Introduction

The benefits that an accurate vineyard yield estimation can bring to the entire grape and wine production chain are well-known (Clingeleffer et al., 2001; Whitty et al., 2017). Among others, the following applications of vineyard yield estimation should be highlighted: planning cluster thinning needs (in order to prevent excessive production and consequent poor wine quality); planning and organisation of the harvest (labour, equipment, others); planning cellar needs (scheduling grape intake, allocating tank space, purchasing tanks, barrels, oenological products, bottles and others); planning purchases and/or grape sales; establishment of grape prices and management of wine stocks; management of grape and wine market and programming investments and development of marketing strategies. This multiplicity of potential applications and benefits means that yield estimation and forecasting play an essential role in viticulture and the wine industry.

Of the several methods that can be used to estimate vineyard yield, those based on the estimation of yield components (e.g., number of berries or bunches and bunch weight) are the most widely used (Whitty et al., 2017). These methods can be used at different phenological stages, from bud break to harvest. One of the best examples is the method based on the counting of bunches in canopy sample segments, combined with bunch weight evaluation obtained by destructive sampling at onset of veraison (lag phase) (Clingeleffer et al., 2001; Martin et al., 2003; de la Fuente et al., 2015). Based on historical data, bunch weight at harvest can then be obtained using a berry growth factor and yield at harvest estimated by multiplying the estimated number of bunches by the estimated bunch weight at harvest (Clingeleffer et al., 2001). However, besides being destructive, this practice is manual, and thus very labour-intensive, and it can prove inaccurate, as it extrapolates for the entire vineyard the assessment performed on a small percentage of bunches.

Recent advances in machine vision and machine learning technologies have provided a multitude of automated tools for recognition and quantitative estimation of several yield components, mainly bunch and berry traits in 2D images (for a review see Seng et al., 2018). The images are taken using RGB digital cameras deployed on terrestrial vehicles, and they are then processed using machine vision technologies in order to differentiate the yield components (e.g., shoots, flowers, bunches and berries) from other canopy components. After an accurate segmentation, multiple traits and metrics can automatically be extracted from the yield components images, which can then be used as proxies for vineyard yield estimation.

Total berry number is also considered a very good vineyard yield predictor as it includes variations in the number of bunches and in the number of berries per bunch (Clingeleffer et al., 2001; Whitty et al., 2017). The detection and counting of visible grape berries in 2D images (#vBe) is nowadays feasible by automated non-invasive image analysis and machine learning techniques (e.g., Nuske et al., 2014; Diago et al., 2015; Rose et al., 2016; Aquino et al., 2017; Aquino et al., 2018; Pérez-Zavala et al., 2018; Zabawa et al., 2020); for example, when using neural network semantic segmentation, Zabawa et al. (2020) achieved between 85.6 % and 94.0 % accuracy (depending on the training system) for non-invasive berry detection. However, as only a fraction of the berries is visible in an image, estimating actual total berry number per bunch (#Be) would require hidden berries to be estimated using modelling techniques to extract 3D information from 2D images (Schöler and Steinhage, 2015; Rist et al., 2018; Xin et al., 2020; among others) or, alternatively, the use of statistical models based on relationships obtained between #vBe and #Be (Liu et al., 2020). Once #Be has been estimated, yield can be estimated using an equation previously obtained from a linear relationship between #Be and actual yield (Nuske et al., 2014), or by multiplying #Be by an estimation of berry mass, which can be obtained using automated 2D and 3D image post-processing techniques (Cubero et al., 2015; Ivorra et al., 2015; Rose et al., 2016; Rist et al., 2018; among others).

Regarding the bunches, several studies have demonstrated that the 2D image bunch area (BuA) - expressed in pixels - is well correlated with bunch weight (Dunn and Martin, 2004; Diago et al., 2012; Font et al., 2015, Hacking et al., 2019; Victorino et al., 2020); this allows regression models to be obtained that show a good fit between estimated and actual yield in conditions of no bunch occlusions by the canopy and/or other bunches. For example, using RGB images and a supervised classifier based on the Mahalanobis distance to assess the yield of field-grown grapevines, Diago et al. (2012) obtained a determination coefficient of 0.76 for the relationship between the actual yield in the images and the corresponding number of pixels.

Bunch volume is also a feature that has been much explored in recent studies using several machine vision and modelling techniques to extract 3D information from 2D images (Nuske et al., 2014; Font et al., 2015; Herrero-Huerta et al., 2015; Ivorra et al., 2015; Tello et al., 2016; Rist et al., 2018). However, the automatic extraction of the volumetric information is a much more complex computational task than the extraction of the bi-dimensional projection of the bunch.

Despite the aforementioned advances in machine vision, 3D modelling and other methods for the automated estimation of grapevine yield components, to obtain a general model to convert image-visible bunch attributes into bunch mass is still challenging. This is mainly due to the naturally large variation that can be found in bunch architecture and compactness among cvs. (Tello et al., 2015).

