3D Reconstruction of Plant Roots from MRI Images to Advance Root-Soil Systems Modelling

Abstract

<strong>Background and Motivation:</strong> Roots are of particular interest for the efficient use of nutrients and water by plants. Therefore, the optimization of root system architecture (RSA) offers large potential in finding more sustainable agricultural practices. Magnetic resonance imaging (MRI) is one of the few phenotyping methods that allows us to observe the 3D RSA in opaque soil. Such volumetric data are essential to investigate favorable RSA traits with functional-structural root architecture models (FSRMs). However, the processing of MRI images and their integration into FSRMs is challenging and limits the use of the data to this day. In this work, we investigated how MRI images of plant roots and related experimental data can be processed more efficiently, and how their meaningful use in FSRMs can be optimized.<br /><strong>Material and Methods:</strong> To alleviate the bottleneck in MRI image processing, we deployed a novel approach for automated root system reconstruction. The approach combines a semantic segmentation of raw MRI images into roots and soil with a root reconstruction algorithm. We evaluated the results by comparing them with state-of-the-art manual expert reconstructions. In the next step, we investigated if the current soil process descriptions in FSRMs are adequate to derive realistic root water uptake (RWU) predictions for RSAs derived from MRI images. We performed a soil grid convergence study of our default modelling approach in CPlantBox and implemented an alternative approach for RWU calculation. The results were evaluated by comparing them to a numerical reference solution. Finally, we explored new methods for the virtual replication of MRI experiments in FSRMs. We devised a novel parameterization method for mimicking root growth based on MRI time series. By combining the measured root growth with additional experimental data, we performed a virtual repetition of an MRI experiment.<br /><strong>Results:</strong> We observed that the U-Net segmentation improved reconstruction performance in manual and automated workflows of root system reconstruction and allowed us to process MRI images more efficiently. Furthermore, the segmentation enabled the application of the automated reconstruction algorithm for MRI images with a low contrast-to-noise ratio. The soil grid convergency study highlighted that root system scale models are not able to spatially resolve the steep soil water potential gradients near plant roots during water uptake. This resulted in large errors in simulated RWU for dry soil conditions. The implemented alternative approach for RWU calculation showed the best agreement with the reference solution, while the computational cost was kept low. Mimicking root growth based on MRI time-series data with the novel parameterization method allowed us to derive time-dependent root system metrics and to create a functional representation of growing root systems. By combining this functional representation of growing root systems with additional experimental data, we have created a parameterization framework that allows a data-driven replication of the observed RWU in CPlantBox.<br /><strong>Conclusions:</strong> We were able to improve several aspects of the 3D reconstruction of plant roots from MRI images and their integration into root-soil-system models. The improvements to manual and automated workflows for RSA reconstruction will facilitate the parameterization of RSA submodels with MRI data. In addition, the ability to derive RSAs from low CNR images broadens the general scope of MRI experiments. The grid convergence study raised awareness for errors related to current RWU modelling paradigms under drought conditions. Using the alternative approach for RWU calculation makes it possible to bring the level of detail of FSRMs closer to that of MRI-based RSAs. The novel parameterization method for virtual replication of MRI experiments facilitates the parameterization of RSA submodels based on time-dependent root system metrics. Furthermore, the parameterization method refines our ability to validate the mechanisms and assumptions underlying RWU in FSRMs.