Combining Machine Learning with Images from RGB and Multispectral Cameras to Estimate Fresh Weight and Height of Spinach for Processing Use

Abstract

Spinach for processing use should be systematically harvested according to the fresh weight and height standards required by the processing plant; thus, monitoring spinach growth in fields is crucial. However, ground measurements require substantial labor. In recent years, unmanned aerial vehicles (UAVs) have been established as effective labor-saving tools for monitoring crop growth. This study aimed to clarify: 1) which model and image type are appropriate for mono-regression in the estimation of fresh weight and height of processing spinach, 2) whether a combination of images from RGB and multispectral cameras and the application of machine learning can improve the estimation accuracy in multivariate regression, and 3) which model shows the highest estimation accuracy over extrapolation data. In the mono-regression, the highest fresh weight estimation accuracy was achieved with the soil adjusted vegetation index (SAVI) using a multispectral camera at 50 m altitude (test root mean squared error (RMSE) = 0.795). The highest height estimation accuracy was achieved with the modified green red vegetation index (MGRVI) using an RGB camera at 50 m altitude (test RMSE = 5.01). Estimation accuracy was highly dependent on image type, suggesting that the effectiveness of appropriate selection for accurate estimation. In multivariate regression, machine learning models showed higher accuracy for fresh weight estimation compared to multiple linear regression. Although most support vector regression models exhibited higher accuracy than that of the multiple linear regression model, random forest regression models exhibited lower accuracy for height estimation. These findings suggest that the selection of an appropriate model type is critical. Although combining images from RGB and multispectral cameras improved accuracy in machine learning models, the effects were inconsistent because of other factors including target traits, model types, and UAV altitude. After comparing all models with the highest accuracy in each category, the fresh weight estimation model developed using support vector regression with images from an RGB camera at 30 m altitude, and the height estimation model based on support vector regression with images from both cameras at 30 m altitude showed the highest accuracies (test RMSE = 0.720 and 4.19, respectively). These machine learning-based models can thus facilitate precise monitoring of the fresh weight and height of spinach for processing use.

Introduction

Recently, in Japan, the demand for domestic vegetables has decreased, whereas demand for processing and commercial use has increased (MAFF, 2023). Among vegetables, spinach has a high demand for processing use (Kobayashi, 2006). Cultivation of spinach for processing use differs from normal spinach cultivation. The major difference is in the use of harvesting standards, as fresh weight and height standards are required by processing plants. Processing plants require the arrival of predetermined quantities of spinach for smooth operation; therefore, spinach should be harvested systematically according to the fresh weight in the field. According to height standards, spinach for household consumption should be approximately 25 cm high, whereas spinach for processing use should be at least 40 cm high (Kobayashi, 2006). However, planned yields and height standards for processing spinach are sometimes not achieved because the fields and harvest times are decided by the farmer.

To harvest spinach systematically according to processing standards, a detailed understanding of spinach growth is required. Some studies have used growth simulation models to estimate spinach growth (Kamada et al., 2021; Koyano and Kamada, 2022). These studies used meteorological information on the daily mean temperature and daily global solar radiation to estimate spinach growth and showed that fresh weight could be predicted during the harvest season. However, Kamada et al. (2021) suggested that the prediction model requires calibration by field measurements as it exhibits inaccuracy depending on weather conditions, such as heavy precipitation. Although simulation models are useful tools for roughly predicting spinach growth, field measurements are necessary to address unforeseen weather situations.

To measure spinach growth in the field, non-destructive methods using handheld cameras or sensors have been applied instead of the classic destructive methods used for other crops as ground measurements require substantial labor (Jones et al., 2007; Liu et al., 2006). To reduce labor costs, unmanned aerial vehicles (UAVs) have been established as effective tools for monitoring crop growth. The advantages of using UAVs include the efficiency of capturing several hectares simultaneously, the convenience of capturing images at any time, and the accuracy of captured images with a high resolution of several millimeters. In addition, UAVs have the ability to evaluate the physical properties of the crop canopy, such as canopy height, using the structure obtained from the motion and multi-view stereo (SfM-MVS) process. However, few studies have monitored the growth of processing spinach using UAVs, which is vital for determining suitable harvesting periods.

To develop growth monitoring using UAV, image capture and processing methods need to be carefully examined. Many studies have examined the appropriate image types and regression algorithms according to the plant type and purpose (Jin et al., 2017; Moeckel et al., 2018; Tsukaguchi et al., 2022). Furthermore, in recent years, the UAV-monitoring accuracy of crop yields and traits has been improved by several new approaches, including the combination of different image types and application of machine learning algorithms (Astor et al., 2020; Cheng et al., 2022; Yang et al., 2021). However, machine learning can cause difficulty in terms of over extrapolation in some cases (Naghdyzadegan Jahromi et al., 2023). To avoid such situations and select a model with improved applicability, model evaluation of data over different growth periods is needed.