The aim of this paper is to assess and compare the explanatory power of the two most used image-based bunch attributes (#vBe and BuA) in order to predict grapevine bunch weight at harvest. 2D image-based bunch attributes were extracted in laboratory conditions by visual photo-interpretation (#vBe) and manual segmentation (BuA) from 300 images of four grapevine cultivars. These variables were then used to find accurate predictive models for bunch weight, which can also be useful in assisting the algorithms used for vineyard yield estimation based on fully automated image analysis.

Materials and methods

2. Laboratory Assessments

In the laboratory, each bunch was positioned in a random position in front of a white background and imaged with a corresponding tag, using a digital RGB camera located at a constant distance from the bunches. Then actual bunch weight (BuW) and volume (BuV) were assessed using a table scale (KERN FCB v1.4) and the water displacement method respectively. The bunches were then destemmed by hand and the following attributes were assessed per bunch: number of berries (#Be), total berry weight (BeW) and rachis main axis length (RaL). Average weight per berry (avBeW) and a bunch compactness (BuC) index were derived (Table 1) from the obtained data.

The images were manually labelled using ImageJ® software and the total area of grapes was computed in pixels and converted into cm² of bunch area using the tag dimensions (8.0 x 1.8 cm coloured rectangle) to calculate the conversion ratio. Furthermore, the number of visible berries in each image (including partially hidden ones) was visually assessed and manually labelled and counted.

For model validation, two independent bunch datasets were used; one from the white cv. ‘Viosinho’ collected in the 2014 season (VI14; n = 100) and the other from the red cv. ‘Cabernet-Sauvignon’ collected in the 2015 season (CS15; n = 109). Both validation datasets were collected in the same vineyard plots and subjected to similar cultural practices to those described for the training datasets.

FIGURE 1. 220dpi. Examples of three bunches of different sizes (small, medium and large) of each of the five cultivars used in this work, showing the bunch morphological diversity of each cultivar. Actual tag dimensions: 8.0 x 1.8 cm.

Save View full size Expand inline Collapse inline

Results

1. Variability and correlations of bunch attributes

The boxplots with descriptive statistics of the actual and image-based bunch attributes for the four cvs. show considerable variability in all the assessed bunch attributes in all training datasets (Figure 2). Regarding BuW, the VI dataset had the highest median value, while TN showed the lowest (Figure 2A). #Be was very similar among cultivars with medians ranging from 79 for TN to 94 for VI (Figure 2B). The avBeW variable was found to have much higher variability among cvs., with the highest median value found for VI, closely followed by SY, and the lowest for TN (Figure 2C). The highest BuV median was also found for the cvs. VI and SY and the lowest for TN (Figure 2D). The highest RaL median was found for cv. SY, closely followed by VI and TN, and the lowest for AL (Figure 2E). The highest median for number of berries per rachis length - an objective measure of bunch compactness (Hed et al., 2009; Tello and Ibáñez, 2014) - was found for cv. AL, while the other three cvs. (Figure 2F). Of the actual bunch attributes, BuW, #Be and BuV showed the highest variability, with the SY and AL datasets having the highest coefficient of variation (CV) (from 42.5 % to 46.6 %), and VI the lowest (from 35.1 to 37.8 %). The avBeW had the lowest variability, with the highest CV being found for TN (16.8 %) and the lowest for AL (8.5 %). CV was intermediate for the variables RaL and BuC.

Regarding the image-based attributes, the relative differences among cvs. for the variables #vBe and BuA mirrored those described for #Be and BuW respectively (Figures 2G and 2H). The cv. SY showed the highest median for the derived variable percentage of visible berries (%vBe), closely followed by TN, while VI had the lowest median and AL an intermediate value (Figure 2I).

Figure 2. Box plots showing descriptive statistics of the actual and image-based bunch attributes for each of the four cultivars used for the training dataset (n = 75). VI = ‘Viosinho’, AL = ‘Alvarinho’, SY = ‘Syrah, and TN = ‘Touriga Nacional’.

Save View full size Expand inline Collapse inline

Regarding the CV, #vBe and BuA had the highest variability, with the SY dataset having the highest CV (~ 38 %) and the other three cultivars showing a CV of similar magnitude (ranging from 30.1 to 32.9 %). The %vBe variable had the lowest variability, with SY having the highest CV (19.0 %) and VI the lowest (12.3 %).

Table 2 shows the correlation matrix (for original and log-transformed data) for a selected set of variables that encompasses actual bunch attributes and image-based features extracted from the 2D images for the pooled dataset of 300 bunches from the four cultivars. The variables representing actual bunch attributes were all significantly (p < 0.001) and positively correlated with BuW, with BuV having the highest correlation coefficient (r), followed by #Be. The avBeW variable was the least correlated with BuW, while the r value was intermediate but significant and positive for the calculated variable BuC.

In terms of the correlations between the variables extracted from the 2D images (BuA and #vBe) and BuW, very high, positive and significant (p < 0.001) r values were found. The correlation between the derived variable %vBe and BuW showed a lower and negative r value (Table 2).

