Assessing tree root distributions using ground‐penetrating radar and machine learning algorithms

Abstract Tree cultivation provides food, raw materials, carbon sequestration, and many other ecosystem services. Developing innovative approaches for tree analysis to help optimize their management is crucial. Cherry trees provide numerous health and economic benefits, with Michigan home to 75% of the cherry trees grown in the United States. In this study, we investigated the coarse root architecture of tart cherry trees using non‐invasive imaging techniques to reconstruct their spatial distribution and extent. Roots from matured orchards in Michigan were imaged using ground‐penetrating radar (GPR) with an 800 MHz antenna. The processed radiograms were analyzed using MALA Vision software, through which a three‐dimensional cube was generated. Depth slices extracted from this cube were subsequently analyzed using convolutional neural networks—a novel approach employed to identify and extract root patterns from the imaging data. A nondestructive, controlled root experiment was conducted to validate and assess the detection capabilities of the GPR frequency employed. The findings from this experiment informed the image interpretation process used to reconstruct root geometry. Results indicated that the GPR could detect and reconstruct coarse roots with diameters as small as 4.3 cm. To establish an allometric relationship between root systems and canopy size, an unmanned aerial vehicle was utilized to estimate tree canopy dimensions. Comparative analysis revealed that the lateral extent of coarse roots was approximately 1.2 times larger than the canopy area. Finally, a separate experiment involving root proxies was developed to create a predictive model for root biomass, achieving an accuracy of 95%.