Crop Yield Estimation


Current industry practices for crop estimation and management–manually counting and weighing clusters–is labor-intensive, expensive, imprecise, does not scale to large vineyards, and is sometimes destructive.

The typical process is for workers to sample a certain percentage of the vineyard, counting the average number of clusters per vine, average weight per cluster and extrapolate based on the number of vines per acre.

Besides the difficulty and expense of getting a large enough sample size, and its destructive nature, the process is inaccurate. The weight of the clusters is constantly increasing until harvest, so the vineyard manager must guess at what percentage of the final cluster weight is the current measurement.

We propose to use sensors to automatically forecast the crop yield with the precision and accuracy, and sufficiently in advance of harvest time, to enable vineyard managers to differentially adjust the crop for optimum fruit quality and yield.

The Problem:

Our Approach : Capture Vines and Detect and Count Grape Crop with Camera Imagery

Left is a graph showing our predictions of the harvest weight of rows in a vineyard. Rows 1 to 4 have 24 Traminette vines each. Rows 5 to 8 have 32 Riesling vines each. Predictions are generated from the functions mapping berry count to crop weight that were calibrated on data from other rows. Our yield estimates have a mean error of 9.8% of the weight of the row. Producing yield predictions at this accuracy at the resolution of a single row surpasses the coarse sampling currently used in vineyards.

Video of detection performance

One of the main challenges in our work is that color is not a viable cue on its own. All varieties of grapes start with green berries after fruit-set and remain green until véraison (beginning of ripening). This make it difficult to identify green fruit against the green background.

Images of clusters and berries were collected post-véraison, processed, and compared to actual harvest yield components. Computer generated berry counts vs. actual harvest crop weights.

Data collected for 224 vines (approximately 1600ft of vines); 96 Traminette vines and 124 Riesling vines.

Correlation r2 = 0.74.

By comparison, the typical manual predictions would take a measurement at a small fraction of the vines and extrapolate, whereas we can take measure every vine non-destructively.

We use shape as the primary cue for detecting green grapes among green leaves. A radial symmetry transform is used to detect possible berries, and contextual constraints are used to account for the shape of grape clusters.

We also use multiple light sources to improve berry detection. Images from the different light sources are combined to recover boundary of depth discontinuities (occlusion boundaries).

Validation: Correlating to Harvest Yield

The following results were collected for a one acre plot of mixed V. vinifera and inter-specific hybrid grape varieties with vertically shoot positioned (VSP) training and basal leaf removal in the cluster zone.

(See video from the experiment below.)