Besides the correlations with BuW, it is worth highlighting some other relationships, mainly those between the variables extracted from 2D images. There were very strong, positive and significant (p < 0.001) correlations between the variable BuA and the variables #Be, BuV, RaL and #vBe. Very high, positive and significant correlations were also found between the variable #vBe and the variables #Be, BuV and RaL. Furthermore, significant (p < 0.01) but negative correlations were found between %vBe and all the other assessed variables (Table 2A).

The correlation coefficients between the log-transformed variables (Table 2B) were in general of higher magnitude than between the original pairs of corresponding variables, indicating stronger linearity in the relationships within the log-transformed data; this is a result that will be subsequently used.

In general, the correlation analyses performed separately per cv. (n = 75) yielded results that mirror those reported in Table 2 for the pooled data. However some differences among cvs. were observed in terms of the magnitude and significance of the correlation coefficients linked to BuW (Figure 3A). Regarding actual bunch attributes, the r value was much lower (0.17) and non-significant when correlating avBeW with BuW in the SY dataset compared to the other three cv. datasets, in which higher and significant r (p < 0.001) was obtained for the same variables. The opposite result was obtained for RaL in correlation with BuW: the SY dataset showed the highest r value and VI the lowest. Regarding the correlation of image-based variables with BuW, the r value was very high and significant (p < 0.001) for all four cv. datasets, with the highest r (0.95) for BuA in cv. AL and the lowest (0.82) for #vBe in cv. TN (Figure 3A).

Table 2. Correlograms showing Pearson correlation coefficients between pairs of a selected set of variables for the pooled dataset of the four cvs. [original (A) and log-transformed data (B); n = 300]. The set of variables include actual bunch attributes [BuW = bunch weight (g), #Be = number of berries per bunch, avBeW = average berry weight (g), BuV = bunch volume (mL), RaL = rachis length (cm), BuC = bunch compactness] and bunch attributes extracted from 2D images [BuA = bunch area (cm2), #vBe = number of visible berries] or derived (%vBe = #vBe/#Be x 100). All correlation coefficients of magnitude greater than |0.11| are significantly different from zero (p < 0.05). Correlations of similar magnitude have comparable colours (according to the colour code shown at the bottom).

Save View full size Expand inline Collapse inline

2. Testing the use of visible berries on the image as a proxy of actual number of berries

The average values regarding the proportion of visible berries on the 2D images were similar for VI and AL (49.0 and 50.5 % respectively), but they were lower than those obtained in the TN and SY datasets (55.8 and 57.5 % respectively).

When creating the scatterplot fitted with linear regressions of the #Be (response variable) over #vBe (predictor variable), a linear relationship between both variables appeared acceptable for all datasets, despite there being a slight curvature. However, the constant variance assumption underlying the regression analysis was violated. The logarithmic transformation of both variables stabilised the variance and improved the linearity in the four datasets (Figure 3B). Table 3 shows the fitted regression models of the log-transformed data, with very high and significant coefficients of determination (R²) in all cases.

Figure 3. (A) Pearson correlation coefficients between bunch weight and all other bunch assessed attributes of the four cvs. With the exception of the correlation between avBeW and BuW in cv. ‘Syrah’, all coefficients are significant (p < 0.01). VI = ‘Viosinho’ (green), AL = ‘Alvarinho’ (blue), SY = ‘Syrah’ (red), TN = ‘Touriga Nacional’ (purple). For variable legend see Table 2. (B) Scatter plot showing the fitted linear relationship of logarithm of actual number of berries per bunch [ln(#Be)] versus the logarithm of the number of visible berries on the image [ln(#vBe)], for each cv. (thin coloured lines) and for pooled data (thick black line). n = 75 per cv.

Save View full size Expand inline Collapse inline

Table 3. Fitted power law models (in the original, non-logarithmic scales) resulting from the fitted linear regressions of the logarithm of actual number of berries per bunch (#Be) over the logarithm of the number of visible berries on the 2D image (#vBe), for each of the four cvs. (n = 75) and for the pooled data (n = 300). Values of the coefficient of determination (R2) are relative to the linear regression in the log-scale and are always significant (p < 0.001).

View popup Expand inline Collapse inline

Dataset	Fitted model (Eq. nº)	R2
‘Viosinho’	#Be = 1.162 * #vBe1.152 (6)	0.91
‘Alvarinho’	#Be = 0.727 * #vBe1.27 (7)	0.91
‘Syrah’	#Be = 0.792 * #vBe1.215 (8)	0.92
‘Touriga Nacional’	#Be = 0.715 * vBe1.254 (9)	0.92
Pooled	#Be = 0.809 * #vBe1.230 (10)	0.90

3. Using image-based variables to estimate bunch weight

3.1. Number of visible berries on the 2D image

Figure 4A shows a scatterplot of the variables BuW and #vBe for all training datasets, showing an increase in the scatterpoint dispersion as the values of both variables increase, as well as a slightly curvilinear trend. The violation of the constant variance assumption was noticed for all datasets, and therefore the logarithmic transformations of both variables were adopted to stabilise variances and improve linearity (Figure 4B). The fitted regression models (Table 4) show that the highest R² was obtained for the AL model and the lowest for the TN and pooled models. Standard t-tests on the ANCOVA model, which allows each cultivar to have its own linear regression between ln(BuW) and ln(#vBe), indicate that the differences between the cultivar-specific slope (exponent in Table 4) for cv. AL and the remaining cvs. are slightly significant (0.02 < p < 0.10), but they are not significant for the other cvs.