Therefore, this study aimed to clarify: 1) which model and image type are appropriate for mono-regression in the estimation of fresh weight and height of processing spinach, 2) whether a combination of images from RGB and multispectral cameras and the application of machine learning regression can improve the estimation accuracy in multivariate regression models, and 3) which model shows the highest estimation accuracy over the extrapolation data.

Materials and Methods

Spinach cultivation and ground measurement

The spinach cultivar ‘Kuronosu’ was grown at the Kyushu Okinawa Agricultural Research Center, NARO in Miyakonojo, Miyazaki Prefecture, Japan (31°45′N, 131°00′ E) in 2020 to 2021. Field comprised black soil, with a pH of 6.2, NO₃-N of 35.9 mg·kg⁻¹, NH₄-N of 6.0 mg·kg⁻¹, available-P of 184 mg·kg⁻¹, and available-K of 266 mg·kg⁻¹ before cultivation. Spinach was seeded in three different seasons (October 2, 2020, November 9, 2020, and February 3, 2021), with an inter row space of 30 cm and an intra row space of 7–8 cm in four rows. Four different nitrogen conditions were applied with three replicates each: conventional fertilization as a control, additional fertilization, split fertilization, and basal fertilization, as summarized in Table 1. Additional fertilizers were applied approximately one month after seeding. Phosphorus and potassium were applied in conventional amounts of 16 and 8 kg·10a⁻¹, respectively. The daily mean temperature and precipitation during the cropping season are summarized in Figure 1. We used agrometeorological grid square data from NARO (Ohno et al., 2016) (https://amu.rd.naro.go.jp/).

Table 1

Fertilizer application rate in each treatment.

Fig. 1

Daily mean temperature and precipitation in growing seasons.

The fresh weight (kg) and height (cm) of the spinach populations per area were measured every two weeks by destructive sampling. Each measured population consisted of two internal rows in four rows, each 0.5 m long, with an area of 0.3 m⁻². Twelve populations (four treatments with three replicates each) were measured simultaneously. Four measurements were taken for the October and February seedings, and eight for the November seedings.

Image acquisition with UAV

A UAV (Phantom 4 Pro; DJI, Hong Kong, China) with an onboard RGB camera and a mounted multispectral (MS) camera (Sequoia, Parrot) was prepared. An onboard RGB camera was used to capture RGB images with a high resolution of 20 MP, and an MS camera was used to capture MS images with monocameras of green 530–570 nm, red 640–680 nm, red edge 730–740 nm, and NIR 770–810 nm of 1.2 MP (a Sequoia RGB camera was not used in the current study). The flight route was set by autonomous flight application (DJI GS PRO; DJI), and flight altitudes were set at 30 m and 50 m. Front lap and side lap were set as 80% and 70%, respectively, based on the onboard RGB camera. The images were captured 1–3 d before the ground measurement day from 10:00 to 15:00. All flights were conducted automatically, and images were captured by time lapse setting every two seconds.

Image processing

Captured images were orthomosaiced by the SfM-MVS process using Pix4D Mapper software (Pix4D S.A.; Lausanne, Switzerland). The RGB camera obtained an orthomosaic RGB image and a digital surface model (DSM), and the MS camera obtained orthomosaic reflectance images of green, red, red edge, and NIR by autonomous calibration with an external light sensor without a calibration panel. In the SfM-MVS process, four ground control points (GCPs) in the field corner were used to calibrate the geoinformation. For image analysis, the normalized orthomosaic RGB images were generated by dividing each digital number (DN) of R, G, and B by the sum of the DNs from the RGB image, as shown in Equations (1–3): DSM images were used for height image (deemed as H) generation to obtain the canopy height above the ground surface as follows: the DSM image of each day minus the DSM image after seeding. Images of vegetation indices were generated from calculations of normalized RGB images and MS reflectance images. Thirteen images of vegetation indices that were used in previous studies for plant growth monitoring were obtained (Table 2). WDRVI has a parameter of a (ranging from 0.1 to 0.2), and a smaller a value can mitigate the saturation problem of the NIR band for high biomass conditions. We set the a to 0.1, expecting to deal with it if a saturation problem was encountered.

Table 2

Summary of vegetation indices and formulae used in this study (r, g, b: the digital numbers (DNs) of normalized RGB images) (Green, Red, Red Edge, NIR: reflectances of MS images).

