 0.860) and the maximum RMSE (0.856 and 1.207) were found in the training and testing stages, respectively. The LR model developed by the D1 dataset could predict fruit weight with a 1.70% and 4.19% increase in R² and a reduction of 16.82% and 12.68% in RMSE in the training and testing stages, respectively, compared to the LR model developed by the D3 dataset. Apart from the two datasets (D1 and D3), the LR model also performed better in predicting fruit weight using the D2 and D4 datasets. The measured and predicted values of fruit weight in the D1, D2, D3 and D4 datasets for the LR model in the testing stage are illustrated in scatter plots and time series graph (Figure 6). According to Figure 6, it is shown that the measured data fit well with the predicted data for the D1 dataset and the values were very close to the line 1:1 compared to the D2, D3, and D4 datasets. Lee et al. [40] established a linear regression model to predict strawberry volume with coefficients of determination of 0.866 and 0.603 in the training and testing stages, respectively. Omid et al. [34] estimated weight by measured volume using an image processing technique and found the coefficient of determination for lemon, lime, orange, and tangerine, which were 0.962, 0.970, 0.985, and 0.959, respectively, which were almost similar results compared to this study. The mass model of pomegranate developed by volume using a water displacement method was explained as a linear form of mass (M) = 0.96 volume (V) + 4.25 [66]. Similar experiments were also conducted by employing digital image processing techniques to predict the volume and mass of oranges [38], lettuce [35] and tomatoes [39]. These studies established a linear relationship between the volume and weight of fruit; however, in this study, we measured the strawberry fruit weight using a non-destructive method with high accuracy.

3.3. SVR Model Performance for Fruit Weight Prediction

The support vector regression (SVR) can process the non-linear data effectively and fit very well; as a result, the performance is comparatively better than linear models [30]. Since LR is a linear-based regression model, a nonlinear-based regression model, i.e., SVR has been performed in predicting fruit weight and evaluating the performance of the two models. For the non-linear approach, the SVR model is widely used in a number of studies to measure the fruit quality parameters [3,56,67]. Like the LR model, the four datasets, i.e., D1, D2, D3, and D4, were also used to develop the SVR model and to compare the performance among the datasets for predicting fruit weight.

The performance of the SVR model for each dataset in predicting fruit weight is shown in Table 3. Based on the R² and RMSE values, the findings of this study showed that the SVR with D1 dataset provided better performance compared to D2, D3, and D4. According to the outcomes of SVR, the highest R² (training = 0.942 and testing = 0.856) and the lowest RMSE (training = 0.891 and testing = 1.239) values were obtained for D1, which indicates that SVR could explain a maximum of 94.2% and 85.6% in the training and testing period, respectively. However, the worst performance was obtained for the D3-based SVR model (Table 3). Moreover, the SVR model developed on the D1 dataset could predict fruit weight with a 1.30% and 3.13% increase in R² and 9.45% and 11.63% reduction in RMSE in training and testing, respectively, compared to the D3 dataset. Nyalala et al. [68] estimated the tomato weight using the computer vision and radial basis function–SVR (RBF-SVR) and Bayesian artificial neural network (Bayesian-ANN), and the proposed technique obtained an average accuracy of 95% and 96% using RBF-SVR and Bayesian-ANN models, respectively, for weight prediction, which were almost similar results compared to the present study. Moreover, Khojastehnazhand et al. [69], El Hariri et al. [70], and Fellegari and Navid [71] determined the volume and weight of tangerine, tomato, and orange, respectively, using image processing with machine learning algorithms (SVR). According to the findings of these three studies, it can be concluded that the performances of the SVR model developed in our study using D1, D2, D3, and D4 in estimating strawberry fruit weight were within the acceptable levels. In addition, the actual and predicted data on fruit weight obtained from SVR with D1, D2, D3, and D4 were shown in the time series graph and scatter plots (Figure 7). As shown in Figure 7, the predicted data and the observed data fit well; similarly, the values were near the 1:1 line, suggesting that SVR had a high accuracy level in estimating the fruit weight of strawberry.

