New digital technologies for managing winegrape production and quality

Abstract

Four areas of viticultural management were addressed through this project, which further developed technologies from predecessor projects CSA 1601 and CSA 1602.

• The viability of using machine vision for hands-off yield estimation at scale in a range of viticultural situations was clearly demonstrated with an accuracy >90%, but will require significant computational efficiencies to be commercially practicable. In contrast, utilising a radar approach to assess yield by ‘seeing’ through the canopy was not successful, despite examining a wide range of frequencies and methods.
• The ‘ProxiCrop’ prototype low-cost static sensor for irrigation management support was refined and tested, and proved to be capable of monitoring canopy temperature and subsequent estimation of vine water status.
• On-the-go field estimation and mapping of fruit maturation and composition was demonstrated using an NIR spectral sensor and calibration libraries that were extended and further validated.
• The use of dual-axis spinning LiDAR with CSIRO SLAM and ‘raycloudtools’ libraries to map canopy structure in three-dimensions was finalised and demonstrated in a trial with commercial growers.

Summary

CSIRO undertook two projects with Wine Australia as part of the Rural R&D for Profit scheme, which ended in 2020. CSA 1601 developed ‘New technologies for dynamic canopy and disease management’ and CSA 1602 developed ‘New non-destructive technologies for simultaneous yield, crop condition and quality estimation’. The work described in this report continued the testing and development of the technologies utilised in those projects. This encompassed five potential solutions to four industry problems, all five aimed at in-field use with four being on-the-go.

Two technologies addressed the need for accurate yield estimation. This is a key part of horticultural production, including winegrapes, with supply chains and fruit value significantly impacted when production volume does not match expectations. Current yield estimates rely significantly on the experience of the vineyard manager, are labour intensive and typically have an error rate of up to 30%. Yield can be split into three components, berry weight, berry number per bunch and bunch number per vine/vineyard area. Berry weight is only fixed at harvest, whereas berry number is determined by fruit-set early in the season and bunch number by budburst, at the start of the season.

Machine learning methods trialled in CSA 1602 for assessing yield potential by counting inflorescences in video data post-budburst were further refined, proving highly accurate (approximately 5% error) without any manual adjustment or additional calibration. Methods for estimating berry number and size were explored and data sets collected, but the capability available for pre-harvest yield estimation was primarily directed at estimating bunch mass from video in combination with bunch counts. This was undertaken in collaboration with the CSIRO Machine Learning and Artificial Intelligence Future Science Platform and produced a complete data pipeline demonstration for the determination of fruit yield from video data, again without manual adjustment or additional data collection.

Radar has the potential to ‘see through’ canopy to detect bunches that are occluded by leaves based on wood, fruit and leaves having a different radar cross section. Initial results from testing in CSA 1602 demonstrated differences in the radar returns depending on whether fruit were present in a canopy and between vine components. However, detailed testing in the current project of both transmission and reflection modes across a wide range of frequencies did not suggest that it is possible to obtain a quantitative relationship with fruit mass due to the scatter of signal within the canopy.

The second industry problem addressed was in-field determination of fruit maturity and quality. More than any other perennial crop, the commercial potential of winegrapes depends on composition. If objective assessment of fruit quality can be done in the field prior to harvest and transport, harvest timing and winery planning can be optimised and a number of potential points of friction between grower and winery avoided.

As part of CSA 1602, a lab based spectral classification tool was developed to allow grapevine parts (fruit, leaves, canes, other) to be identified using only their reflectance spectra. In addition, a calibration for prediction of key fruit composition parameters from their reflectance spectra was developed using four winegrape varieties. The current project expanded the calibration out to eight varieties, which allowed model validation by successfully predicting parameters for varieties excluded from the model. The field spectrometer-based instrument was further developed, and fitted with a lighting system that allowed it to be used in full daylight. Finally, a data processing method was trialled to allow the lab-based calibration to be used with the field instrument. The combination of these tools successfully predicted fruit sugar maturity when used on a continuously moving farm vehicle.

The third problem addressed was objective measures of canopy structure to allow consistent and repeatable canopy management. As part of CSA 1601, a dual-axis spinning LiDAR was successfully used with CSIRO’s Simultaneous Localisation and Mapping (SLAM) tools and a novel technique using the ‘ray-cloud’ rather than just returned laser pulses to map canopy density in 3D. Through the current project this technique was packaged into an open-source library and released by CSIRO. In addition, a demonstration event was held in the Coonawarra, where two vineyards were scanned with the LiDAR system, the data processed overnight, and canopy density maps provided to growers at an information session the following morning.

Finally, in recent years there has been significant interest in using plant-based sensing in addition to soil moisture measurements for irrigation management. A component of Wine Australia-CSIRO project CSA 1701-1.5 developed such a sensor, with the aim of quantifying vine water status, based on canopy temperature. In contrast to existing solutions, which have a single point measurement of a few leaves or less, the sensor used a low cost thermal imaging sensor, fused with RGB imaging, to obtain canopy temperatures of large areas (approximately one panel of vines). This work was moved into the current project, with the sensor tested under commercial field conditions and developed accordingly. The resulting canopy temperature data were well correlated with direct measurements of vine water stress.

In conclusion, four of the five technologies further developed through this project show potential for use in commercial growing systems, although the technology readiness of these varies.