(Equation 1) r = R / (R + G + B)

(Equation 2) g = G / (R + G + B)

(Equation 3) b = B / (R + G + B)

Model development

As explanatory variables of models, the mean values for each ground-measured location were calculated across all images using 21 preprocessed images, which included eight original images (r, g, b, and H obtained from the RGB camera and green, red, red edge, and NIR obtained from the MS camera) and 13 images of vegetation indices (5 obtained from the RGB camera and eight obtained from the MS camera). The objective variables were ground-measured fresh weight and height. Two types of regression models were used: mono-regression and multivariate regression. In the mono-regression model creation, linear and exponential regressions were compared and the model with the lowest root mean squared error (RMSE) was selected. We compared 21 models: nine models obtained from the RGB camera and 12 models obtained from the MS camera. In the multivariate regression model creation, three models were used: multiple linear regression (MLR) and two machine learning regression models, random forest regression (RF) and support vector regression (SVR). We compared three patterns of explanatory variables: 1) four images of r, g, b, and H from the RGB camera; 2) four images of Green, Red, Red edge, and NIR from the MS camera; and 3) the above eight images from both cameras. For standardization, the training data were standardized and their mean and standard deviation were used to standardize the test data. To avoid multicollinearity problems, the feature selection method of recursive feature elimination (RFE) was applied before estimating the MLR parameters. The hyperparameters of the machine learning regression models were determined by a grid search. During training, mean squared error (MSE) was used as the accuracy index.

To determine the parameters of all models including mono-regression and multivariate regression models, 144 October and November seeding data points were used for training, and a five-part cross-validation was performed. To validate the model accuracy, the R-squared (R²) values of the training data were compared. To validate the model accuracy and applicability to extrapolation data, 48 data points from the February seeding were used as the test data, and the RMSEs of the test data were compared.

Results

Fresh weight and height of ground spinach

The changes in fresh weight and height of the ground spinach in each treatment and cropping season are shown in Figure 2. The fresh weight ranged from 0.502 to 6.04 kg·m⁻² in the October seeding, 0.116 to 8.47 kg·m⁻² in the November seeding, and 0.0334 to 7.43 kg·m⁻² in the February seeding. The height ranged from 15.2 to 50.1 cm in the October seeding, 9.72 to 56.8 cm in the November seeding, and 5.44 to 78.9 cm in the February seeding. The height increased rapidly in the October seeding and reached 40 cm at approximately 50 days, while approximately 100 and 60 days were required to reach 40 cm in November and February, respectively. The fresh weight and height gradually increased during each cropping season, but the rate of increase differed. In the October seeding, spinach grew rapidly in the earlier growing season, and grew slowly in the later growing season because of low temperatures. In the November seeding, spinach grew slowly because of low temperatures for most of the growing period. In the February seeding, spinach grew rapidly in the later growing season owing to high temperatures. The fresh weight and height varied slightly depending on the nitrogen conditions.

Fig. 2

Changes in fresh weight and height for each treatment in each cropping season (Error bars represent standard error, n = 3).

Fresh weight estimation

The accuracy of each mono-regression model for fresh weight estimation is presented in Table 3. The estimation accuracies differed according to the image type. With images from the RGB camera, R² and RMSE ranged from 0.312 to 0.844 and 0.925 to 2.75, respectively, at 30 m altitude, and from 0.418 to 0.688 and 1.00 to 3.02, respectively, at 50 m altitude. With images from the MS camera, R² and RMSE ranged from 0.181 to 0.880 and 1.07 to 4.25, respectively, at 30 m altitude, and from 0.278 to 0.875 and 0.795 to 5.04, respectively, at 50 m altitude. The lowest RMSE of 0.795 was achieved using the SAVI with an MS camera at 50 m UAV flight altitude, following those of OSAVI (RMSE = 0.815) with an MS camera at 50 m, and EVI2 (RMSE = 0.825) with an MS camera at 50 m.

Table 3

Estimation accuracy of fresh weight and height of spinach for processing using mono-regression for each camera type, UAV flight altitude, and vegetation index type.

The accuracy of each multivariate regression model for fresh weight estimation is presented in Table 4. The estimation accuracies differed with the model and image type. Comparing RMSE values from the same camera type and UAV altitude conditions, the RMSE values of the RF and SVR models were lower than those of the MLR models. When using images from the RGB camera, the SVR showed the lowest RMSE of 0.720 (R² = 0.781) at 30 m altitude, and the lowest RMSE of 1.41 (R² = 0.410) at 50 m altitude. With images from the MS camera, the SVR showed the lowest RMSE of 1.24 (R² = 0.925) at 30 m altitude, and the lowest RMSE of 1.44 (R² = 0.834) at 50 m altitude. For images from both cameras, SVR showed the lowest RMSE of 0.778 (R² = 0.877) at 30 m altitude, and the lowest RMSE of 0.752 (R² = 0.829) at 50 m altitude.

Table 4

Fresh weight estimation accuracy using multivariate-regression models for each image type, UAV flight altitude, and regression algorithms (RGB: four images of R, G, B, and H from the RGB camera; Multipspectral: four images of Green, Red, Red edge, and NIR from the MS camera; Both: the above eight images from both cameras).

