LEVERAGING SEEDLING CHESTNUT ORCHARDS TO RAPIDLY BREED HIGHER YIELDING VARIETIES USING DRONE- BASED AERIAL IMAGERY AND GENOMIC PREDICTION

A key crop in U.S. agroforestry systems is Chinese chestnut (Castanea mollissima), and its interspecific hybrids with Japanese (C. mollissimaxcrenata) and European chestnut (C. mollissimaxsativa). Chestnuts are a unique temperate nut crop with a starchy, rather than oily, texture (Li et al. 2022a). They can serve as a staple food and as a replacement for maize in processed food and industrial applications (Paciulli et al. 2018; Raczyk et al. 2021). Modern chestnut cultivars are precocious annual bearers (Hunt et al. 2009) and produce nuts with an amino acid balance similar to milk and eggs (Sacchetti et al. 2004). The global market for chestnuts is currently $7.5 billion USD and has been growing at an average annual rate of 3.3% since 2017 (FAOSTAT 2022). Like all woody perennial crops, cultivation of chestnuts also provides a number of ecological benefits. Their perennial root system can help capture excess nutrients and reduce eutrophication of surface waters, and as a permanent landscape cover, they provide habitat for birds, beneficial insects, and other wildlife (Roces-Diaz et al. 2018). And perhaps most significantly, chestnuts store large amounts of carbon: they sequester >0.75 t carbon/acre in woody biomass over their first five years, scaling to more than 8 t carbon/acre sequestered by maturity (Wolz et al. 2018).Aim 1: Development of remote-sensing methods to unlock on-farm genetic resources using two mature chestnuts orchards;Aim 2: Utilization of high-density molecular markers to calculate genomic-estimated breeding values of seedling trees currently growing in mature chestnut orchards.Aim 1. Development of remote-sensing methods to unlock on-farm genetic resources using two mature chestnuts orchards?Background: A limitation with evaluating chestnuts within a traditional breeding program is the substantial acreage required to grow large seedling populations, due to their size at maturity, and the decades required for trees to reach their mature yield-potential (Hunt et al. 2009). However, in contrast to most tree crops, which rely on the cultivation of clonal germplasm, farmers are currently growing large seedling populations, due to the risk of graft failure in the Midwest climate (Jaynes 1975). These seedling trees primarily represent half-sib families descended from high-performing maternal parents (Rutter et al. 1991), and as such, they comprise the types of progeny families that would be evaluated within many breeding program designs. Such existing on-farm genetic diversity therefore offers a unique opportunity to implement a participatory breeding program, wherein phenotypic data is collected from existing mature orchards (Revord et al. 2022). This performance data, combined with genotypic information, can then be used to make genomic predictions of complementary parents for future crosses, as well as cheaply screen future seedling populations to improve the overall performance of seed lots grown by farmers. A key trait in this regard is yield density, but the substantial labor requirements associated with manually collecting nuts from individual trees - which often drop over a multi-week period - and normalizing these measurements for total canopy size, has historically prohibited farmers and researchers from measuring this trait quantitatively.Maximally utilizing this on-farm genetic diversity will require the development of protocols that can be applied to thousands of trees across multiple locations, in order to exploit the substantial extant genetic diversity (Li et al. 2022b), minimize the contribution of spatial heterogeneity to environmental variance components (Heslot et al. 2014; Oakey et al. 2016; Mao et al. 2020), and ultimately optimize training population composition (Isidro et al. 2015; Akdemir and Isidro-Sánchez 2019; Berro et al. 2019). This project addresses the current lack of scalable methods to rapidly and accurately perform such phenotyping for key agronomic traits such as canopy burr density, burr size, and total canopy volume. Specifically, this approach will combine manual, ground-truthing evaluation of trees with remote sensing using drones equipped with high-resolution cameras (see example images in Fig. 1). Recent advances in drone imagery have made application of these methods both economically efficient (Weinstein et al. 2019), and scalable to the evaluation of large numbers of individual trees in horticultural environments (Johansen et al. 2018; Tu et al. 2019; Dong et al. 2020). Specific to this project, previous research suggests phenotyping of key agronomic traits in chestnuts is now feasible using such remote sensing methods (Di Gennaro et al. 2020; Pádua et al. 2020). In addition, deep learning methods have been successfully used to identify chestnut burrs using RGB images taken from the ground (Adão et al. 2019). Aim 2. Utilization of high-density molecular markers to calculate genomic-estimated breeding values of seedling trees currently growing in mature chestnut orchardsBackground: The immediate manner in which these seedling chestnut orchards can be utilized within a breeding program is through the identification of superior parents. These would then be crossed to produce improved progeny families for cultivation by farmers. Given this ultimate objective, it is critical to develop methods to not simply select trees exhibiting high yield density, but rather to select on the basis of maximal breeding value for yield density. Particularly for a trait such as yield, where non-additive variance components are known to be large (Hardwick and Andrews 1980), it is essential to partition variances, and select specifically on the basis of additive genetic value (Lynch and Walsh 1998). An ideal method for doing so in a diverse set of half-sib families, such as represented in these two orchards, is to use dense molecular markers to calculate genomic-estimated breeding values (Bernardo 2002). Such an approach will be preferred to alternative methods aimed at detecting QTL (such as genome-wide association, or multi-parent linkage mapping) that could then be selected for using marker-assisted selection. For traits such as yield density, which have multiple underlying components, and are therefore likely to be under highly polygenic control, the phenotypic variance explained by detectable QTL is expected to be very low, despite their potentially high heritabilities (Brachi et al. 2011).