Improved yield prediction for the Australian wine industry

Abstract

This project presented a system for yield estimation using image processing using consumer cameras and mobile phones to tackle high errors in yield estimation pervasive across the industry. The system was tested on four vineyard blocks and shown to outperform manual yield estimation at all stages, achieving 5.5% error at flowering. Image processing-based shoot counting, flower counting and berry counting techniques showed high levels of accuracy and have great potential to reduce the manual labour required in Australian vineyards. However, to achieve this, improvements in automated data collection and management are crucial to drive the accuracy and uptake forward.

Summary

Industry practices for yield estimation in winegrape vines have been developed over many years, however they are still error-prone, time consuming and hence costly. Time pressures frequently prohibit detailed analysis of the data collected over many seasons, so understanding of the effect of different components on the final yield estimates is lacking.

This project developed several image processing-based systems for the purpose of yield estimation on an individual block basis. The systems were evaluated against manual yield estimates, using tonnage of fruit delivered to the winery as the objective. Analysis of non-bearing proportions of the block and harvester efficiency were included to provide inputs to all the yield estimates.

Experiments were conducted on two common varieties, Chardonnay and Shiraz, at two sites, Orange (NSW) and Clare (SA) over three growing seasons from 2015 - 2017. The blocks chosen were all spur pruned, with VSP trellis in Orange and ‘Aussie sprawl’ in Clare.

The final novel field-mobile system developed in the third year of the project was able to outperform the industry standard manual yield estimation method, relying predominantly on data available within the three project seasons. The system involves the use of GoPro cameras for videoing an entire block at the shoot stage in combination with mobile phone photos of marked bunches at flowering, pea-sized and harvest.

Shoot counting by image processing from consumer grade cameras mounted on farm vehicles such as Gators and quadbikes was demonstrated to have an accuracy of 88%. This translated to an error of 20% in final yield estimates, five months prior to harvest. The system can potentially be mounted on a tractor during an early spray pass, and the optimum time for filming was found to be EL9.

At the flowering stage, an average error of 5.5% was achieved across the four blocks, which is close to the winemaker target of 5%. The error increased to 14% and 12% respectively at pea-sized and harvest, as the approach is fundamentally based on counting florettes or berries per shoot. Variation in shoot number throughout the season influences the final result, hence later measurement of shoot or bunch density would further improve accuracy of yield estimation.

Of the many image processing components that went into the yield estimation, an improved flower counting system suitable for Australian varieties shows the most promise as a standalone output. An accuracy of 84% was achieved across 12 different datasets, and the limitations of the algorithm in respect of development stage and variety are discussed in the report.

Berry counting at pea-sized showed an accuracy of 90%, speeding up data collection and greatly reducing the manual labour needed at pea-sized estimates. Berry counting at harvest using an existing 3D reconstruction algorithm achieved 88% accuracy on a single image without the need for any calibration and is applicable to multiple varieties. Across a few dozen images, the count error reduced to less than 10% and in two blocks was less than 2%, showing the feasibility of automated berry counting from mobile phone images.

Berry diameter measurement was briefly investigated and shown to have an accuracy of 95% from single images of bunches.

The project showed that achieving less than a 15% estimation error from industry standard manual sampling even at harvest is not realistic when relying on a purely data driven approach. In fact, using long term averages is often more accurate than manual estimates, which calls into question the effort put into making manual estimates. The addition of subjective farm-specific knowledge and longer term historical records can improve manual estimates in some cases, but all manual measurements are subject to bias.

Stratified sampling was shown to substantially reduce the number of samples required, and is a first step in improving accuracy without image processing. Bunch to shoot ratios and bunch to inflorescence ratios were found to contribute the most uncertainty to early season forecasts. Counts of berries per bunch were the cause of the greatest inaccuracy in forecasts at pea-sized, suggesting that methods for reducing bias in bunch selection are necessary. Estimated berry diameter at harvest was also poorly predicted, suggesting image processing methods for tracking this over time would be beneficial.

Improving data custodianship or management practices is critical to driving the accuracy of yield prediction up across the industry. Smartphone apps for flower counting and berry counting using image processing have the potential to not only improve accuracy by speeding up data capture but also provide a front-end for entering these data into a cloud-based yield estimation system.

Given the limited amount of data available to determine prediction factors, it is recommended that longer term trials be undertaken. Greater potential for improved accuracy exists in cane pruned vines due to greater visibility of shoots and fruit and industry is encouraged to implement cane pruning systems.

Finally, the report recommends further development and commercialisation of a number of image processing technologies that will improve yield estimation and provide key tools to reduce labour requirements. Shoot counting and early season mapping have the potential to greatly assist growers to manage blocks differentially and hence improve yield and quality across the Australian wine industry.