Figure 3 summarizes the fresh weight estimation models with the lowest RMSE for each UAV flight altitude, regression algorithm, and image type, to compare the accuracy of the best models in all categories. The fresh weight estimation model with the lowest RMSE was the SVR model with an RGB camera at 30 m altitude (R² = 0.781, RMSE = 0.720) (Table 4; Figs. 3 and 4). Comparing the estimation accuracies of the mono- and multivariate regression models when using images from the RGB camera at 30 m altitude, the multivariate regression of SVR had a lower RMSE (0.720) than that of the mono-regression of H (RMSE = 0.925), which showed the lowest RMSE in mono-regression (Fig. 3; Tables 3 and 4). With images from the RGB camera at 50 m altitude, the multivariate regression of SVR had a higher RMSE (1.41) than that of the mono-regression of NGRDI (RMSE = 1.00) (Fig. 3; Tables 3 and 4). With the images from the MS camera, the multivariate regression of SVR had a higher RMSE (1.24) than that of the mono-regression of NDVI (1.07) at 30 m altitude, and the multivariate regression of SVR had a higher RMSE (1.44) than that of the mono-regression of SAVI (RMSE = 0.795) at 50 m altitude (Fig. 3; Tables 3 and 4).

Fig. 3

Summary of fresh weight estimation accuracy of the best models for each UAV flight altitude, regression algorithm, and image type.

Fig. 4

Comparison of ground measured and estimated fresh weight of test data using the SVR model with images from an RGB camera at 30 m flight altitude (Figure represents the seeded month (February) and days after seeding of test data).

Height estimation

The accuracy of each mono-regression model for height estimation is presented in Table 3. The estimation accuracies differed according to the image type. With the images from the RGB camera, R² and RMSE values ranged from 0.417 to 0.942 and 5.26 to 15.2, respectively, at 30 m altitude, and from 0.448 to 0.810 and 5.01 to 17.6, respectively, at 50 m altitude (Table 3). With images from the MS camera, R² and RMSE ranged from 0.261 to 0.823 and 8.96 to 28.0, respectively, at 30 m altitude, and from 0.328 to 0.832 and 6.86 to 32.4 respectively at 50 m altitude. The lowest RMSE of 5.01 was achieved by MGRVI with the RGB camera at a UAV flight altitude of 50 m, followed by that in H (RMSE = 5.26) with the RGB camera at 30 m, and GLI (RMSE = 6.11) with the RGB camera at 30 m. The same RMSE values were observed in GLI and G (Table 3), whereas GLI was 6.106 and G was 6.111.

In the accuracy comparison of each multivariate regression model for height estimation, the estimation accuracies differed with the image and model type (Table 5). When using images from the RGB camera, SVR showed the lowest RMSE of 5.11 (R² = 0.944) at 30 m altitude and the lowest RMSE of 4.88 (R² = 0.814) at 50 m altitude (Table 5). With images from the MS camera, SVR showed the lowest RMSE of 11.3 (R² = 0.745) at 30 m altitude and the lowest RMSE of 11.1 (R² = 0.767) at 50 m altitude. For images from both cameras, SVR showed the lowest RMSE of 4.19 (R² = 0.956) at 30 m altitude, and MLR showed the lowest RMSE of 8.18 (R² = 0.681) at 50 m altitude.

Table 5

Height estimation accuracy using the multivariate-regression models for each image type, UAV flight altitude, and regression algorithms (RGB: four images of R, G, B, and H from the RGB camera; Multipspectral: four images of Green, Red, Red edge, and NIR from the MS camera; Both: the above eight images from both cameras).

Figure 5 summarizes the height estimation models with the lowest RMSE for each UAV flight altitude, regression algorithm, and image type, to compare the accuracy of models in all categories. The height estimation model with the lowest RMSE was the SVR model with images from both cameras at 30 m altitude (R² = 0.956, RMSE = 4.19) (Table 5; Figs. 5 and 6). Comparing the estimation accuracies of the mono- and multivariate regression models, when using images from the RGB camera at 30 m altitude, the multivariate regression of SVR had a lower RMSE (4.19) than that of the mono-regression of H (RMSE = 5.26), which had the lowest RMSE in mono-regression (Fig. 5; Tables 3 and 5). With the images from the RGB camera at 50 m altitude, the multivariate regression of MLR had a lower RMSE (4.88) than that of the mono-regression of MGRVI (RMSE = 5.01) at 50 m altitude (Fig. 5; Tables 3 and 5). With the images from MS camera, the multivariate regression of SVR had a higher RMSE (11.3) than the mono-regression of EVI2 (9.09) at 30 m altitude, and the multivariate regression of SVR had a higher RMSE (11.1) than that of the mono-regression of NIR (6.86) at 50 m altitude (Fig. 5; Tables 3 and 5).

Fig. 5

Summary of height estimation accuracy of the best models for each UAV flight altitude, regression algorithm, and image type.

Fig. 6

Comparison of ground measured and estimated height of test data using the SVR model using images from both cameras at 30 m flight altitude (Figure represents the seeded month (February) and days after seeding of test data).

Discussion

This study aimed to clarify the optimal estimation of the fresh weight and height of processing spinach using UAV imagery. After comparing mono-regression models with different image types, the estimation accuracy depended on image type and the optimal image type for a simple algorithm was identified. Comparing the multivariate regression models, machine learning regression was found to improve the estimation accuracy. The combination of images from the RGB and MS cameras conditionally improved the estimation accuracy. Comparing all models, the highest accuracy was found to be achieved by the machine learning model, and the optimal image type and regression algorithm were selected according to the estimation targets. The details are as follows.

