Segmenting Growing and Dying Woody Roots in a Forest Using Deep Learning-Based Software

Fine root dynamics of woody plants are important for understanding the carbon cycle in forests. Fine root dynamics have been studied by manually tracing fine roots in soil profile images; however, manual tracing is very time consuming. Automated fine root tracing by deep learning segmentation of roots from an image has been proposed to reduce the time needed to identify fine roots and analyze their dynamics. The accuracy of this approach in detecting fine root production and mortality is still unknown. We evaluated and compared the accuracy of automated tracing programs that use different deep learning methods for fine root segmentation. We assessed four segmentation methods, ARATA, RootPainter, SegRoot, and TrenchRoot-SEG. Fine roots from a time series image of a soil profile were segmented using each of these methods to detect changes in the fine roots area over time. The results from each of the methods were compared with the manual segmentation results and with each other. The highest correlation in the changes in fine root areas was observed between ARATA and manual tracing. This finding suggests that ARATA may be the most suitable method for tracing fine root dynamics from soil profile images among the automated segmentation methods we tested. Our results will help develop methods for automated tracing of fine roots and promote the study the fine root dynamics in the future.