Figure 4. Scatter plots showing (A) the fitted power relationship of bunch weight (BuW) versus the number of visible berries on the 2D image (#vBe), and (B) the fitted linear relationship of logarithm of bunch weight [ln(BuW)] versus the logarithm of the number of visible berries on the image ([ln(#vBe)], for each cv. (thin coloured lines) and for pooled data (thick black line). Colour codes: green = ‘Viosinho’, blue = ‘Alvarinho ‘, red = ‘Syrah’, and purple = ‘Touriga Nacional’. n = 75 per cv.

Save View full size Expand inline Collapse inline

Table 4. Fitted power law models (in the original, non-logarithmic, scales) resulting from the fitted linear regressions of the logarithm of bunch weight (BuW; g) over the logarithm of the number of visible berries on the 2D image (#vBe), for each of the four cvs. (n = 75) and for the pooled data (n = 300). Values of the coefficient of determination (R2) are relative to the linear regression in the log-scale and are always significant (p < 0.001).

View popup Expand inline Collapse inline

Dataset	Fitted model (Eq. nº)	R2
‘Viosinho’	BuW = 1.513 * #vBe1.232 (11)	0.85
‘Alvarinho’	BuW = 0.659 * #vBe1.403 (12)	0.87
‘Syrah’	BuW = 1.410 * #vBe1.210 (13)	0.86
‘Touriga Nacional’	BuW = 1.183 * #vBe1.201 (14)	0.79
Pooled	BuW = 1.077 * #vBe1.276 (15)	0.80

3.2. Bunch area in the 2D image

Figure 5A shows the scatter plot of BuW and BuA for all training datasets. Visually, there is an increase in the scatterpoint dispersion as the value of both variables increases and also a slight curvilinear trend. Indeed, the constant variance assumption of the linear regression model relating BuW (response variable) and BuA (predictor variable) appears unrealistic, as was also observed with #vBe relationships, in particular for the VI and SY datasets. In order to improve linearity and stabilise variances, the logarithmic transformation of both variables was used in all cases, thus ensuring comparable models for all four datasets and for the pooled data (Figure 5B).

Table 5 shows the R² values for the linear regressions of ln(BuW) on ln(BuA), as well as the fitted power relationships that result from converting back to the original (non-logarithmic) scales for both variables. In all four datasets, R² was very high and significant. The ANCOVA model allowing each cv. to have its own relationship indicates that the slopes (exponents in Table 5) for the cvs. AL and VI are significantly different (p = 0.0025), with less or non significant differences among other pairs of cvs.

Figure 5. Scatter plots showing: (A) the fitted power relationship of bunch weight (BuW) versus the bunch area on the 2D image (BuA), and (B) the fitted linear relationship of logarithm of bunch weight [ln(BuW)] versus the logarithm of bunch area on the 2D image [ln(BuA)], for each cv. (thin colored lines) and for pooled data (thick black line). Colour codes: green = ‘Viosinho’, blue = ‘Alvarinho ‘, red = ‘Syrah’, and purple = ‘Touriga Nacional’. n = 75 per cv.

Save View full size Expand inline Collapse inline

Table 5. Fitted power law models (in the original, non-logarithmic, scales) resulting from the fitted linear regressions of the logarithm of bunch weight (BuW; g); over the logarithm of bunch area (BuA; cm2), for each of the four cvs. (n = 75) and for the pooled data (n = 300). Values of the coefficient of determination (R2) are relative to the linear regression in the log-scale and are always significant (p < 0.001).

View popup Expand inline Collapse inline

Dataset	Fitted model (Eq. nº)	R2
‘Viosinho’	BuW = 1.085 * BuA1.119 (16)	0.83
‘Alvarinho’	BuW = 0.454 * BuA1.357 (17)	0.91
‘Syrah’	BuW = 0.574 * BuA1.258 (18)	0.92
‘Touriga Nacional’	BuW = 0.497 * BuA1.288 (19)	0.89
Pooled	BuW = 0.653 * BuA1.238 (20)	0.88

The comparison of the R² values between the two single-predictor regressions (Table 4 vs Table 5) shows that, for three cvs., the best individual predictor of ln(BuW) is ln(BuA), the exception being VI.

3.3. Combining image-based explanatory variables

Based on the high performance of the two variables extracted from the 2D images (BuA and #vBe) a multiple linear regression model was used to estimate ln(BuW), with both log-transformed variables as predictors. As shown in Table 6, in general the R² values of the two-predictor regressions were significantly higher than those for the best single-predictor regressions, except in the case of the predictor ln(BuA) for SY and TN, which showed an R² of similar magnitude. While these improvements in the R² coefficient were significant, they were often modest and more pronounced for the models based on the predictor #vBe (gains ranging from 0.02 for VI to 0.10 for TN) than for the models based on BuA, (gains of 0.04, 0.02 and 0.01 in VI, AL and pooled models respectively, or near-zero for SY and TN).