First, we examined the estimation accuracy of the conventional mono-regression model using various image types. For fresh weight estimation, the lowest RMSE was achieved with the SAVI using an MS camera at 50 m altitude. For height estimation, the lowest RMSE was achieved with MGRVI using an RGB camera at 50 m altitude. The estimation accuracy was highly dependent on the image type, suggesting that selecting an appropriate image type is necessary to accurately estimate spinach growth. Comparing the image type for fresh weight estimation, the three lowest RMSE values were achieved with SAVI, OSAVI, and EVI2 from the MS camera. The reason for their high accuracy could be attributed to the NIR bands, which reflect plant biomass changes more than visible bands (Gitelson et al., 2002; de Oliveira et al., 2022; Zhou et al., 2018). Furthermore, SAVI and OSAVI have the ability to reduce the soil background noise, and they provided higher accuracy in estimating biomass and LAI (Liang et al., 2015; Ren et al., 2018). EVI2 can avoid spectral saturation problems (Bolton and Friedl, 2013; Jiang et al., 2008). These characteristics can contribute to higher accuracy compared to that with other image types. For height estimation, the three lowest RMSE values were achieved with MGRVI, H, and GLI from the RGB camera. Some past studies indicated that MGRVI showed a high correlation with plant growth parameters (Feng et al., 2022; Hammond et al., 2023), partially supporting the ability of MGRVI in accurate height estimation. However, the reason remains unclear, and further research is needed to clarify these details. To estimate plant height, past studies often used a crop surface model captured by an RGB camera or a light detection and ranging (LiDAR) sensor (ten Harkel et al., 2020; Sumesh et al., 2021). Similar to past studies, H (canopy height) from the RGB camera had the second highest accuracy in this study, suggesting that H could also be appropriate for estimating the height of spinach for processing use. In addition to these features, other factors such as ground sample distance (GSD) could have affected estimation accuracy. The GSD of UAV imagery depends on the camera resolution and UAV flight altitude. In this study, images from the RGB camera had a higher GSD than that from the MS camera, and images acquired at 30 m altitude had a higher GSD than that acquired at 50 m altitude. Although higher GSD generally contributed to high accuracy for crop growth monitoring (Zhai et al., 2023), this trend was not clear in the current study. These findings suggest that the effects of image type were more significant than those of GSD. Further comprehensive analysis is needed to clarify the individual and synergistic effects of image type and GSD on estimation accuracy. This study demonstrated that the best mono-regression model for fresh weight estimation was an exponential model with SAVI from an MS camera at an altitude of 50 m and that for height estimation was an exponential model with MGRVI from an RGB camera at an altitude of 50 m.

Next, we examined the application of machine learning and a combination of images from RGB and MS cameras to improve the estimation accuracy in multivariate regression. For fresh weight estimation, the machine learning models of RF and SVR showed a higher accuracy than that of the MLR model; this trend was consistent in different conditions in terms of camera type or UAV altitude. The higher accuracy of machine learning models could be attributable to their estimation ability. These models can deal with the non-linear relationship between image features and plant growth parameters (Liang et al., 2022; Zhang et al., 2022), in contrast to the MLR model, which assumes a linear relationship between them. Additionally, the MLR model could not use all the variables to avoid a multicollinearity problem, resulting in lower accuracy. The effectiveness of machine learning models using UAV imagery on plant growth estimation has also been reported in several studies (Li et al., 2019; Liang et al., 2022), supporting the results of this study. This study indicated that the machine learning model can contribute to improve the fresh weight estimation accuracy of processing spinach. For height estimation, although some of the machine learning models showed higher accuracy than that of the MLR model, some showed lower accuracy. In detail, the SVR model increased accuracy, whereas the RF model decreased accuracy, in five of six situations, compared to that of the MLR model. These results suggest that the effectiveness of the machine learning model depends on the model type, and that the SVR model could be better for height estimation. One of the reasons for lower accuracy in RF models could be overfitting. Indeed, the results showed that the RF models had higher R² values and lower RMSE values than those of the MLR model, suggesting an overfitting trend. A past study also observed overfitting in RF models on plant growth estimation because of high sensitivity over noise in the dataset (Taşan et al., 2022), supporting the current lower accuracy results. Accordingly, this study demonstrated that the application of machine learning could improve estimation accuracy for fresh weight and height of processing spinach and indicates that the selection of appropriate model types is important. Focusing on the combination of images from RGB and MS cameras, some of the models using images from both cameras showed higher accuracy for fresh weight and height estimation, compared to those with models using images from the RGB or MS camera alone. In the MLR model, accuracy did not change because the selected features were the same as in the model with images from the RGB camera. In the RF and SVR models, some models using image combination improved accuracy, but the effects were inconsistent. This may be because of the different conditions of target traits, model types, and UAV altitude, suggesting that further research is needed to clarify these details. Accurate estimation in some models could be obtained by providing more information from both cameras. Indeed, images from both cameras contain visible spectrum information, morphological information, and reflectance information relevant to crop growth. Past studies have revealed that more images could capture more details regarding plant growth status, resulting in improved yield and LAI prediction accuracy (Qiao et al., 2022; Wan et al., 2020). This study and past studies suggest that the combination of images can be helpful in the appropriate conditions. In summary, this study demonstrates that the selection of a suitable machine learning model is effective to improve the accuracy of fresh weight and height estimation for processing spinach. Although limited, the combination of images from RGB and MS cameras could increase estimation accuracy in appropriate conditions.