3.4. Comparison LR and SVR Model’s Performance and Proposed Model

Regarding the results of comparison between the LR and SVR models with four datasets D1, D2, D3, and D4, it was observed that the performances of both models were almost similar for all four datasets. However, among the four datasets, the D1 (Maehyang) performed best for both LR and SVR models in predicting fruit weight. This finding indicated that the shape of the strawberry cultivars may affect the pixel numbers of captured images, which are directly related to fruit weight. Uniform shape of any object gives a good quality image for acquiring precise information [72,73,74]. The Maehyang and Seolhyang strawberry cultivars have a conical and almost uniform shape, whereas Santa is a large size strawberry cultivar and sometimes follows an irregular shape [60]. This may be one of the main reasons for lower performance of the D3 (Santa) dataset compared to D1 (Maehyang), D2 (Seolhyang), and D4 (combined three datasets) in the current study. Similar results were also reported by Agulheiro-Santos et al. [75] and Guo et al. [76].

Moreover, the LR model had a slightly better performance compared to the SVR model in predicting fruit weight of strawberries (Figure 8). Based on the performance metrics (R² and RMSE), the results of the current study showed that the selected LR model based on D1 could predict fruit weight with a 2.23% and 4.67% higher R² and with 20.10% and 14.93% lower RMSE in training and testing periods, respectively, compared to the SVR model with the D1 dataset. One of the main reasons might be a certain portion of the relationship between the explanatory (pixel numbers) and response (fruit weight) variables exiting a linear relation, and the linear data could play a significant role in the better performance of the LR model. Some studies also noted that SVR models could not increase the prediction accuracy compared to simple linear models due to the linear nature of the variables [45,77,78]. In addition, the cumulative distribution function that resulted from the LR model achieved a residual value of 78.33% of the observed and predicted data of fruit weight between the ranges from −1.0 to 1.0, whereas it was 63.67% for the SVR model in the same boundary limit (Figure 9). As shown in Figure 9, there was a linear relationship between the actual and predicted data of fruit weight. Moreover, models developed in the current study using machine learning algorithms might be subjected to noise and model inference errors. It is thus highly desirable to represent uncertainty in a trustworthy manner in any machine learning-based process [79,80]. Therefore, it is vital to evaluate the reliability and efficacy of machine learning models by conducting more experiments with a large number of samples; thereafter, they could be applied in real fields.

4. Conclusions

The experiment was conducted to assess the performance of machine learning models, i.e., LR and SVR models for the fruit weight prediction of three strawberry cultivars using image processing techniques. The finding of this study showed that there was a significant variation in length, diameter, and fruit weight between the Maehyang and Santa cultivars. The pixel numbers of fruit images acquired from the image processing techniques were used as input variables for developing fruit weight prediction models, and the performance of the models was measured using R² and RMSE. The results of the study noted that the prediction accuracy of the LR model was slightly better than SVR model in estimating fruit weight, indicating a linear relationship between the pixel numbers and fruit weight. Among the four datasets (D1, D2, D3, and D4), LR and SVR models achieved the best performance for the D1 (Maehyang) dataset, which may be due to its conical and almost uniform shape. The LR and SVR models with the D1 dataset could explain a maximum of 96.3% and 94.2% in the training stage and 89.6% and 85.6% in the testing stage, respectively, of the variations of the pixel numbers and fruit weight data. In conclusion, strawberry fruit weight was satisfactorily predicted on the basis of the pixel numbers. Hereafter, more samples from various strawberry cultivars might need to be examined to determine whether the LR and SVR models performed better than the other regression-based models, such as Elastic Net, Decision Tree, Random Forest, in predicting fruit weight. This new method is promising non-destructive, time-saving and cost-effective for timely monitoring of fruit weight; hence, the next experiment will give attention to its application in the real agriculture field.