Progressive Automation of Minirhizotron Image Processing through Advanced Contextualization and Machine Learning

Minirhizotron image collection has benefited from technological advances in camera design and operation, bulk data storage, and remote camera control enabling high-volume image collection. The technology used in processing these minirhizotron images into useful data metrics has not progressed at the same pace. Individual images are often manually interpreted, and researchers are accumulating images faster than can be processed, which restricts their ability to advance critical research. Phase I work will demonstrate solutions to the most significant and technically challenging bottlenecks in the image processing workflow including 1) reliably distinguishing roots from soil and other background materials in images, and 2) quantifying and tracking individual roots and metrics through a time series of images. The proposed approach will include both adaptations of mature image processing techniques used in other disciplines and machine learning techniques customized for minirhizotron imagery. A convolutional neural network will be developed and trained to automate root separation from background soil in minirhizotron images leveraging a technique called, transfer learning, to apply the network to minirhizotron images from new ecosystems with minimal re-training. Feature quantification will be performed by tuning existing image processing tools to capture the root phenology of most use to researchers and the metrics will be stored in an adaptable data structure that can accommodate complex time series data. The value of Phase I accomplishments will be demonstrated through comparative testing in relation to baseline technology, and quantitative evaluation of receiver operating characteristic curves. A fully developed tool will be integrated into a minirhizotron workflow in a software as a service model. This will ensure the tool is based in cloud-computing infrastructure, an important consideration for keeping digital tools relevant, accessible, and useful to the research community. In Phase II we will have developed the software and database backend capabilities to manage the data volume and feature cataloging associated with high-throughput processing of time-series images, including newer high-resolution imaging systems.