Finally, we compared the estimation accuracy of all models including mono- and multivariate-regression to identify the most accurate model considering practical use in the field. To achieve this, model accuracy was evaluated in this study using extrapolation data. From the comparison of all models with the highest accuracy in each category, some of the machine learning models achieved higher accuracy compared to that of conventional mono-regression models. The model with the highest accuracy for fresh weight estimation was the SVR model with images from the RGB camera at 30 m altitude, and that for height estimation was the SVR model with images from both cameras at 30 m altitude. These models could thus be suitable for precise measurements of the fresh weight and height of processing spinach. If necessary, the developed models can handle the demand for lower cost or higher efficiency by selecting a model with images from the RGB camera or with images obtained at 50 m altitude. Our findings could help reduce labor costs for fresh weight and height measurements in spinach fields, allowing systematic harvesting of spinach according to processing standards. However, further research is necessary for more practical applications to achieve higher accuracy. To reflect the growth status more and improve the robustness of the developed model, the model should be retrained with various datasets, including different growth periods (e.g. February and September seedings), cultivar conditions, and planting densities. Additionally, more precise models such as deep learning or collaboration with different growth indicators could be used to increase applicability over extrapolation data (Desloires et al., 2023; Zhu et al., 2022). Such issues should be explored in future studies.

Acknowledgements

The authors acknowledge the members of Operation Unit 3 of the Technical Support Center, and Mr. Noriyoshi Kamou, Ms. Tomoko Higashinaka, and Ms. Kanako Hachikubo for their technical assistance with spinach cultivation and surveys.

