Rhizomatic: Next Generation Image Processing for In Situ Fine Root Measurement

The characteristics of fine roots, as observed by minirhizotron imaging, are vital for the fields of both ecology and agronomy and factor into a number of DOE research efforts studying effects of emissions and radiation on plant systems. This imaging provides a wealth of information about the presence, extent, morphology and progression of root systems and other phenomena. The current bottleneck for analyzing such imagery is the laborious process of manually marking roots. The purpose of the present work is to develop a system using machine learning and image processing to automatically detect and characterize roots in minirhizotron imagery. Root image analysis is currently a time and labor-intensive process that involves individual technicians and researchers manually identifying root locations, often using outdated and error prone software. A trained labeler can process between 5 and 7 images per hour, often working on datasets with thousands of images, and their labels can be prone to fatigue-related errors and inconsistency between different labelers. Manually labeling images is therefore a costly diversion from scientific analysis of imagery and a barrier preventing broader use of minirhizotron imagery both in the research and commercial markets. This research approach merges classical computer vision with machine learning to create a pipeline that can automatically detect and characterize root systems using non-destructive minirhizotron imagers. It doesn’t require even more laborious pixel-level mask creation and instead draws its training data from existing point-to-point labeling commonly used by researchers. While this labeling is not itself pixel accurate and often noisy and inaccurate, the custom processing pipeline is able to produce pixel-accurate segmentation, which can then be processed to generate the root characteristics desired by researchers. As there are a wide variety of root and soil combinations, creating a universal model is difficult and likely to trade accuracy in favor of universality. The proposed research recognizes and acknowledges this by focusing on a process that allow both the potential for a universal model, but also an ability to generate more targeted models quickly and with limited user intervention. It has been validated in Phase I research by training successful models against two different datasets from different imaging systems and different environments. In Phase II the successful machine learning methodology that was developed in Phase I will be continued to create "Rhizomatic", a cloud service for automatically detecting and characterizing roots. Placing Rhizomatic in the cloud maximizes the number of users it can serve, while keeping it commercially viable by minimizing deployment and maintenance costs. In addition, the prototype system will include integration with RootSnap!, the freely available root labeling software provided with many minirhizotron systems, to provide users a well-supported front-end interface.