Current and Future Technologies for Wine Grape Crop Estimation Luis Sanchez E&J Gallo Winery 16th Annual Enology & Viticulture Conference & Tradeshow Penticton, BC / July 20, 2015
Vineyard Yield Forecasting Business need for improved methods: Effectively manage grape supply Know the quantity Anticipate quality Estimate the cost Harvest logistics and winery capacity efficiency Metric based cultural practices Pruning Irrigation Shoot and cluster thinning
Business goal: Vineyard Yield Forecasting Wineries would like +/- 5% accuracy In Australia, an reduction in error from 33% to 20% was valued as having an $85MM annual value for the industry (Updated: $100-200MM) The accuracy of our current crop estimation methods typically ranges between 15% and 35% average
Challenges for yield forecasting Vineyard spatial variability Annual variability in cluster & berry weight Hang time/berry desiccation in red cultivars Variable machine harvesting efficiency Time and cost of measure-based methods
Yield (tons/acre) Cluster weight (g) Spatial variability 250 200 150 100 50 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 Vine # 20 16 12 8 4 0 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 Vine #
0.6 Cluster weight at commercial harvest (lbs) 0.5 0.4 0.3 0.2 0.1 2004 2005 2006 0.0 DLM B6 DLM B10 Lodi 1 Lodi 3 Laguna N Laguna Ripp 5B Ripp 4A Schmierer Leventini R14 6A1 6A2 Block Seasonal bunch weight estimates may vary up to 50% from historical averages 1A 4D D23 T12 3B 5A 3B 9B D03 D14
Cluster weight (% of maximum) Hang time (Extended maturation) 100 Lodi Merlot 90 80 70 60 34% Max 25 Brix 26 Brix 27 brix 28 Brix 50 2004 2005 2006 2007 2008 Maximum weight at 21.1 to 23.6 Brix
Yield components For count/measurebased methods Yield per acre Y ie ld p e r vin e V in e s per acre C lu ste rs p e r vin e In-season measurement C lu ste r w e ight Inflorescence s p e r sh o o t Shoot s p e r vin e B e rrie s per cluster B e rry w e ight Cluster prim ordia p e r n o d e N o d e s p e r vin e P e rce n t b u d b re a k F lo w e rs p e r inflorescenc e P e rce n t fruit set B e rry abscisio n Pre-season measurement Modified after Tim Creagh, EIT, NZ
Yield components Cluster weight Yield/vine = Clusters/vine x berries/cluster x berry weight 60% 30% 10% Contribution to variation in yield
MAPE = Mean Absolute Percentage Error Source: GMS, blocks > 3 acres, all varieties 60 50 40 Bunch Count Forecast GVI SJV COASTAL 30 20 10 0 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 60 50 40 Bunch Count Forecast Non-GVI SJV COASTAL 30 20 10 0 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
Mean vs. absolute difference Block Estimated Tons/acre Delivered Mean difference (%) Absolute difference (%) A 12.5 9.5 31.6 31.6 B 5.5 7.5-26.7 26.7 C 3.0 4.0-25.0 25 D 7.8 6.0 30.0 30 E 6.0 7.8-23.1 23.1 F 14.0 12.0 16.7 16.7 G 7.0 9.2-23.9 23.9 Average 8.0 8.0-2.9 25.3
9 All GVI Mean weighed yield (tons per acre) 8 7 6 5 4 Forecast Actual 10% 27% 20% 17% 18% 11% 15% 22% Mean absolute percent differences by block 2002 2003 2004 2005 2006 2007 2008 2009 Figure 1. Estimated vs. actual mean weighed GVI yield per acre and mean absolute percent differences
Research on measure-based forecasting Pre-season Weather modeling Bud dissection In-season Trellis tension monitors Grape Forecaster Pronofrut CMU sensor External Spain, France, Australia, Germany
Research on measure-based forecasting Pre-season Weather modeling Bud dissection In-season Trellis tension monitors Gallo-Commercialized Grape Forecaster Pronofrut CMU sensor External Spain, France, Australia, Germany
100 spurs 50 canes Bud dissection size = 3 nodes long size = 10-15 nodes long
Bud dissection Potential fruitfulness = cluster primordia per node Dissect under stereomicroscope Count inflorescence primordia per node Report inflorescence primordia per spur or cane
Potential fruitfulness vs. actual yield
2002-2011 yearly mean yields for SJV Barbera (tons/acre) 2002-2011 yearly mean yields for SJV Barbera (tons/acre) Bud dissection block selection Individual vs. average regional yields for selected SJV Barbera blocks Jean Hardy Barbera Ray Pool Barbera 15 15 14 14 13 13 12 12 11 11 10 10 9 8 7 y = 0.5419x + 3.9437 R² = 0.9023 9 8 7 y = 0.3864x + 6.2292 R² = 0.7007 6 6 4 6 8 10 12 14 16 4 6 8 10 12 14 16 2002-2011 yearly yields for block (tons/acre) 2002-2011 yearly yields for block (tons/acre)
Barbera Cab. Sauv. Chard Chenin Blanc Fiesta F. Colombard Grenache Merlot Muscat of A. Petite Sirah Pinot gris Pinot noir Rubired Riesling Sauv. blanc Syrah Symphony Teroldego Thompson White Zin Zinfandel Current bud dissection blocks S. Joaquin Valley 2 8 6 5 4 6 2 7 8 2 6 3 6 4 3 3 2 2 3 2 7 North Coast 9 3 4 3 2 Central Coast 3 4 1 2 Washington 4 4 4 1 Total 135 blocks (from 24 in 2004 / 66 in 2011 / 87 in 2012) SJV (91) / N. Coast (22) / C. Coast (12) / WA (13) More emphasis on varieties by region
