Increasing the efficiency of forecasting winegrape yield by using information on spatial variability to select sample sites Andrew Hall, Research Fellow, Spatial Science Leo Quirk, Viticulture Extension Officer Mark Wilson, Technical Officer Jim Hardie, Professor, Winegrowing Innovation NWGIC, Wagga Wagga Introduction Grower estimates of seasonal winegrape production can differ from winery deliveries by an average of 33%. The impact of this mismatch is high, not only for wineries, which spend considerable amounts of money on harvest preparation and logistics, but also for suppliers, who need to match inventory to demand, and for the seasonal negotiation of grape prices, for which accurate supply forecasts are needed. The collective cost of this uncertainty in grape supply volume to Australian wineries has been estimated at $200 million a year. Greater levels of accuracy in forecasting seasonal grapevine yield would obviously benefit wine producers and other sectors of the industry. An important cause of inaccuracy in yield forecasting is the spatial variability within vineyard blocks. The underlying cause of spatial variability in grapevine performance is often the inherently variable nature of soils particularly in terms of depth and water-holding capacity but it may also be topography or management practices, such as spatial variations in irrigation dripper rates. One yield-monitoring study of several regions of Australia found variation of up to 10 fold within Cabernet Sauvignon vineyards. For this reason, extensive sampling is needed to produce data that are representative of this variability within vineyards. Estimates of yield in vineyards presently rely on random or group sampling of vines and their developing grape bunches between budburst and harvest. The average of the factors used to calculate yield namely, bunch number and bunch weight per vine in the sample set is then multiplied by the number of vines in the block to produce the forecast for the whole block. Increasing the numbers of samples taken can improve the difference between the expected and actual yields, but this requires increased cost and effort. These impediments can be overcome by reducing the number of sampled vines by segregating the vine population into zones of known variation using precision technologies such as electromagnetic (EM) surveying of soil conditions, aerial remote sensing of vine canopies, or harvest yield monitoring. The yield forecast study We now have some preliminary results from a study that is investigating the benefits of using spatial information to reduce the number of samples needed to describe the
variability of yield within a vineyard block and hence reduce the sampling time taken to obtain accurate estimates of yield. The study uses two trial sites, one in the Hunter Valley and the other in the Riverina, both comprising mature (more than 10 years old) Shiraz grapevines. On both blocks, 54 sample sites are hand-harvested within a week before the blocks are commercially harvested. The total fruit weight at each site is recorded. The variability of the Hunter Valley site was initially mapped by an EM-38 soil survey conducted while the soil profile was relatively moist (Figure 1). Geographic information system software was then used to split the block into three zones of equal variability. Each of these three zones was further divided into six sub-zones. Three grapevines were selected from within each of the sub-zones to give 54 sample sites. Figure 1 (a) EM-38 soil survey of Hunter Valley vineyard site. (b) Three-zone map made from the EM-38 data. (c) Eighteen-subzone map, showing the 54 sample sites selected (three within each subzone). The Riverina site has 54 sample sites that are categorised in a similar way, but the sample sites change because they are selected each season from canopy vigour data derived from remote-sensing images taken of the site (Figure 2). Each grapevine is individually described using a computer program and ranked by canopy vigour. Sample grapevines are then selected at intervals on the ranked list of grapevines to best characterise the overall variability of the block. Each grapevine in the Riverina trial is placed into categories that are equivalent to the Hunter Valley site s zone and subzone designations.