Table 6. Fitted power law models resulting from the multiple linear regression of the logarithm of bunch weight (BuW, g) over two predictors: the logarithm of bunch area (BuA; cm2) and the logarithm of the number of visible berries on the image (#vBe). Models are shown for each of the four cvs. (n = 75) and for the pooled data (n = 300). Values of the coefficient of determination (R2) are relative to the linear regression in the log-scale. Values of the adjusted R2 differ by, at most, 0.01. The exponent of predictor BuA is in all cases significantly different from zero (p < 0.001). The p-values in the standard t-tests for the exponents of #vBe are shown in the right column and are not significantly different from zero for ‘Syrah’ and ‘Touriga Nacional’.

View popup Expand inline Collapse inline

Dataset	Fitted model (Eq. nº)	R2	p-value for the coefficient of the predictor ln(#vBe)
‘Viosinho’	BuW = 1.105 * BuA0.501 * #vBe0.725 (21)	0.87	< 0.001
‘Alvarinho’	BuW = 0.420 * BuA0.873 * #vBe0.557 (22)	0.93	< 0.001
‘Syrah’	BuW = 0.605 * BuA1.134 * #vBe0.129 (23)	0.92	0.392
‘Touriga Nacional’	BuW = 0.501 * BuA1.257 * #vBe 0.032 (24)	0.89	0.834
Pooled	BuW = 0.611 * BuA0.939 * #vBe0.358 (25)	0.89	< 0.001

3.4. Model Validation

In order to test the potential of the above fitted pooled models by applying different conditions and cultivars than those used during the fitting process, a validation was performed using two independent datasets (VI14 and CS15). Figure 6 shows the scatter plots of the observed vs fitted values (on the original scale) of #Be (Figure 6A) and BuW (Figures 6B, 6C and 6D) using the fitted models for the pooled data (Equations 10, 15, 20 and 25 respectively). It can be observed that overall there is a good agreement between the observed and fitted values for the two datasets, as well as some variability that increases with an increase in magnitude of the observed values.

Regarding the model that estimates #Be using #vBe as a predictor (Figure 6A), the scatter points are very close to the 1/1 line, indicating a very good overall fit for both datasets. However, a closer look reveals an underestimation of #Be in the VI14 dataset and an overestimation in CS15 (less pronounced). The statistical measures of validation confirmed this visual appraisal showing higher values of MAE (12.7 vs 7.8), NMAE (0.13 vs 0.09), RMSE (17.5 vs 10.0) and Pbias (-8.8 vs 4.4) and a lower EF (0.89 vs 0.94) for VI14 when compared to CS15.

The scatterplot related to the model that estimates BuW based on the predictor #vBe (Figure 6B, Equation 15) clearly indicates a different kind of model bias for each validation dataset, with an underestimation of BuW in VI14 and an overestimation in CS15. In the other two scatterplots involving the estimation of BuW (Figures 6C and 6D), the dispersion of points in relation to the 1/1 line, seems to show that the model based on both predictors produces the best fit for both cvs.

The linear regression between observed (dependent variable) and estimated (independent variable) values obtained for each of the four pooled models (Equations 10, 15, 20 and 25) also shows very high and significant R² and an intercept and slope very close to 0 and 1 respectively (data not shown).

The results shown in the scatterplots with the fitted values from the models for BuW prediction are corroborated by the statistical validation measures, which generally indicate a fairly good fit (Table 7). When comparing the validation performance of the two single predictor models, it can be observed that, for the combined validation dataset (both cultivars), the model based on #vBe (Equation 15) has a higher MAE and NMAE, a similar RMSE and EF and a smaller Pbias, than the model based on BuA. When comparing the same single predictor models with each validation dataset, it can be observed that for VI14 the model based on the predictor #vBe showed smaller errors (MAE, NMAE and RMSE), an almost similar Pbias and a higher EF than the model based on the predictor BuA. For the CS15 dataset a better performance of the model based on the predictor BuA was observed in all the statistical measures of validation. It also becomes apparent that the much smaller Pbias of the combined validation dataset using predictor #vBe masks strong biases of opposite signs for each cultivar. Regarding the validation performance of the model based on the two explanatory variables (Equation 25), smaller errors (MAE, NMAE and RMSE) and higher EF were observed in the combined validation dataset compared to the two single predictor models. When individual cultivars were considered, in the CS15 dataset the single-predictor BuA model performed slightly better than the model with both predictors, but there were clear advantages on all indicators in the VI14 dataset (Table 7).

