Application of ImageJ Software in the assessment of flowering Intensity and growth Vigor of Pear Trees

The research was carried out in the Experimental Orchard of the National Institute of Horticultural Research in Skierniewice, Poland (latitude 51°57′ N, longitude 20°08′ E, altitude 120 m). The study was performed on young ‘Conference’ pear (Pyrus communis L.) trees grafted on Quince C (Cydonia oblonga Mill.). The experimental trees were planted in 2013 at a spacing of 3.5 m × 1.5 m. The trees were trained as a spindle to a height of 2.5–3 m. Lateral branches were bent with strings and fastened to the trunk to reduce tree vigor and induce flower bud initiation. To keep the crown size within the allotted space, the trees were pruned annually in March (at the stage of swollen buds/bud burst). Only strongly growing shoots on the inside of the crown were removed.

For assessing flowering intensity, manual counting of flower clusters and taking of photographs of the trees were conducted at full bloom in 2015. In 2016, photographs were taken 3 days before full bloom and also during the full bloom of pear trees. The number of flower clusters on individual trees was manually counted on the day of image acquisition. The calibration model for assessing flowering intensity by image analysis was based on measurements and photographs taken for nine selected trees differing significantly in the number of flower clusters. Quality assessment of the model was carried out on 26 nonselected trees. For the assessment, the flowering trees were photographed against a black background, always from the same distance. In order to scale the size of objects in the photograph, a reference marker of known dimensions was placed on the background surface.

Tree growth vigor was estimated by manually measuring the total length of the central leader and shoots of individual trees. The calibration model for assessing the vigor of trees by image analysis was based on manual measurements and photographs taken for nine selected trees differing in the total length of shoots. Quality assessment of the model was carried out on 26 nonselected trees. Photographs of trees were taken after all the leaves had fallen off. Due to the large size of trees in 2016, the assessment of their size was carried out only in 2015.

For the assessment of total shoot length, trees were photographed from the same distance. In order to scale the size of objects in the photograph, a reference marker with known dimensions was placed near the tree. The photographic images were recorded with a Nikon CoolPix P520 digital camera equipped with a 4.3–180 mm objective lens.

The image analysis procedures are schematically shown for the assessment of flowering intensity in Figure 1, and for tree growth in Figure 2. All image processing was performed offline using ImageJ software. Images were first loaded from the camera's memory card to a computer. Then, thanks to the marker placed on the background (for the photographs taken during flowering) or near the tree (photographs taken in the winter to assess the total length of tree shoots), the scale for a specific photograph was set.

Figure 1

Figure 2

In the next step, only the area where the flowers were visible was duplicated from the entire photograph. To estimate the number of flower clusters on the trees, it was decided to use the contrasting white color of pear perianth petals. Areas with the white color were separated using the command “Color Threshold” (Fig. 3). The total surface of the separated area was then measured (in cm²) by using the command “Measure”.

Figure 3

After processing the photographs, a calibration regression model was built to relate flower surface area (cm²) to the number of flower clusters. The calibration model determined in this way was used to assess the intensity of flowering of subsequent trees.

After setting the scale in the photograph, the image of the entire tree was copied. The next step was to eliminate the background by using “Subtract Background” and “Color Threshold” commands, and generate a tree skeleton using “Make Binary” and “Skeletonize Process” (Fig. 4). The total length of branches (in cm) of the individual tree was measured by using “Measure_Skeleton_Length_Tool” plugins.

Figure 4

Finally, the manual and automated measurements were compared. Quality assessment of the developed models was carried out by determining the mean absolute error (MAE) and the mean absolute percentage error (MAPE) of estimation.

MAE is the average of all absolute errors.

MAPE is a statistical measure of how accurate a forecast system is. (2) MAPE=100%n∑i=1n| yi−xi | MAPE = {{100\% } \over n}\sum\nolimits_{i = 1}^n {\left| {{y_i} - {x_i}} \right|}

n – the number of errors

Based on the description of the correlation between the total surface area of perianth petals (white color) determined by analyzing photographic images of pear trees and the actual (counted) number of inflorescences, a calibration model was developed in 2015 that was used for estimating the flowering intensity of pear trees (Fig. 5). This calibration model was described by a second-degree polynomial at a 98% level of determination.

Figure 5

In the next step, the quality of estimating the flowering intensity of pear trees was assessed using the previously determined calibration model. Analysis of the results obtained showed that the estimates were strongly correlated (determination coefficient of 83%) with those determined by image analysis (Fig. 6).

Figure 6

Preliminary analyses in 2016 (unpublished data) showed that the use of the calibration model developed in the previous season for assessing flowering was burdened with a large error. These observations indicate the need for individual calibration of estimation models in each season of assessment. This is a characteristic feature of this type of method. Calibrations should be performed (or at least tested and modified if necessary) independently for species (even cultivars) and growing seasons, to be able to estimate the desired parameters accurately (Adamsen et al. 2000; Treder et al. 2016).