Literature Cited

•  Astor,  T.,  S.  Dayananda,  S.  Nautiyal and  M.  Wachendorf. 2020. Vegetable crop biomass estimation using hyperspectral and RGB 3D UAV data. Agronomy 10: 1–16.
•  Bendig,  J.,  K.  Yu,  H.  Aasen,  A.  Bolten,  S.  Bennertz,  J.  Broscheit,  M. L.  Gnyp and  G.  Bareth. 2015. Combining UAV-based plant height from crop surface models, visible, and near infrared vegetation indices for biomass monitoring in barley. Int. J. Appl. Earth Observ. Geoinform. 39: 79–87.
•  Bolton,  D. K. and  M. A.  Friedl. 2013. Forecasting crop yield using remotely sensed vegetation indices and crop phenology metrics. Agric. Forest Meteorol. 173: 74–84.
•  Cao,  Q.,  Y.  Miao,  H.  Wang,  S.  Huang,  S.  Cheng,  R.  Khosla and  R.  Jiang. 2013. Non-destructive estimation of rice plant nitrogen status with Crop Circle multispectral active canopy sensor. Field Crops Res. 154: 133–144.
•  Cheng,  Q.,  H.  Xu,  S.  Fei,  Z.  Li and  Z.  Chen. 2022. Estimation of maize LAI using ensemble learning and UAV multispectral imagery under different water and fertilizer treatments. Agriculture 12: 1267. DOI: 10.3390/agriculture12081267.
•  de Oliveira,  R. P.,  M. R.  Barbosa Jr.,  A. A.  Pinto,  J. L. P.  Oliveira,  C.  Zerbato and  C. E. A.  Furlani. 2022. Predicting sugarcane biometric parameters by UAV multispectral images and machine learning. Agronomy 12: 1992. DOI: 10.3390/agronomy12091992.
•  Desloires,  J.,  D.  Ienco and  A.  Botrel. 2023. Out-of-year corn yield prediction at field-scale using Sentinel-2 satellite imagery and machine learning methods. Comput. Electron. Agric. 209: 107807. DOI: 10.1016/j.compag.2023.107807.
•  Feng,  H.,  H.  Tao,  Z.  Li,  G.  Yang and  C.  Zhao. 2022. Comparison of UAV RGB imagery and hyperspectral remote-sensing data for monitoring winter wheat growth. Remote Sens. 14: 3811. DOI: 10.3390/rs14153811.
•  Fitzgerald,  G. J.,  D.  Rodriguez,  L. K.  Christensen,  R.  Belford,  V. O.  Sadras and  T. R.  Clarke. 2006. Spectral and thermal sensing for nitrogen and water status in rainfed and irrigated wheat environments. Precision Agric. 7: 233–248.
•  Gitelson,  A. A. 2004. Wide dynamic range vegetation index for remote quantification of biophysical characteristics of vegetation. J. Plant Physiol. 161: 165–173.
•  Gitelson,  A. A.,  Y. J.  Kaufman and  M. N.  Merzlyak. 1996. Use of a green channel in remote sensing of global vegetation from EOS-MODIS. Remote Sens. Environ. 58: 289–298.
•  Gitelson,  A. A.,  Y. J.  Kaufman,  R.  Stark and  D.  Rundquist. 2002. Novel algorithms for remote estimation of vegetation fraction. Remote Sens. Environ. 80: 76–87.
•  Hague,  T.,  N. D.  Tillett and  H.  Wheeler. 2006. Automated crop and weed monitoring in widely spaced cereals. Precision Agric. 7: 21–32.
•  Hammond,  K.,  R.  Kerry,  R. R.  Jensen,  R.  Spackman,  A.  Hulet,  B. G.  Hopkins,  M. A.  Yost,  A. P.  Hopkins and  N. C.  Hansen. 2023. Assessing within-field variation in alfalfa leaf area index using UAV visible vegetation Indices. Agronomy 13: 1289. DOI: 10.3390/agronomy13051289.
•  Huete,  A. 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25: 295–309.
•  Jiang,  Z.,  A. R.  Huete,  K.  Didan and  T.  Miura. 2008. Development of a two-band enhanced vegetation index without a blue band. Remote Sens. Environ. 112: 3833–3845.
•  Jin,  Z.,  R.  Prasad,  J.  Shriver and  Q.  Zhuang. 2017. Crop model- and satellite imagery-based recommendation tool for variable rate N fertilizer application for the US Corn system. Precision Agric. 18: 779–800.
•  Jones,  C. L.,  P. R.  Weckler,  N. O.  Maness,  R.  Jayasekara,  M. L.  Stone and  D.  Chrz. 2007. Remote sensing to estimate chlorophyll concentration in spinach using multi‐spectral plant reflectance. Trans. ASABE 50: 2267–2273.
•  Kamada,  E.,  T.  Ishii and  K.  Okada. 2021. Effect of temperature and solar radiation on dry matter production of spinach for processing use. Hor. Res. (Japan) 20: 423–432 (In Japanese with English abstract).
•  Kobayashi,  S. 2006. The trend of vegetable demand by intended purposes and the main issues of domestic production. J. Agric. Policy Res. 11: 1–27 (In Japanese with English abstract).
•  Koyano,  H. and  E.  Kamada. 2022. Development of a system for predicting and simulating the growth of spinach (Spinacia oleracea). Agric. Inf. Res. 31: 1–12.
•  Li,  S.,  F.  Yuan,  S. T.  Ata-UI-Karim,  H.  Zheng,  T.  Cheng,  X.  Liu,  Y.  Tian,  Y.  Zhu,  W.  Cao and  Q.  Cao. 2019. Combining color indices and textures of UAV-based digital imagery for rice LAI estimation. Remote Sens. 11: 1763. DOI: 10.3390/rs11151763.
•  Liang,  L.,  L.  Di,  L.  Zhang,  M.  Deng,  Z.  Qin,  S.  Zhao and  H.  Lin. 2015. Estimation of crop LAI using hyperspectral vegetation indices and a hybrid inversion method. Remote Sens. Environ. 165: 123–134.
•  Liang,  Y.,  W.  Kou,  H.  Lai,  J.  Wang,  Q.  Wang,  W.  Xu,  H.  Wang and  N.  Lu. 2022. Improved estimation of aboveground biomass in rubber plantations by fusing spectral and textural information from UAV-based RGB imagery. Ecol. Ind. 142: 109286. DOI: 10.1016/j.ecolind.2022.109286.
