Combining image analyses tools for comprehensive characterization of root systems from soil-filled rhizobox phenotyping platforms

Functional plant traits have been recognized to be effective predictors of ecosystem function and plant growth strategies. In this context, root traits have gained significant attention in plant and soil research. Root traits such as root angle, specific root area, root diameter, root length density and total root length are essential for plant productivity, particularly under conditions of limited resource availability, and in turn influence the characteristics of the soil and ecosystem. In cropping systems, cultivars with site-adapted root systems are expected to enhance water and nutrient use efficiency, yield stability, and resilience during periods of climate change. In the context of trait-based breeding, the integration of root systems, however, implies two major challenges: i) measurement/screening of large populations to rank candidates according to their performance, and ii) targeting those traits which are most promising for further selection and crossing. Root systems are complex organs of various sizes and structures resulting from spatial and temporal factors, cellular-level processes of elongation, branching and bending (Hodge et al., 2009). As a consequence, plant roots can also be characterized by parameters measured at different observation scales, from composite descriptors (e.g. total root length or shape of the root system, Kashiwagi © 2021 Institute of Agrophysics, Polish Academy of Sciences M. ALSALEM et al. 258 et al., 2006, Freschet et al., 2021) to single traits (lateral branching number, emergence angle of laterals, Chen et al., 2017). In addition, the different types of root descriptors used in comparative root studies are related to different classification schemes which have emerged over the course of the entire history of root research (branching topology, geometrical shape, developmental order, see Freschet et al. (2021) for a recent review concerning root classification schemes and measurement protocols). Finally, the relationship between root traits and the agronomic/breeding target of improved drought resistance due to higher root water uptake is complex (Vadez, 2014). Thus, optimal measurement strategies for root systems would also consider the linkage between root descriptors and root functionality. The number of root datasets has increased to a significant extent with the advancement of image-based phenotyping. Advanced methods for observing root systems non-destructively such as MRI imaging or X-ray tomography are expensive, require specialized equipment and are still limited in resolution at the fine root scale. Optical imaging approaches, using digital cameras or a scanner, are thus more frequently used to identify morphological, physiological, anatomical and biochemical traits. Rhizobox imaging is a root phenotyping approach which involves plants growing in soil-filled containers beyond the seedling/juvenile stages. This setup aims to approximate field-growing conditions (Nagel et al., 2012; Bodner et al., 2017). It may be considered to be an intermediate approach between root phenotyping on artificial media (filter paper, agar plates) and field root imaging with minirhizotrons (Johnson et al., 2001). The roots are only partially visible about 20-30% (Pfeifer et al., 2014; Bodner et al., 2018) to 75-85% of the total root length (Bontpart et al., 2020). The complexity of the data sets is increased due to discontinuities within the visible root system structures (Chen et al., 2019) and the increasing overlap of the dense mature root systems (Kimura et al., 1999). Image-based root phenotyping is expected to improve information concerning root system characteristics beyond classical field data (biomass, length, surface area per depth, specific root length/area, diameter). It provides an in-situ observation methods to track the development of root architecture over time (lateral numbers, angles, topology, Downie et al., 2015) as well as root-soil interactions influencing the shape of the root systems (inorganic nutrients, Wagner et al., 2020). These root architectural and functional traits are highly relevant to achieving better induction of potentials for improved stress resistance: e.g., a steep root angle essentially drives better access to subsoil water in dry ecosystems (Manschadi et al., 2006), while high lateral branching facilitates the superior exploitation of non-mobile nutrients such as phosphorous (Lynch, 2011). This is particularly important for legumes that have a comparatively high P demand (Pang et al., 2018), while – compared to the fibrous root system of monocots – the root density tends to be lower (Haling et al., 2016). While image analysis has been identified as a general ‘bottleneck’ of current plant phenotyping (Minervini et al., 2015) which has led to many image analysis tools being developed over the last few decades (Lobet, 2017), this still holds particular true for root image analysis. For instance, Delory et al. (2017) compared different root length estimates from digital image analyses using the software packages WinRhizo® (Regent, Quebec, Canada) and ImageJ (NIH, USA) with the macro IJ_Rhizo (Pierret et al., 2013). They concluded that a comparison of root length measurements is substantially influenced by the software used due to the different underlying methods. Rose and Lobet (2019) compared IJ_Rhizo and WinRhizo with regard to the accuracy of the root diameter and volume estimates at different image resolutions. Similarly, they concluded that both software and image quality have a significant impact on reported root measurements. In recent times, approaches utilizing machine-learning have been presented to improve root segmentation and tracking and also to overcome some of the problems with soil-grown root images. Indeed, in-situ root images can be challenging for image analysis software due to i) the low contrast between root and soil in the case of colour-based segmentation (Wang et al., 2019), ii) the growth in parallel and/or high overlap of root axes when plants are grown over longer time periods (Himmelbauer et al., 2004), and iii) the partial visibility of soil-grown roots at the imaged surface (Chen et al., 2019). Therefore, the utility of root phenotyping critically depends on analytical strategies and adequate software tools to extract quantitative descriptors from root images that capture functional traits. In particular, an imaging and analytical strategy should be essentially based on the relationship between root traits and a specific target function, e.g. improved transpiration or the nutrient uptake of crops under resource limiting conditions (Vadez, 2014; Chen et