The number of flowers is estimated based on the surface area of perianth petals visible in the photograph. This area depends not only on the number of flowers but also on the extent of their development, the annually changing architecture of the tree crown (density of shoots), and the extent of development of the leaves that partially cover the flowers. Detailed research conducted in 2016 showed that due to the changing surface area of perianth petals during the flowering period, calibration should be performed individually for each assessment date. Figure 7 presents two calibrations performed with the same trees three days apart, on April 22 and 25, 2016. On both dates, very high regression coefficients were obtained between the surface area of petals measured on the photographs and the number of inflorescences counted, r² = 0.98 and 0.97, respectively. The presented linear calibration models, however, differ from each other by the values of the coefficients a and b. In the case of the calibration performed on the first date, 1 cm² of petals measured on the digital image corresponded to approximately 0.21 of an inflorescence, and after 5 days, when the flowers had developed more, it was 0.17 of an inflorescence. A lower absolute error of estimation was obtained on the first calibration date when the flowers were just before full bloom (Table 1). The percentage values of the MAPE, depending on the year and time of measurement, ranged from 21.8% in 2015 to 14.0% on the first assessment date in 2016.

Figure 7

Techniques for remote detection of flowering have been demonstrated in several crops. For example, Viña et al. (2004) demonstrated the use of a spectral index (computed from several light bands) to detect the emergence of tassels in maize. Attempts to use the image processing approach for remote monitoring of flowering patterns in lesquerella canopies were made by Adamsen et al. (2000) and Thorp and Dierig (2011). Guo et al. (2015) described a method for detecting flowering panicles of paddy rice in RGB images taken under natural field conditions. According to the authors, the proposed method can quantify the daily amount and the diurnal changes of flowering, and even identify daily peaks of flowering.

Irregular flowering and fruiting is a common trait in some important horticultural plants such as apple, pear, or plum (Monselise & Goldschmidt 1982). The annual variation in the bloom density of individual trees can cause a large variation in orchard yield (Bukovac et al. 2010). For that reason, attempts have been made to automate the process of estimating the number of flowers (clusters) in apple orchards. Aggelopoulou et al. (2011) analyzed images of blooming apple trees. An image processing-based algorithm was developed that predicted the fruit yield of an apple tree by analyzing a picture of the tree at full bloom. The algorithm showed an 18% error in the predicted yield, so optimization of it (as well as of the image-taking process) was, as stated by the authors, a necessary further step. Another attempt was made by Hočevar et al. (2014), who estimated the number of flower clusters of individual trees in a high-density apple orchard on the basis of image analysis with HSL (hue, saturation, and luminance) thresholding as an element of site-specific management (spraying) decisions. These results indicate that such methods have the potential to make possible intensive sampling of flowering in a variety of crop plants, and fruit yield might be predicted by analyzing flowering distribution maps.

The assessment of tree growth vigor by means of image analysis was performed after leaf fall only in the autumn of 2015. In the autumn of 2016, the trees were so overgrown that some shoots overlapped the neighboring trees, which in the case of digital analysis prevented them from being assigned to a specific tree.

To be able to estimate the growth vigor of trees by means of image analysis of digital photographs of them, a calibration model was determined (Fig. 8). It described the correlation between the actual length of shoots measured individually for each tree and the values obtained by analyzing the photographs.

Figure 8

In the next step, the quality of estimating the growth vigor of pear trees using image analysis was assessed. The estimates were strongly correlated (determination coefficient of 93%) with those determined by image analysis (Fig. 9). The MAE of estimation was 93.1 cm, and the MAPE was 12.9%.

Figure 9

In recent years, more and more attention has been paid to the measurement of plant morphological parameters using the image and video acquisition and processing systems. Much effort has been focused on nondestructive determination of leaf surface area, though there have been some attempts to use the image-processing approach to analyze plant shoot length. Polder et al. (2010) adopted the ImageJ application for measuring total shoot length from photographs of several aquatic plant species. High correlation (r² > 91) with manual measurements proved the usefulness of the method of digital imagery and the developed analytical procedure. Brunes et al. (2016) used image processing (and the Matlab^® mathematical tool) for determining the length of shoots of wheat seedlings. The proposed method was used by the authors for separating the cultivars into different levels of vigor. Wang et al. (2011) proposed a method for modeling tea shoots with images and calculation models for nondestructive monitoring of the growth of tea shoots. A more complex system based on an image processing technique (images captured using multiple cameras) for monitoring the growth and nondestructively predicting the age and fresh weight of mustard plants was proposed by Saputra et al. (2017).

Less work has been done on fruit tree species (Lordan et al. 2015). The main reason for estimation uncertainty while using image acquisition and processing in fruit orchards is the overlapping of branches of neighboring trees (Hočevar et al. 2014) and the dependence of the applied method on illumination, which was also experienced in our study. Because the work described here was our first attempt to use digital image analysis for estimating pear tree vigor and flowering, more work is necessary to improve the image taking and processing procedure as well as to optimize the developed algorithms.