•  Liu,  Y. J.,  Y. P.  Tong,  Y. G.  Zhu,  H.  Ding and  F.  Smith. 2006. Leaf chlorophyll readings as an indicator for spinach yield and nutritional quality with different nitrogen fertilizer applications. J. Plant Nutr. 29: 1207–1217.
•  Louhaichi,  M.,  M. M.  Borman and  D. E.  Johnson. 2001. Spatially located platform and aerial photography for documentation of grazing impacts on wheat. Geocarto Int. 16: 65–70.
• MAFF. 2023. ‘The situation of Processing and Commercial use of Vegetables’. <https://www.maff.go.jp/j/seisan/kakou/yasai_kazitu/attach/pdf/index-12.pdf> browsed on July 24, 2023 (In Japanese).
•  Moeckel,  T.,  S.  Dayananda,  R. R.  Nidamanuri,  S.  Nautiyal,  N.  Hanumaiah,  A.  Buerkert and  M.  Wachendorf. 2018. Estimation of vegetable crop parameter by multi-temporal UAV-borne images. Remote Sens. 10: 1–18.
•  Naghdyzadegan Jahromi,  M.,  S.  Zand-Parsa,  F.  Razzaghi,  S.  Jamshidi,  S.  Didari,  A.  Doosthosseini and  H. R.  Pourghasemi. 2023. Developing machine learning models for wheat yield prediction using ground-based data, satellite-based actual evapotranspiration and vegetation indices. Eur. J. Agron. 146: 126820. DOI: 10.1016/j.eja.2023.126820.
•  Ohno,  H.,  K.  Sasaki,  G.  Ohara and  K.  Nakazono. 2016. Development of grid square air temperature and precipitation data compiled from observed, forecasted, and climatic normal data. Clim. Biosph. 16: 71–79 (In Japanese with English abstract).
•  Qiao,  L.,  D.  Gao,  R.  Zhao,  W.  Tang,  L.  An,  M.  Li and  H.  Sun. 2022. Improving estimation of LAI dynamic by fusion of morphological and vegetation indices based on UAV imagery. Comput. Electron. Agric. 192: 106603. DOI: 10.1016/j.compag.2021.106603.
•  Ren,  H.,  G.  Zhou and  F.  Zhang. 2018. Using negative soil adjustment factor in soil-adjusted vegetation index (SAVI) for aboveground living biomass estimation in arid grasslands. Remote Sens. Environ. 209: 439–445.
•  Steven,  M. D. 1998. The sensitivity of the OSAVI vegetation index to observational parameters. Remote Sens. Environ. 63: 49–60.
•  Sumesh,  K. C.,  S.  Ninsawat and  J.  Som-ard. 2021. Integration of RGB-based vegetation index, crop surface model and object-based image analysis approach for sugarcane yield estimation using unmanned aerial vehicle. Comput. Electron. Agric. 180: 105903. DOI: 10.1016/j.compag.2020.105903.
•  Taşan,  S.,  B.  Cemek,  M.  Taşan and  A.  Cantürk. 2022. Estimation of eggplant yield with machine learning methods using spectral vegetation indices. Comput. Electron. Agric. 202: 107367. DOI: 10.1016/j.compag.2022.107367.
•  ten Harkel,  J.,  H.  Bartholomeus and  L.  Kooistra. 2020. Biomass and crop height estimation of different crops using UAV-based LiDAR. Remote Sens. 12: 17. DOI: 10.3390/rs12010017.
•  Tsukaguchi,  T.,  H.  Kobayashi,  Y.  Fujihara and  S.  Chono. 2022. Estimation of spikelet number per area by UAV-acquired vegetation index in rice (Oryza sativa L.). Plant Prod. Sci. 25: 20–29.
•  Tucker,  C. J. 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 8: 127–150.
•  Wan,  L.,  H.  Cen,  J.  Zhu,  J.  Zhang,  Y.  Zhu,  D.  Sun,  X.  Du,  L.  Zhai,  H.  Weng,  Y.  Li,  X.  Li,  Y.  Bao,  J.  Shou and  Y.  He. 2020. Grain yield prediction of rice using multi-temporal UAV-based RGB and multispectral images and model transfer—a case study of small farmlands in the South of China. Agric. For. Meteorol. 291: 108096. DOI: -->10.1016/j.agrformet.2020.108096.
•  Yang,  K.,  Y.  Gong,  S.  Fang,  B.  Duan,  N.  Yuan,  Y.  Peng,  X.  Wu and  R.  Zhu. 2021. Combining spectral and texture features of UAV images for the remote estimation of rice LAI throughout the entire growing season. Remote Sens. 13: 3001. DOI: 10.3390/rs13153001.
•  Zhai,  W.,  C.  Li,  Q.  Cheng,  B.  Mao,  Z.  Li,  Y.  Li,  F.  Ding,  S.  Qin,  S.  Fei and  Z.  Chen. 2023. Enhancing wheat above-ground biomass estimation using UAV RGB images and machine learning: multi-feature combinations, flight height, and algorithm implications. Remote Sens. 15: 3653. DOI: 10.3390/rs15143653.
•  Zhang,  X.,  K.  Zhang,  Y.  Sun,  Y.  Zhao,  H.  Zhuang,  W.  Ban,  Y.  Chen,  E.  Fu,  S.  Chen,  J.  Liu and  Y.  Hao. 2022. Combining spectral and texture features of UAS-based multispectral images for maize leaf area index estimation. Remote Sens. 14: 331. DOI: 10.3390/rs14020331.
•  Zhou,  Z.,  M.  Jabloun,  F.  Plauborg and  M. N.  Andersen. 2018. Using ground-based spectral reflectance sensors and photography to estimate shoot N concentration and dry matter of potato. Comput. Electron. Agric. 144: 154–163.
•  Zhu,  W.,  E. E.  Rezaei,  H.  Nouri,  Z.  Sun,  J.  Li,  D.  Yu and  S.  Siebert. 2022. UAV-based indicators of crop growth are robust for distinct water and nutrient management but vary between crop development phases. Field Crops Res. 284: 108582. DOI: 10.1016/j.fcr.2022.108582.