Image-based analysis of yield parameters in viticulture

Abstract

Yield estimation is of great interest in viticulture, since an early estimation could influence management decisions of winegrowers. The current practice involves destructive sampling of small sets in the field and a subsequent detailed analysis in the laboratory. The results are extrapolated to the field and only approximate the actual conditions. Therefore, research in recent years focused on sensor-based systems mounted on field vehicles since they offer a fast, accurate and robust data acquisition. However many works stop after detecting fruits, rarely the actual yield estimation is tackled. We present a novel yield estimation pipeline that uses images captured by a multi-camera system. The system is mounted on a field phenotyping platform called Phenoliner, which has been built from a modified grapevine harvester. We use a neural network whose output is used to count berries in single images. In contrast to other existing methods we take the step from the single vine image processing to the plant level. The information of multiple images is used to acquire a count on plant level and the approach is extended to the processing based on the whole row. The acquired berry counts are used as input for the yield estimation, and we explore the limitations and potentials of our pipeline. We identify the variability of the leaf occlusion as the main limiting factor, but nonetheless we achieve a mean absolute yield prediction error of 26% for plants in the vertical shoot positioned system. We evaluate each described stage comprehensively in this study. • Yield estimation from number of visible berries in images. • Automatic, objective and non-invasive high throughput pipeline for viticulture. • Leaf occlusion of yield components has high influence on yield estimate.