Figure 6. Relationship between observed and estimated (A) berry number per bunch (#Be) and (B, C, D) bunch weight (BuW) for two independent datasets: ‘Viosinho’, 2014 (n = 100, green circles) and ‘Cabernet-Sauvignon’, 2015 (n = 109, red squares) using the pooled models represented in equations 10 (A), 15 (B), 20 (C) and 25 (D). The dotted black lines represent the 1:1 lines.

Save View full size Expand inline Collapse inline

Table 7. Statistical indicators of goodness of fit of the predicted pooled models for bunch weight (BuW) estimation of the two independent datasets: CS15 (‘Cabernet-Sauvignon’, 2015; n = 109) and VI14 (‘Viosinho’, 2014; n = 100). BuA = bunch area (cm2), #vBe = number of visible berries on the 2D image, MAE = mean absolute error (g), NMAE = normalised mean absolute error; RMSE = root mean square error (g), Pbias = percent bias and EF = modelling efficiency.

View popup Expand inline Collapse inline

Fitted model (Eq. nº)	Dataset	MAE (g)	NMAE	RMSE (g)	Pbias (%)	EF
BuW =1.077 * #vBe1.276 (15)	CS15	27.44	0.24	33.73	22.81	0.67
	VI14	32.78	0.19	44.90	-16.36	0.79
	combined	30.00	0.21	39.47	-0.04	0.79
BuW = 0.652 * BuA1.238 (20)	CS15	16.92	0.15	23.65	10.22	0.84
	VI14	39.94	0.23	52.70	14.97	0.71
	combined	27.94	0.19	40.26	12.99	0.78
BuW = 0.611 * BuA0.939 * #vBe0.358 (25)	CS15	17.95	0.16	25.45	14.01	0.81
	VI14	28.52	0.16	39.56	7.25	0.84
	combined	23.01	0.16	32.96	10.07	0.85

Discussion

Conclusions

To obtain accurate predictive models for automated vineyard yield estimation, this paper compared the explanatory power of the two most used image-based bunch attributes (number of visible berries and bunch area in 2D images) for predicting bunch weight. The models were fitted with data from four grapevine cultivars. The high goodness of fit of all the simple linear regression models indicates that either one of the image-based variables can be used as an accurate proxy of actual bunch weight. The goodness of fit showed by the models based on the pooled data indicates that a general model is also suitable. However, its application across different cultivars and seasons should be approached with caution, as indicated by some observed differences in the statistical measures of validation among the two independent datasets. Furthermore, the significant differences found when comparing the pooled models using cultivar as an additional factor indicate that cultivar-specific models can improve predictions.

The comparison of the explanatory power of the two image-based bunch attributes for predicting bunch weight showed that the models based on the predictor #vBe had a slightly lower R² than the models based on BuA, whether for each single cultivar (except for Viosinho) or for the pooled dataset. The combination of the two image-based explanatory variables in a multiple regression model produced predicted models with similar or noticeably higher R² than that obtained for single-predictor models. However, adding a second variable to the regression produced a higher and more generalised gain in accuracy for the simple regression models based on the predictor #vBe than for the models based on BuA. Despite the aforementioned differences and some contrasting results obtained in the validation tests of the three BuW pooled predictive models, based on our results it is possible to recommend using models based on the two image-based variables because they were generally more accurate and robust than the single variable models. When the gains in accuracy produced by adding a second image-based feature are small, a single predictor should be used. In such a case, our results indicate that BuA would be a more accurate and less cultivar dependent option than the #vBe. Further work is needed in order to test these models in field conditions at earlier phenological phases and with datasets from other cultivars and seasons.

Acknowledgements

This research has received funding from the European Community’s Seventh Framework Programme (SME 2013-2), grant agreement nº 605630, Project VINBOT. We would also like to thank the master students Filipe Silvério and Emilio Giuliani for their help in data collection and João Graça for his technical assistance.

