Elsevier

Geoderma

Volume 406, 15 January 2022, 115507
Geoderma

Characterizing plant root parameters with deep learning-based heat pulse method

https://doi.org/10.1016/j.geoderma.2021.115507Get rights and content

Highlights

  • A deep learning-based heat pulse method was developed for root parameters characterization.

  • Numerical model was leveraged to generate the synthetic data in training deep learning models with virtually no cost.

  • Transfer learning was employed to adapting the knowledge gained from numerical model to actual real-world problems.

  • The proposed method gave satisfactory results for root fragment number identification and diameters estimation.

Abstract

Accurate quantification of root structure is essential for understanding plant ecological system. Nevertheless, the lack of in-situ root characterization methods has limited further research of the relevant field. Heat pulse (HP) technique has emerged as a promising, cost-effective in situ monitoring approach. The existing HP technique relies on the analytical solution of heat transfer model with the assumption that the medium under test is homogeneous. Thus, its performance is compromised in dealing with heterogeneous formations. The objective of this study is to establish a novel in-situ HP method to characterize root parameters based on a deep learning technique. With the fully connected neural networks (FCNN), the number of root fragments was firstly estimated by solving a classification problem, and then the size and location of each fragment were estimated by solving regression problems. All FCNNs were firstly pre-trained using synthetic dataset generated by the numerical heat transfer model. Although FCNNs showed promising accuracy on the synthetic testing dataset, they failed to provide satisfactory results on real-world experimental data due to the model error and observation noise, i.e., the deviations between the numerical model and actual situation. To this end, FCNNs were fine-tuned with some experimental data through transfer learning for performance improvement. Under controlled indoor experiment in a sandy soil, the proposed method provided satisfactory results for the estimation of root fragment number and diameters with 83.3% accuracy and RMSE of 1.87 × 10−4 m, respectively. The position estimation had relatively larger error, i.e., RMSE 2.41 × 10−3 m. This study is the first step in translating heat pulse signals to root parameters, which indicates the potential of combining HP technique and deep learning in studying root system.

Introduction

The plant root is an important organ for material and energy transmission of soil–plant system, as it transfers water and nutrients from soil to plants, and regulates the rhizosphere micro-environment by changing the soil structure, water conditions and releasing root exudates into the soil. Root morphology parameters (including root length, root diameter, root surface area and root volume, etc.) are important indicators reflecting plant growth and development, and closely related to physiological functions such as the absorption and utilization capacity of nutrient and water, carbon source distribution, and the ability to adapt environmental stresses (al Hagrey and Petersen, 2011, Zhao et al., 2019). Thus, it is vital to monitor root parameters. However, soil is a non-transparent and complex medium, and the topological structure formed during root growth is very complicated. How to quickly and accurately acquire the root morphology information in situ remains a challenge in plant and soil science.

Several noninvasive imaging methods have been developed to measure plant root length, root density, root diameter, lateral root number and so on, e.g., minirhizotrons methods (Shen et al., 2020), ground penetrating radar (Borden et al., 2014), electrical imaging technique (Zhao et al., 2019), X-ray computed tomography (XCT, Mairhofer et al., 2013). Limited by expensive and technical difficulty. There is still a need for characterization methods of individual roots.

As a cost-effective approach, heat pulse (HP) technique also has the advantages of slight impact on soil structure, automatic, rapid and in-situ measurement, and thus received extensive attention in recent years (He et al., 2018). It has been widely used for determining soil thermal properties, electrical conductivity, water content (θ), bulk density (ρb) and subsurface evaporation at various spatial and temporal scales in both the laboratory and field (Ren et al., 1999, Heitman et al., 2003, Heitman et al., 2008, Lu et al., 2018). However, HP method measures soil properties by analyzing temperature changes with analytical solutions of heat transfer equation (Kluitenberg et al., 1993, Knight et al., 2012). These solutions are based on the assumption of an infinite, homogeneous isothermal medium within an effective measurement volume. Errors of HP method due to soil heterogeneity across a plane would be non-negligible especially when the heterogeneous interface lies between the two probes (Knight et al., 2007, Philip and Kluitenberg, 1999). Therefore, the applicability of HP method is limited to bare soils or soils with sparsely distributed vegetation (Heitman et al., 2008, Deol et al., 2014). Developing new method for interpreting heat pulse signals will open up a new opportunity of HP technology for heterogeneous soil or rhizosphere soil.

Recently, deep learning has emerged as a powerful method in a wide range of problems including objective identification by automatically learning feature representation from given data (Shen, 2018, Sun and Scanlon, 2019, Zhong et al., 2020). Smith et al. (2019) showed that an automated image segmentation method using a convolutional neural network architecture could identify plant roots from images with higher quality than traditional methods. In the field of industrial production, deep-learning-based thermal imaging or electrical imaging techniques have been proposed to quantitatively characterize the material defects. Based on the contrasting thermal properties between defects and surrounding medium, Yang et al. (2019) used laser infrared thermal imaging to record temperature changes on the material surface, and used convolutional neural networks to identify material cracks at different depths. Zhao et al. (2020) employed convolutional neural network and fully connected neural networks (FCNN) to invert the shape, location and size information of defects from electrical impedance imaging and electrical conductivity spatial information, respectively.

