High-throughput phenotyping of yield parameters for modern grapevine breeding

Grapevine is grown on about 1% of the German agricultural area requiring one third of all fungicides sprayed due to pathogens being introduced within the 19th century. In spite of this requirement for viticulture a reduction is necessary to improve sustainability. This objective can be achieved by growing fungus resistant grapevine cultivars. The development of new cultivars, however, is very time-consuming, taking 20 to 25 years. In recent years the breeding process could be increased considerably by using marker assisted selection (MAS). Further improvements of MAS applications in grapevine breeding will come along with developing of faster and more cost efficient high-throughput (HT) genotyping methods. Complementary to genotyping techniques the quality, objectivity and precision of current phenotyping methods is limited and HT phenotyping methods need to be developed to further increase the efficiency of grapevine breeding through sensor assisted selection. Many different types of sensors technologies are available ranging from visible light sensors (Red Green Blue (RGB) cameras), multispectral, hyperspectral, thermal, and fluorescence cameras to three dimensional (3D) camera and laser scan approaches. Phenotyping can either be done under controlled environments (growth chamber, greenhouse) or can take place in the field, with a decreasing level of standardization. Except for young seedlings, grapevine as a perennial plant needs ultimately to be screened in the field. From a methodological point of view a variety of challenges need to be considered like the variable light conditions, the similarity of fore- and background, and in the canopy hidden traits. The assessment of phenotypic data in grapevine breeding is traditionally done directly in the field by visual estimations. In general the BBCH scale is used to acquire and classify the stages of annual plant development or OIV descriptors are applied to assess the phenotypes into classes. Phenotyping is strongly limited by time, costs and the subjectivity of records. Therefore, only a comparably small set of genotypes is evaluated for certain traits within the breeding process. Due to that limitation, automation, precision and objectivity of phenotypic data evaluation is crucial in order to (1) reduce the existing phenotyping bottleneck, (2) increase the efficiency of grapevine breeding, (3) assist further genetic studies and (4) ensure improved vineyard management. In this theses emphasis was put on the following aspects: Balanced and stable yields are important to ensure a high quality wine production playing a key role in grapevine breeding. Therefore, the main focus of this study is on phenotyping different parameters of yield such as berry size, number of berries per cluster, and number of clusters per vine. Additionally, related traits like cluster architecture and vine balance (relation between vegetative and generative growth) were considered. Quantifying yield parameters on a single vine level is challenging. Complex shapes and slight variations between genotypes make it difficult and very time-consuming. As a first step towards HT phenotyping of yield parameters two fully automatic image interpretation tools have been developed for an application under controlled laboratory conditions to assess individual yield parameters. Using the Cluster Analysis Tool (CAT) four important phenotypic traits can be detected in one image: Cluster length, cluster width, berry size and cluster compactness. The utilization of the Berry Analysis Tool (BAT) provides information on number, size (length and width), and volume of grapevine berries. Both tools offer a fast, user-friendly and cheap procedure to provide several precise phenotypic features of berries and clusters at once with dimensional units in a shorter period of time compared to manual measurements. The similarity of fore- and background in an image captured under field conditions is especially difficult and crucial for image analysis at an early grapevine developmental stage due to the missing canopy. To detect the dormant pruning wood weight, partly determining vine balance, a fast and non-invasive tool for objective data acquisition in the field was developed. In an innovative approach it combines depth map calculation and image segmentation to subtract the background of the vine obtaining the pruning area visible in the image. For the implementation of HT field phenotyping in grapevine breeding a phenotyping pipeline has been set up. It ranges from the automated image acquisition directly in the field using the PHENObot, to data management, data analysis and the interpretation of obtained phenotypic data for grapevine breeding aims. The PHENObot consists of an automated guided tracked vehicle system, a calibrated multi camera system, a Real-Time-Kinematic GPS system and a computer for image data handling. Particularly developed software was applied in order to acquire geo referenced images directly in the vineyard. The geo-reference is afterwards used for the post-processing data management in a database. As phenotypic traits to be analysed within the phenotyping pipeline the detection of berries and the determination of the berry size and colour were considered. The highthroughput phenotyping pipeline was tested in the grapevine repository at Geilweilerhof to extract the characteristics of berry size and berry colour using the Berries In Vineyards (BIVcolor) tool. Image data acquisition took about 20 seconds per vine, which afterwards was followed by the automatic image analysis to extract objective and precise phenotypic data. In was possible to capture images of 2700 vines within 12 hours using the PHENObot and subsequently automatic analysis of the images and extracting berry size and berry colour. With this analysis proof of principle was demonstrated. The pilot pipeline provides the basis for further development of additional evaluation modules as well as the integration of other sensors.