Hero Image

Berry size distribution

Estimating berry size distribution using image analysis
 
Mark Battany, UC Cooperative Extension


Previously, a method for counting berries using image analysis was described at the following website:

http://cesanluisobispo.ucanr.edu/Viticulture/Berry_counting.htm

If you haven't read the above website, please do so first as it will make the information below easier to understand.

The same image analysis method can also be used for estimating relative berry size distributions. In the example below, I detail the steps that I followed to test how the berry size distribution based on individual berry weights compared to the berry size distribution based on estimated berry volumes from image analysis. Note that if berry density is constant, then normalized values of berry weights and volumes should in theory be identical; hence the expectation is that both methods will produce identical relative size distributions.

Step 1. All berries from a single Pinot Noir cluster were removed, and each of the 238 berries was individually weighed. Each berry weight was then divided by the weight of the heaviest berry, to produce a list of normalized berry weights; in this case, the heaviest berry weighed 1.01 g, so the normalized values are virtually the same as the actual berry weights in grams. Note that this particular cluster had a large degree of 'hens and chicks', with very many chicks.

Copy of DSC_2538 - hens and chicks

Step 2. The same berries were then scanned following the method outlined on the berry counting website, with the only difference that in the very last step, when using the ANALYZE PARTICLES command, that an additional check was placed in the box marked DISPLAY RESULTS. By enabling this feature, the ImageJ software now produced a separate window listing the area of each individual particle (or berry) in the image; this "Results" file was then saved in the default Excel format.

Copy of DSC_2545 - berries on scanner

Step 3. Each individual berry area from the "Results" files was then converted to an equivalent berry volume, using standard geometric equations, assuming that the berries were spherical. Each equivalent berry volume (in cubic pixels) was then normalized, by dividing each value by the largest volume in the list.

Step 4. Each set of normalized berry sizes was then separated into five different size categories (0-0.2, 0.2-0.4, 0.4-0.6, 0.6-0.8, and 0.8-1.0), and the number of berries in each category were counted to make histograms of berry size distribution. The two histograms for the berry sample, measured with the two different methods, are shown on the same chart below.

Histograms

Step 5. The numbers of berries in each histogram size category were compared with a paired t-test. The two-tailed P value was > 0.9999, indicating no significant difference between the histogram category values.

Thus, both methods produced essentially the same results in this test, which demonstrates the potential that image analysis has for replacing hand weighing to estimate berry size distributions. The big difference in methods was the time required; weighing each individual berry by hand was very time consuming and tedious, whereas the image analysis method was very fast. If one is already counting berries using the image analysis method, the additional effort to estimate the size distribution is minimal.