Like structural defects, root has a similar impact on surrounding soil heat transfer. Dry roots usually have larger specific heat capacities (>1.2 kJ kg−1 K−1) and smaller thermal conductivities (0.25 W m−1 K−1) than those of most soil minerals (<1.0 kJ kg−1 K−1 and 1–12 W m−1 K−1, respectively) (de Vries, 1963, Horai, 1971, Tabil et al., 2003). In most cases, roots differ considerably from the surrounding soil in water content, thermal properties and hydraulic properties (Lu et al., 2020). When HP method is applied to the root zone soil, the presence of a root system will inevitably change the thermal and hydraulic properties of the root-zone soil, thus provides significant contribution to heat transfer process. It has been shown that the temperature signal monitored by HP sensor will respond differently to roots with different diameters, positions and lengths (Fu et al., 2020). Inspired by above study on structural defects detection, it is of interest to develop a deep learning-based HP technique to characterize in situ plant root parameters by leveraging the contrasting thermal properties between soil and root system and using temperature signals as carrier.

However, the success of deep learning model relies on the availability of a large amount of data. Limited by cost in practical applications, nevertheless, high-quality experimental data are often in short supply, which makes building a high-quality model very challenging. Leveraging multiple datasets (like large synthetic/simulated dataset obtained from numerical models and small experimental dataset) can serve as a powerful method to solve this problem. Transfer learning is commonly used to apply a pre-trained network on a large dataset to solve another problem with some commonalities but insufficient training data. This facilitates a practically fast application for changed environment (Yang et al., 2020).

There usually exits system difference between model outputs and observations of actual word, which can be caused by simplification and/or misrepresentation of real systems, the uncertainty of model parameters and boundary conditions, as well as experimental errors. The characterization of error sources remains particularly challenging (Xu et al., 2017). Common practices often focus on parameter uncertainty while neglecting other factors. The pre-training method can improve model robustness and decrease uncertainty (Hendrycks et al., 2019), thus it is of interest to investigate the application of transfer learning in real-world problems.

In this study, the goal is to inverse root parameters by interpreting HP measuring signals with deep learning method. Limited by the scarcity of experimental data, we use the synthetic dataset generated with heat transfer model to obtain a well-developed FCNN model, i.e., pre-training model. Then the transfer learning is used to adapt this model to experimental conditions, finally the fine-tuned FCNN is experimentally validated.

Section snippets

Heat pulse measurements on soil-root system

The controlled laboratory experiment was conducted on 12 soil columns with root included to obtain HP measurement signal. Three replicated measurements were made on soil columns. Soil texture was sand with sand fraction of 94% and clay fraction of 5%, which was determined with the laser diffraction method (Gee and Or, 2002). Soil organic matter content was 0.9 g kg−1, which was measured with the Walkley-Black titration method (Nelson and Sommers, 1982).

The soybean roots at the maturity stage

Monitoring location design

The existence of roots changes the thermal properties of the block between the heater and sensor probe, which in turn, results in the change of HP signals. For the soil (θ = 0.2 m3 m−3) studied in this research, with Croot (3.32 MJ m−3 K−1) > Cbulk-soil (1.87 MJ m−3 K−1) the root can absorb more heat than bulk soil, causing a much less temperature change in HP signals. With λroot (0.56 W m−1 K−1) < λbulk-soil (1.52 W m−1 K−1), the root can significantly hinder the heat conduct efficiency

Conclusions

A deep learning-based HP method has been developed in this work for root parameter characterization. Using synthetic dataset generated by numerical heat transfer model, the FCNNs were pre-trained to estimate root parameters from temperature signals. We demonstrated that the proposed FCNN model gave satisfactory results for root fragment number and diameters estimation. Next, to further improve the accuracy of FCNN model in practical application, transfer learning was employed and we fine-tune

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work is supported by the National Natural Science Foundation of China (grant 42107313 and 41771254).

References (36)

  • D.A. de Vries

    Thermal properties of soils

  • G.W. Gee et al.

    Particle-size analysis. Methods of Soil Analysis: Part 4—Physical Methods

  • H. He et al.

    Development and application of the heat pulse method for soil physical measurements

    Rev. Geophys.

    (2018)
  • J.L. Heitman et al.

    Field evaluation of the dual probe heat-pulse method for measuring soil water content

    Vadose Zone J.

    (2003)
  • J.L. Heitman et al.

    Sensible heat measurements indicating depth and magnitude of subsurface soil water evaporation

    Water Resour. Res.

    (2008)
  • D. Hendrycks et al.

    Using pre-training can improve model robustness and uncertainty

    Proceedings of the International Conference on Machine Learning

    (2019)
  • K.-I. Horai

    Thermal conductivity of rock-forming minerals

    J. Geophys. Res.

    (1971)
  • G.J. Kluitenberg et al.

    Error analysis of the heat pulse method for measuring soil volumetric heat capacity

    Soil Sci. Soc. Am. J.

    (1993)
  • Cited by (3)

    • Streambed water flux characterization through a Deep-Learning-Based approach considering data worth analysis: Numerical modeling and sandbox experiments

      2022, Journal of Hydrology
      Citation Excerpt :

      Therefore, the estimation results may also suffer the uncertainties resulting from model structural errors of the numerical model. To overcome this limitation, physical information and some experimental data can be introduced into the neural network (Raissi et al., 2019; Xie et al., 2022) to develop a physics-informed and experimental-data-tuned CNN-DWA framework in future studies. Finally, we stress that the CNN-DWA approach is tested with data produced with numerical and well controlled lab-scale experiments that are easily reproducible.

    View full text