References

• Aquino, A., Diago, M. P., Millán, B. & Tardáguila, J. (2017). A new methodology for estimating the grapevine-berry number per cluster using image analysis. Biosystems Engineering, 156, 80-95. https://doi.org/10.1016/j.biosystemseng.2016.12.011
• Aquino, A., Millan, B., Diago, M.P. & Tardaguila, J. (2018). Automated early yield prediction in vineyards from on-the-go image acquisition. Computers and Electronics in Agriculture, 144, 26–36. https://doi.org/10.1016/j.compag.2017.11.026
• Buayai, P., Saikaew, K.R. & Mao X. (2021). End-to-end automatic berry counting for table grape thinning. IEEE Access, 9, 4829-4842. https://doi 10.1109/ACCESS.2020.3048374
• Chen, X., Ding, H., Yuan, L.M, Cai, J.R., Chen, X. & Lin, Y. (2018). New approach of simultaneous, multi-perspective imaging for quantitative assessment of the compactness of grape bunches. Australian Journal of Grape and Wine Research, 24, 413-420. https://doi: 10.1111/ajgw.12349
• Clingeleffer, P., Dunn, G., Krstic, M. & Martin, S. (2001). Crop development, crop estimation and crop control to secure quality and production of major wine grape varieties: A national approach. Technical Report, Grape and Wine Research & Development Corporation, Australia.
• Cubero, S., Diago, M.P., Blasco, J., Tardaguila, J., Prats- Montabalbán, J.M., Ibáñez, J., Tello, J. & Aleixos, N. (2015). A new method for assessment of bunch compactness using automated image analysis. Australian Journal of Grape and Wine Research, 21, 101–109. https://doi: 10.1111/ajgw.12118
• de la Fuente, M., Linares, R., Baeza, P., Miranda, C., & Lissarrague, J. R. (2015). Comparison of different methods of grapevine yield prediction in the time window between fruitset and veraison. OENO One, 49(1), 27–35. https://doi.org/10.20870/oeno-one.2015.49.1.96
• Diago, M. P., Tardaguila, J. Aleixos, N. Millan, B. Prats-Montalban, J. M. Cubero, S. & Blasco, J. (2015). Assessment of cluster yield components by image analysis. Journal of the Science of Food and Agriculture, 95, 1274–1282. https://doi.org/10.1002/jsfa.6819
• Diago, M.P., Correa, C., Millán, B., Barreiro, P., Valero, C. & Tardaguila, J. (2012). Grapevine yield and leaf area estimation using supervised classification methodology on RGB images taken under field conditions. Sensors, 12, 16988-17006. https://doi:10.3390/s121216988
• Dunn, G. M. & Martin, S. R. (2004). Yield prediction from digital image analysis: A technique with potential for vineyard assessments prior to harvest. Australian Journal of Grape and Wine Research, 10, 196-198. https://doi.org/10.1111/j.1755-0238.2004.tb00022.x
• Font, D., Tresanchez, M., Martínez, D., Moreno, J., Clotet, E. & Palacín, J. (2015). Vineyard yield estimation based on the analysis of high resolution images obtained with artificial illumination at night. Sensors, 15, 8284–8301. https://doi:10.3390/s150408284
• Hacking, C., Poona, N., Manzan, N. & Poblete-Echeverría, C. (2019). Investigating 2-D and 3-D proximal remote sensing techniques for vineyard yield estimation. Sensors, 19(17), 3652. https://doi.org/10.3390/s19173652
• Hed, B., Ngugi, H.K. & Travis, J.W. (2009). Relationship between cluster compactness and bunch rot in Vignoles grapes. Plant Disease, 93, 1195-1201. https://doi.org/10.1094/PDIS-93-11-1195
• Herrero-Huerta, M., González-Aguilera, D., Rodriguez-Gonzalvez, P. & Hernández-López, D. (2015). Vineyard yield estimation by automatic 3D bunch modelling in field conditions. Computers and Electronics in Agriculture, 110, 17–26. https://doi.org/10.1016/j.compag.2014.10.003
• IVV (2011). Catálogo das castas para vinho cultivadas em Portugal. Instituto da Vinha e do Vinho, Lisboa, Portugal.
• Ivorra, E., Sánchez, A.J., Camarasa, J.G., Diago, M.P. & Tardaguila, J. (2015). Assessment of grape cluster yield components based on 3D descriptors using stereo vision. Food Control, 50, 273–282. https://doi.org/10.1016/j.foodcont.2014.09.004.
• Keene, O. N. (1995). The log transformation is special. Statistics in Medicine, 14(8), 811-819. https://doi:10.1002/sim.4780140810
• Liu, S., Zeng, X. & Whitty, M. (2020). A vision-based robust grape berry counting algorithm for fast calibration-free bunch weight estimation in the field. Computers and Electronics in Agriculture, 173, 1-10. https://doi.org/10.1016/j.compag.2020.105360.
• Loague, K. & Green, R. E. (1991). Statistical and Graphical methods for evaluating solute transport models: overview and application. Journal of Contaminant Hydrology, 7, 51-73. https://doi.org/10.1016/0169-7722(91)90038-3
• Lopes, C.M., Egipto R., Zarrouk O. & Chaves M. M. (2020). Carry-over effects on bud fertility makes early defoliation a risky crop-regulating practice in Mediterranean vineyards. Australian Journal of Grape and Wine Research, 26, 290–299. https://doi: 10.1111/ajgw.12437
• Martin, S., Dunstone, R., and Dunn, G. (2003). How to Forecast Wine Grape Deliveries Using Grape Forecaster. Department Primary industries, GWRDC, Adelaide, Australia.
