Novel Deep Learning Methods for Mini Rhizotron Measurements

The goal of this Phase I project is to improve current mini-rhizotron image-processing algorithms and solutions for analysis and production of rapid, automated assessments and quantification of in situ fine root measurement. The foundation for this effort is based on UHVs recent success in developing novel advanced machine learning algorithms which were designed to identify and segment fine roots in soil. The keys to success include; working with a highly skilled team, so that the ability of a skilled human being to detect roots in rhizotron images is transferred to machine learning / artificial intelligence algorithms, and having a very large dataset of mini-rhizotron images to work with. The work plan includes the following objectives: To create ground truth datasets of mini-rhizotron images for evaluating and training neural networks from a skilled team of scientists. To evaluate both state-of-the art neural networks and UHVs novel neural network with the newly created ground truth datasets. To integrate quantification methods for root measurement with the output from the neural networks. To design a software application with an automation pipeline for rapid assessment to be used by current users of mini-rhizotrons. To propose designs for a commercial software for Beta testing in Phase II of this project. The advantages of the proposed technology include the development of a fully automated software toolkit with a data pipeline for the quantification of root system traits. These types of novel artificial intelligence algorithms were recently created and developed by UHV to identify and quantify root traits and soil. This software will be used on images captured from mini-rhizotrons to evaluate rooting responses to abiotic stress and climate change using mini rhizotrons for field experimentation