Laser ablation tomography
In-season measurements
Lag-time method for estimating yield 2004 2006 Tons per acre (% of maximum) 100 80 60 40 20 Merlot Sonoma 511 578 643 Degree-day bio-fix for 50% crop weight too inconsistent to provide accurate estimates 0 0 500 1000 1500 2000 2500 Growing Degree Days Chiotti CH2 Barrelli Creek A04 Barrelli Creek A16
Grape Forecaster Measure-based software system designed for vineyard sampling and yield forecasting Result of a 10-year research effort by the Australian wine industry Stratified random sampling spots in a vineyard block clumping stratified random vs. random
Grape Forecaster Segments of row rather than whole vines Segment = slice of known length across a vine row Minimizes time, difficulty and expense of sampling Ideal length function of: vine age - training pruning 1 meter for GVI Coastal / 60 cm for GVI-SJV
Grape Forecaster Validation Research 2007-2010 2007: Tested GF on 38 blocks System statistically sound Software was made more Gallo-friendly Accuracy 10-20% 2008: Shortcut studies (row and cluster sampling) Harvest efficiency determinations Coastal GVI for cluster counting only ~ 190 blocks 2009: Coastal GVI : 300+ blocks 2010: Coastal+SJV GVI 450+ blocks
Impact of Grape Forecaster Implementation Performance in Napa Eastern Section Percent of blocks at each MAPE class MAPE (%) 2012 2013 2014 5 19% 25% 46% 10 32% 47% 65% 20 52% 77% 87%
Grape Forecaster Summary Can achieve MAPE under 20% Requires dedication and sufficient labor Data QA is essential data automation?
Research on measure-based forecasting Pre-season Weather modeling Bud dissection In-season Trellis tension monitors Research level Grape Forecaster Pronofrut CMU sensor External Spain, France, Australia, Germany
Weather-based forecasting Baldwin, Australia, 1960 s
Adjusted R 2 (%) Factors that Improved Model Fit LODI CABERNET SAUVIGNON GDD + Precipitation + Frost + Prior Year Yield + Like Pattern + Year + Interactions
MAPE (%) Opportunities Current model regression trend-line over time with data increase Current model average MAPE 10% MAPE possible by 2017 with additional years of data if trend continues. Model improves with additional years of data
Pronofrut, DEYANU-Chile Uses current available spatial data Pre-sampling for assessment of variability Systematic sampling sequence Not random Precise spots: every n rows, n vines, n clusters Precise vine count not needed Time consuming
Results for wine grapes in Chile Company Variety Duration (h:m) Area (ha) Error (%) Sta. Emiliana Carmenere 1:14 3.1 7.6 Undurraga Carmenere 2:42 10.2 1.2 Undurraga Cabernet 9:37 50 + 0.4 Undurraga Cabernet 6:44 50 3.3 Juvei Camps Chardonnay 7.7 1.4 Juvei Camps Pinot Noir 11.4 + 6.0
Trellis tension monitors
CMU sensor Automation of berry counting
CMU sensor
Berry imaging system
Accuracy Berry Imaging System can identify and count 99.9% of the berries present in an image Due to canopy occlusion, system measures between 15% and 30% of the actual berries per vine Need block calibration to adjust berry count to actual yield Currently characterizing and modeling canopy occlusion for GVI vineyards
Yield Monitor September 2013 Image Estimate June 2013 Automation of berry counting Estimated yield = 7.91 t/a Actual yield = 7.10 t/a
Progress since 2010 Collaboration and support of CMU s berry imaging and counting system: Year Coverage/ shift Results turnaround 2010 ½ row 1 month 2012 5 rows 2 weeks 2014 25 acres 24 hours
External research Wall-Ye V.I.N. robot France VineRobot Europe PHENObot Germany USW, Australia
vitisflower Spain Flowers per cluster counting app for smartphones (free) University of La Rioja (Vinetics Research Project)
Knowledge/data General strategy for the future - GVI Forecast Final July May January Yield weather model Bud dissection Historical data Inputs Maturity (IBMP), HiRes NDVI, Weather Berry counts (Grape Forecaster, Pronofrut, CMU Sensor), HiRes NDVI Bunch counts Grape Forecaster, Pronofrut Accuracy
General strategy for the future - GVI Leveraging of GIS data Yield maps HiRes NDVI Canopy volume (LiDAR/ PhoDAR)
Conclusion Spatial and temporal vine variability are the main challenges for correct sampling and accurate yield prediction in vineyards Cost of measured-based forecasting is still a fraction of the potential benefit of accurate estimation Bud dissection and improved count-based forecasting are the main contributions of our research effort Through research partnerships worldwide we will continue developing and testing new platforms for more efficient sampling and accurate yield estimation