A Comparison of Non-destructive Imaging and Destructive Load Cells for Grape Yield Estimation

Vineyard systems are spatially and temporally variable in terms of soil and vine characteristics. However, soil and vine management decisions are applied uniformly, which leads to inefficiencies in water and nutrient use and also limits potential fruit yield and quality. The first step to enabling variable management in vineyards is to be able measure the variability. Many methods are available, but most are complex, expensive and feasible for just a small subsample of the vine population. Ideally measurements should be inexpensive and include the entire vineyard. We present and discuss two sensing strategies for efficiently measuring crop size variation, deployed across an 11-year old, 10.5-hectare Petite Sirah vineyard in Galt, California. One strategy is non-destructive, using continuous images of the fruit zone of the vines taken from a farm vehicle during the growing season. The non-destructive measurements are generated from image-processing algorithms that detect and count the berries within the imagery. The second strategy is a destructive measurement as it is taken at the time the vineyard is machine-harvested, with the fruit continuously weighed by a load cell placed under the belt of the discharge conveyor. Both forms of measurement are geo-referenced, dense and high-resolution, and are processed into yield maps that capture the spatial variability in a vineyard. We compare and contrast the respective yield maps to assess the accuracy of measurements. In particular, we consider the spatial accuracy by comparing the respective yield variation patterns between the maps. We find striation patterns with large variations in yield through the vineyard, running parallel with the rows. The patterns raise a number of topics, such as the origin of the yield patterns (underlying soil variation vs. vineyard management practices) and also the question of how to sub-sample the measurements effectively to reduce the chance that unexpected yield variations are missed.  We study the sub-sampling problem in detail and assess the impact of reducing the measurement density on the overall accuracy of the yield forecast.