Assessing Fine Root Production in Terrestrial Forests: A Comparative Analysis of AI and Human Annotation Using Minirhizotron Images

Fine roots are a major source of the stabilised carbon in soils. However, the response of fine root production to an increase in atmospheric CO2 and its impact on carbon dynamics in terrestrial forests remain poorly understood. Minirhizotrons can help to quantify fine root production and associated carbon dynamics in long-term, in-situ experiments such as Free Air CO2 Enrichment experiments. Yet, using minirhizotrons requires the manual annotation of thousands of images. Artificial Intelligence (AI) technology for image processing is fast developing and has proven to be successful in simple systems, such as agronomous crops. Here, we quantified how AI (RootPainter) annotation compares with humans, and determined the implications in terms of root production and carbon dynamics in a mature deciduous forest (BIFOR-FACE). Firstly, we quantified the variation in outputs of 30 annotated minirhizotron images using AI and human analysts of varying levels of expertise, comparing them to a gold standard established through expert consensus. We find that root annotation varied substantially among humans, with novices and AI over-annotating root length by 244% and 206% respectively, compared to our gold standard. Secondly, we quantified root length for five minirhizotron tubes in March and June (n = 1060 images) using AI and then a trained human analyst. AI over-estimated root length by more than an order of magnitude compared to a trained human user, and there was a poor linear relationship between annotated images with  AI and humans (r² < 0.22 for both months). This over-annotation by AI resulted in inaccurate quantification of root production and mortality, and thus erroneous carbon budget.