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

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.