Digital solutions for grape quality measures at the weighbridge

Abstract

This project applied rapid digital methods to assess grape quality on delivery to the winery, in particular Botrytis infection and presence of matter other than grapes (MOG). The primary method used was hyperspectral imaging, initially laboratory-based with a full spectral linescan camera (400-1000 nm wavelength), using both laboratory-infected Botrytis samples and field samples. Eight wavelengths were then selected for use with a filter-based multispectral snapshot camera, which was more applicable for use at the winery weighbridge and in the vineyard. Spectral imaging identified Botrytis infection and MOG, differentiated types of fungal infection and identified grape sunburn and shrivel.

Summary

The primary objective of this project was to develop digital methods for rapid assessment of grape quality on delivery to the winery. The key parameters were the identification of Botrytis-infected grapes and assessment of the amount of matter other than grapes (MOG) in mechanically harvested loads. A secondary objective was to determine whether these methods could also be applied in vineyards for non-destructive Botrytis assessment of grapes on the vine.

Images were taken with a laboratory-based hyperspectral linescan camera (Specim FX10) that collected spectral information over the 400-1000 nm wavelength range for individual pixels. The spectral information was used to classify each pixel with regard to what material class it belonged to, and this classification was applied as a false colour image to highlight spatial information. The image analysis software packages used were Scyven and ENVI.

Hyperspectral imaging (HSI) could discriminate clean and infected red and white grapes, for both laboratory infections and field samples. The calibrations were not dependent on growing region, grape variety and Botrytis strain. HSI could also discriminate Botrytis infection from sour rot (mixed fungal and bacterial infection of grapes). Washing sporulating berries with grape juice or smashing berries, to simulate mechanical harvesting, did not prevent identification of Botrytis. HSI could also identify sunburn in white grapes and shrivel in red and white grapes. HSI could discriminate other grapevine components that often form MOG in mechanically harvested grape loads (i.e. canes, wood, petioles, leaves and insects).

Key wavelengths that discriminated different materials were used to select filters for a multispectral camera (Ocean Optics SpectroCam) that operated in a snapshot mode, similar to a normal still camera, and did not require the sample to be moved under the camera as with the FX10 linescan camera. This made it more suitable for use in the field, at the winery weighbridge or in the vineyard. Laboratory tests were performed to compare LED and halogen lighting. Images collected under LED lighting indicated insufficient output in NIR wavelengths (790, 860, 918, 972 nm) and only the first four filters (444, 531, 680, 717 nm) could be used for multispectral imaging (MSI). Halogen lighting was suitable over the whole wavelength range.

When samples of grapes and grapevine material were imaged under LED lighting, shadows on the background matrix were classified as infected grapes and separate spectral libraries had to be made for the shadows to correct this. Under halogen lighting, lighter shadows could be included in the general background spectral libraries to provide correct classification but darker shadows remained an issue.

The SpectroCam was tested in a winery environment at a grape weighbridge testing station. Grape loads arrived in very large trucks that were unsuitable for direct imaging, but the testing station used a load sampling system that could deliver samples to a tray for analysis. Samples were imaged as a mass on the sample tray, in small tubs and spread out on a polymer board that was also scanned by the FX10 camera. Most of the intake was at night and samples were imaged with either halogen or LED lighting, but other lighting conditions were also tested: full sunlight with halogen fill-in and the shaded testing tray area with halogen plus some ambient daylight.

When samples were spread out on a board, MOG, clean grapes and infected grapes could be discriminated, but with LED lighting required spectral libraries for shadows on the backboard. Specular reflections were a problem with material from mechanically harvested loads that was wet with grape juice. Some incorrect classification occurred at the edges of objects, where there may have been dark shadows and mixed pixels. Spectral differences in these regions were identified using the full spectral scan from the FX10 camera. When bulk samples were imaged either on the testing station sample tray or in sample tubs, shadows within the samples caused incorrect classification. This could be compensated for to a degree, using spectral libraries to classify shadows, but unlike shadows on a sample board the shadows on a bulk sample contained a variety of backgrounds, so were difficult to define.

The multispectral camera was also tested in the vineyard to image grapes directly on the vine. Multispectral imaging could discriminate infected grapes, clean grapes and grapevine tissue but the system was not practical as the camera filter wheel needed a stable voltage to synchronise well, so a generator was needed. Each wavelength requires manual exposure control, making shifting ambient light from the sun a problem. Sample movement within the canopy was also an issue as the camera takes separate images of the same scene for each filter.

The RotBot app was also tested in the vineyard. RotBot can discriminate clean and Botrytis infected white grapes but must be used with a blue background. For correct infection level classification there must only be grapes visible in the image: bunch rachis, petioles, immature canes are classified as clean grapes; lignified canes, grapevine bark, dried petioles, dried leaves, sunburnt or shrivelled grapes and powdery mildew berry scars are classified as infected grapes. RotBot can handle shadows on samples but overexposed regions are incorrectly classified.

In conclusion, multispectral imaging is capable of discriminating material in mechanically harvested grape loads. However, to avoid problems caused by shadows within the sample, imaging would be best done with samples spread out on a conveyor belt with a linescan camera or spread on a static backing board with a snapshot camera. This provides a uniform background with reduced shadow, and any shadows would be on a constant background. Some issues with image analysis need to be addressed, with mixed pixels causing incorrect material classification on the edge of samples. Specular reflectance on wet samples is an issue but could be mitigated with diffuse lighting. Further work should be done with LED lighting systems suitable for use in the NIR range as a replacement for halogen, which may soon no longer be available in Australia.

An important area of future work to encourage adoption would be to create a robust imaging system where software controls the camera and seamlessly runs the image analysis in the background and reports a result, with minimal user input. A vineyard imaging system requires development of portable devices with appropriate wavelength ranges and with onboard lighting and image analysis capability.

Abbreviations and glossary

MOG – matter other than grapes

SVM – support vector machine

Vis-NIR – visible-near infrared

MIR – mid-infrared

RGB – red green blue

HSI – hyperspectral imaging

MSI – multispectral imaging