• Mayer, D. G. & Butler, D. G.(1993). Statistical validation. Ecological Modelling, 68, 21-32. https://doi.org/10.1016/0304-3800(93)90105-2
• Nuske, S., Wilshusen, K., Achar, S., Yoder, L., & Singh, S. (2014). Automated visual yield estimation in vineyards. Journal of Field Robotics, 31(5), 837-860. https://doi.org/10.1002/rob.21541
• OIV (2007). OIV descriptor list for grape varieties and Vitis species. Organisation Internationale de la Vigne et du Vin, Paris, France.
• Pallioti, A., Gardi, T., Berrios, J., Civardi, S. & Poni, S. (2012). Early source limitation as a tool for yield control and wine quality improvement in a high-yielding red Vitis vinifera L. cultivar. Scientia Horticulturae, 145, 10-16. https://doi.org/10.1016/j.scienta.2012.07.019
• Pérez-Zavala, R., Torres-Torriti, M., Cheein, F. A., & Troni, G. (2018). A pattern recognition strategy for visual grape bunch detection in vineyards. Computers and Electronics in Agriculture, 151, 136–149. https://doi.org/10.1016/j.compag.2018.05.019
• R Core Team (2020). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/.
• Rey-Caramés, C., Diago, M.P., Martin, M.P., Lobo, A. & Tardaguila, J. (2015). Using RPAS Multi-Spectral imagery to characterise vigour, leaf development, yield components and berry composition variability within a vineyard. Remote sensing, 7, 14458-14481. https://doi 10.3390/rs71114458
• Rist, F., Herzog, K., Mack, J., Richter, R., Steinhage, V. & Töpfer, R. (2018). High-precision phenotyping of grape bunch architecture using fast 3D sensor and automation. Sensors, 18(3), 1-13. https://doi.org/10.3390/s18030763
• Rose, J.C., Kicherer, A., Wieland, M., Klingbeil, L., Töpfer, R. & Kuhlmann, H. (2016). Towards automated large-scale 3D phenotyping of vineyards under field conditions. Sensors, 16(12), 1-25. https://doi.org/10.3390/s16122136
• Sabbatini, P. & Howell, G.S. (2010). Effects of early defoliation on yield, fruit composition, and harvest season cluster rot complex of grapevines. HortScience 45, 1804–1808. https://doi.org/10.21273/HORTSCI.45.12.1804
• Schaeffer, D. L. (1980). A model evaluation methodology applicable to environmental assessment models. Ecological Modeling, 8, 275-295. https://doi.org/10.1016/0304-3800(80)90042-3
• Schöler, F. and Steinhage, V. (2015) Automated 3d reconstruction of grape cluster architecture from sensor data for efficient phenotyping. Computers and Electronics in Agriculture 114, 163–177. https://doi.org/10.1016/j.compag.2015.04.001
• Seng, K. P., Ang, L. M., Schmidtke, L. M., & Rogiers, S. Y. (2018). Computer vision and machine learning for viticulture technology. IEEE Access, 6, 67494–67510. https://doi.org/10.1109/ACCESS.2018.2875862
• Tello, J. & Ibáñez, J. (2014). Evaluation of indexes for the quantitative and objective estimation of grapevine bunch compactness. Vitis, 53(1), 9-16. https://doi.org/10.5073/vitis.2014.53.9-16
• Tello, J. & Ibáñez, J. (2018). What do we know about grapevine bunch compactness? A state-of-the-art review. Australian Journal of Grape and Wine Research, 24(1), 6-23. http://doi.org/10.1111/ajgw.12310
• Tello, J., Aguirrezábal, R., Hernaiz, S., Larreina, B., Montemayor, M. I., Vaquero, E. & Ibáñez, J. (2015). Multicultivar and multivariate study of the naturalvariation for grapevine bunch compactness. Australian Journal of Grape and Wine Research, 21(2), 277-289. doi:10.1111/ajgw.12121
• Tello, J., Cubero, S., Blasco, J., Tardaguila, J., Aleixos, N. & Ibáñez, J. (2016). Application of 2D and 3D image technologies to characterise morphological attributes of grapevine clusters. Journal of the Science of Food and Agriculture, 96(13), 4575-4583. http://doi.org/10.1002/jsfa.7675
• Victorino, G., Braga, R., Santos-Victor, J. & Lopes, C.M. (2020). Yield components detection and image-based indicatorsfor non-invasive grapevine yield prediction at different phenological phases. OENO One, 54(4), 833-848. http://doi.org/10.20870/oeno-one.2020.54.4.3616
• Whitty, M., Liu, S., Cossell, S., Jayakody, H., Woods, M., Tang, J., Singh, S., Oerle, P., Wiseham, D., Liu, S., Davidson, A., Stocco, T., Jarrett, J., Jarrett, P., Wotton, C., Shepherd, J., Lim, S., Petrie, P. & Dunn, G. (2017). Improved yield estimation for the Australian wine industry. Final report project number DPI 1401, Wine Australia, Adelaide, South Australia.
• Xin, B., Liu, S. & Whitty, M. (2020). Three-dimensional reconstruction ofVitis vinifera (L.) cvs Pinot Noir and Merlot grape bunch frameworks using a restricted reconstruction grammar based on the stochastic L-system. Australian Journal of grape and Wine Research, 26(3), 207-219. http://doi.org/10.1111/ajgw.12444
• Zabawa, L., Kicherer, A., Klingbeil, L., Töpfer, R., Kuhlmann, H. & Roscherb, R. (2020). Counting of grapevine berries in images via semantic segmentation using convolutional neural networks. ISPRS Journal of Photogrammetry and Remote Sensing, 164, 73-83. https://doi.org/10.1016/j.isprsjprs.2020.04.002