Improving automated root image analysis and its application in fine-root dynamics in a 120-year-old Chamaecyparis obtusa forest using a scanner method

Abstract

Understanding fine-root dynamics and their drivers is crucial for elucidating forest carbon cycling. The scanner method enables non-destructive observation of fine-root dynamics with large soil profile images (216 mm in width × 297 mm in height); however, root extraction using the conventional manual way is significantly effort intensive and time consuming. Recently, automated root segmentation software based on deep learning, A Root Auto Tracing and Analysis (ARATA), has been developed; however, its applicability to site-specific conditions has not been sufficiently tested. In the present study, we evaluated whether transfer learning of site-specific scanned images improves the performance of ARATA. Next, we applied the optimized model to quantify fine-root dynamics and examine their relationships with climatic factors and aboveground growth phenology in a 120-year-old Chamaecyparis obtusa stand. Scanning was conducted periodically in five transparent boxes over 2 years, and fine roots were automatically detected from a total of 115 full-page-size scanned images. Site-specific training greatly improved model performance, as indicated by an increase in the Dice score from 0.63 to 0.76 and in the correlation coefficient from 0.76 to 0.88. In the aged C. obtusa stand, fine-root growth was active in autumn, whereas the mortality was pronounced in summer based on the total root length per imaged area. This seasonal variation was primarily driven by air temperature, with a negative effect. Overall, we quantitatively revealed for the first time that site-specific model training maximizes ARATA performance, thereby accelerating the analysis of fine-root dynamics across a wide range of forest sites.