Brix Prediction Using Hyperspectral Imaging
USING THE MV.C NIR® WITH THE PERCLASS MIRA® SCANNING STAGE
Blueberries are native to North America and were first cultivated for sale in 1916 (USHBC, 2014). The United States is the world’s largest producer of blueberries (FAOSTAT 2016). During 2016, just in the United States alone, a total of 690 million pounds (312,978 metric tons) of cultivated and wild blueberries were produced and utilized (NASS 2020).1
Outside the US, Peru is the largest producing country, followed by Chile, Mexico and Argentina. In Europe, Spain produced 48,520 metric tons (106,968,376 pounds) and Germany produced 11,300 metric tons (24,912,255 pounds)in 2020.2
They are a diverse species with different varieties having wildly different flavor profiles. Often, earlier blooming varieties are sweeter, and good for snacking, while later season varieties are more tart, and preferred for baking.
Traditionally, blueberries are inspected manually, looking for defects, unripe berries, or bruising. The sweetness of the berries is oftentimes measured via destructive methods. With Headwall’s hyperspectral imaging (HSI) technology, blueberries can be inspected based on their condition, and their Brix value can be predicted, allowing for a non-destructive, rapid measurement of large quantities of berries, with results in real time. Hyperspectral imaging allows growers to assign value based on rapid, accurate grading.
VNIR Sorting of Good and Bad Blueberries
Growers assessed the blueberries for ripeness, bruising, and defects, categorizing them as either good quality or low quality. They were scanned by Headwall using the perClass Mira Stage and the MV.C VNIR, and then a classification model was created by training perClass Mira Software to recognize the differences between the grower-sorted blueberries.
The model identifies blueberries and categorizes them as good or bad. As shown in Figure 1 (below), the model correctly identified 24/25 good berries and 23/24 bad berries. This suggests that VNIR hyperspectral imaging can be used to sort good and bad blueberries to a high level of accuracy.
NIR Blueberry Brix Regression
A subsequent study assessed the ability of a Headwall MV.C NIR sensor with perClass Mira Stage and Software to predict the sweetness (Brix) of blueberries. Both early and late season blueberries were measured to test the model quality across an entire growing season.
First, hyperspectral images of the berries were acquired, then the berries were juiced, and a Cole Parmer E81150-48 handheld digital refractometer measured the Brix value for each individual blueberry. The berries were annotated with the measured Brix values, and a regression model was built. Repeat scans of the same berries allowed for an expanded test set and confirmed the strength of the model. The plot of predicted versus measured Brix in Figure 3 combines the results from the early and late season blueberries.
The classification model first classifies each pixel in the scene as either a good berry, bad berry, or background. Then, contiguous regions of at least 1000 berry pixels are segmented from the background as “berry” objects. These objects are then graded based on the percentage of good or bad berry pixels in that object. This classification model had a 92% success rate correctly classifying good berries as good, and bad berries as bad.
Headwall’s hyperspectral imagers paired with the perClass Mira Stage and Software allow users to rapidly build and update classification and regression models. The perClass Mira models extract actionable information in real-time. These can be used on offline quality test stands, inline at packing/processing plants, or in a lab. This allows producers to better grade blueberries and assign value based on product quality.
1 Blueberries; AgMRC, National Resource for Value-Added Agriculture, January 2023; https://www.agmrc.org/commodities-products/fruits/blueberries