Analysis of fine root dynamics in forest ecosystems using artificial intelligence-based image segmentation

Abstract

Plant fine roots are an important component of the carbon cycle in ecosystems. Nondestructive methods, such as tracing fine roots in soil cross-section images, have become the mainstay of fine root dynamics analysis in recent years; however, manual tracing methods are slow and suffer from low consistency. To solve these problems, many artificial intelligence (AI)-based image segmentation methods using deep learning have been developed. We aimed to verify the segmentation accuracy of RootPainter, an AI-based software, in a real forest ecosystem and use it to estimate the production rate and turnover time. The images segmented by RootPainter contained many errors that led to overestimation. However, by manually eliminating the errors caused by incorrectly segmented roots in areas with no actual roots (a type of false positive, referred to as a “phantom error.”), segmentation accuracy was greatly improved. The AI-segmented area was 2.34 times larger than that of the conventional method, but after removing the phantom errors, it decreased to 1.21 times the size. The correlation coefficient between these areas also increased. In addition, the time required was 43%–47% less than that required by the conventional method. Furthermore, this hybrid AI segmentation and manual correction method could estimate production rates and turnover times from soil cross-sectional images of actual forests, and the results were comparable with those obtained using conventional methods. Thus, the AI-based segmentation software was shown to be effective in analyzing fine root dynamics using soil cross-section images in natural ecosystems, with appropriate human error-correction assistance.