Phenoliner: A New Field Phenotyping Platform for Grapevine Research

Abstract

In grapevine research the acquisition of phenotypic data is largely restricted to the field due to its perennial nature and size. The methodologies used to assess morphological traits and phenology are mainly limited to visual scoring. Some measurements for biotic and abiotic stress, as well as for quality assessments, are done by invasive measures. The new evolving sensor technologies provide the opportunity to perform non-destructive evaluations of phenotypic traits using different field phenotyping platforms. One of the biggest technical challenges for field phenotyping of grapevines are the varying light conditions and the background. In the present study the Phenoliner is presented, which represents a novel type of a robust field phenotyping platform. The vehicle is based on a grape harvester following the concept of a moveable tunnel. The tunnel it is equipped with different sensor systems (RGB and NIR camera system, hyperspectral camera, RTK-GPS, orientation sensor) and an artificial broadband light source. It is independent from external light conditions and in combination with artificial background, the Phenoliner enables standardised acquisition of high-quality, geo-referenced sensor data.

Keywords: Vitis vinifera; big data; geo-information; grapevine breeding; plant phenotyping.

Figure 1

Overview of the Phenoliner. (a) Phenoliner construction plan. (b) Scheme of the sensor layout in the tunnel (marked red above); right: Sensor A: consisting of RGB cameras 1–3 and 5, and a NIR camera 4, left: Sensor B consisting of two hyperspectral cameras; (c) Phenoliner in the vine row.

Figure 2

Tasks within IGG (Institute of Geodesy and Geoinformation) Geotagger 2.0. GNSS: Global Navigation Satellite System; Plant ID: Plant identification; EXIF: Exchangeable image file format.

Figure 3

(a) Terrestrial laser scan of the Phenoliner to measure the lever arm between the GPS antenna and the camera system. (b) Evaluation measurement to test the georeferencing accuracy of the system. The coordinates of the black and white targets at the poles are known and can be compared with the target positions seen in the images.

Figure 4

Experiment set up of POIs at the position of vine stems. (a) Image, selected as the one closest to an POI, with the deviation e as the error of the selection process. (b) Distribution of deviations from 10 runs at different speeds.

Figure 5

(a) Basic principle of the system setup for 3D reconstruction of the full vine row. The Phenoliners tunnel is driven over the vine rows. The MCS is oriented parallel to the rows, capturing images automatically while in motion. (b) Reconstructed point cloud of black (left side) and green grape (right side) varieties. The upper images show the rows from a frontal view and the lower images show grapes from a profile view. Single berry elevations of their spherical geometry are clearly distinguishable.

Figure 6

(a) Rectified image acquired with Sensor A (camera (5), only overlapping part of stereo pair shown). (b) Depth map calculated with stereo pair from camera (3) and camera (5). The brightness indicates the distance to the cameras (white for near points, dark gray for far points). Black pixels indicate positions with no depth which can be assumed to be background. (c) First result of a test for classification with manually set thresholds. The RGB image from (a) and the depth map from (b) are used as input. Classes are: blue for “grapes”, green for “canopy”, brown for “cane”, black for “background”.

Figure 7

Pre-processing of hyperspectral image recording: (a) channel image at 1100nm, (b) clustering of spectral data, colour indicates pixel groups, (c) classification for foreground (leaves) vs. background, (d) image after removal of dark areas and marking of vine position from GPS.

Figure 8

Normalized reflectance spectra for the VNIR 400–1000 nm (a) and SWIR 1000–2500 nm (b) range imaged at 2 p.m. and pooled over west and east side.

Figure 9

Classification accuracy for differentiation of sprayed vs. non-sprayed leaves measured from (a) west and east; (b) only west and (c) only east side. Across the results, the spectral reflectance data of the VIS-NIR range seems to be the more robust predictor for the spray status. Machine Learning and LDA approach are compared.

Figure 10

Relevance profile of the visual-near infrared range based on RBF performance at different recording times.

Figure 11

High intensity reflections result in a wrong point positions at object borders (left image). At berry arches, they may corrupt the spherical geometry of the berries during reconstruction.