Figure 2 The Riverina site. This close-up of the remotely sensed image of the vineyard contains computer-generated vine-location information. NVDI, normalised difference vegetation index A computer simulation is then used to calculate the potential error of making yield estimates with fewer sample sites. The simulation selects different sub-sets of samples incorporating rules that take into account the variability of the site. The number of randomly selected samples required to achieve the equivalent potential error were then calculated using an established method. Results of the study so far In the first year of the trial, the yield forecasts that we obtained by multiplying the yields from the hand-harvested samples by the total block row length were far greater than the machine-harvested totals, so it was clear that an adjustment had to be made for the difference between hand- and machine-harvesting. In the second year, when a multiplier factor was used to take this difference into account, we obtained more accurate forecasts for both blocks: 6.4% and 4.6% greater than the actual yields for the Riverina and Hunter Valley sites, respectively. Using all 54 sample vines produced a similar level of accuracy at both sites, which is an encouraging result to date. However, more seasons of data collection are required before we can be confident that the method of selecting the 54 sample vines and the annual modification multiplier are reliable. To achieve the same level of potential error in an estimate, the number of random samples required is consistently greater than the equivalent number of samples selected by using spatial information. For example, the four instances of yield forecasts using nine sample sites that were selected by incorporating spatial knowledge produced potential errors equivalent to those with a sampling technique
that would have needed to use 17, 33, 33 and 13 randomly selected sites. This represents a significant gain in estimation accuracy if the same number of samples is used, or in sampling efficiency, allowing fewer samples to be collected to achieve the same level of potential error. It should be noted that the simulations using nine sample sites selected with the aid of spatial information still produced potential errors of ±10% to ±23%. To achieve estimates with potential errors of about ±5% requires 36 sample sites selected with the aid of spatial information. Take-home messages (1) Our early results show that, compared with random sampling, use of EM-38 soil electrical conductivity maps (used on the Hunter Valley block) or remotesensing imagery (used on the Riverina block) significantly increases the accuracy of yield estimation. Using the same number of samples selected using spatial data increases estimation accuracy over simple random site selection. Alternatively, fewer samples can be used to achieve the same level of accuracy. (2) A multiplication factor is required to take into account differences between estimates based on mean yield of sample vines multiplied by total vineyard area and the actual machine harvested yield. Using multiplication factors generated from the previous year s data, overall estimation accuracy was reduced to about ±5%. (3) The greater the spatial variability in yield, the greater the number of samples that is needed to get accurate yield forecasts. The high level of spatial variability typical of single vineyard blocks (including those used in this study) resulted in large potential errors in the estimates of yield, in the region of ±10% to 25%, using nine sample sites selected with the aid of spatial information describing the block variability. Acknowledgments This study is supported by the Wine Growing Futures Program. Project assisted by Charles Sturt University s Spatial Data Analysis Network (CSU-SPAN), David Lamb, John Hornbuckle, Rod Muldune, Wyndham Estate Wines, and Hancock Farm Company. References and further reading Bramley R.G.V., Hamilton R.P. (2004). Understanding variability in winegrape production systems 1. Within vineyard variation in yield over several vintages. Australian Journal of Grape and Wine Research 10: 32 45.
Clingeleffer P.R., Dunn G., Krstic M., Martin S. (2001). Crop Development, Crop Estimation and Crop Control to Secure Quality and Production of Major Wine Grape Varieties: A National Approach. Grape and Wine Research and Development Corporation, Wayville, SA. Dunn G.M., Martin S.R. (1998). Optimising vineyard sampling to estimate yield components. The Australian Grapegrower & Winemaker 414a: 102 107. Dunn G.M., Martin S.R. (2003). The current status of crop forecasting in the Australian wine industry. In Bell S.M., De Garis K.A., Dundon C.G., Hamilton R.P., Partridge S.J., Wall G.S. (eds.) Proceedings, ASVO Seminar Series: Grapegrowing at the Edge. Tanunda, South Australia, Australian Society of Viticulture and Oenology, Adelaide, SA. Hall A., Louis J., Lamb D. (2003). Characterising and mapping vineyard canopy using high-spatial-resolution aerial multispectral images. Computers & Geosciences 29: 813 822. Martin S., Dunstone R., Dunn G. (2003). How to Forecast Winegrape Deliveries. Department of Primary Industries, Tatura, Victoria. Wolpert J.A., Vilas E.P. (1992). Estimating vineyard yields: introduction to a simple, two-step method. American Journal of Enology and Viticulture 43: 384 388.