Tue, Aug 16, 2022: 5:00 PM-6:30 PM
ESA Exhibit Hall
Background/Question/MethodsThe properties of roots and their changes over time provide a window into plant health, which in turn is an indicator of the health of an ecosystem. While we have many tools to study the above-ground components of plants, the study of roots is limited to minirhizotron imagery. A minirhizotron is a tube placed into the soil into which researchers may periodically insert an imager. Analyzing the resulting imagery requires someone to manually mark the positions of roots. This is extremely tedious and labor intensive, often prohibitively so for many researchers.To address his problem we have developed an algorithm that combines deep learning based image segmentation methods with traditional image processing techniques to automatically detect roots. We plan to provide this technology as a web service to reduce the root marking burden and make root studies accessible to a wider community of ecologists and agronomists.
Results/ConclusionsWe have trained artificial neural networks on six separate root imagery datasets provided by five researchers around the globe. These represent a range of soil types and plant species from trees to lentils. One of our key challenges was that roots are traditionally marked with lines, instead of the pixel masks required to train a net. Furthermore, they are subject to human error. We addressed this by computing a probability cloud around each root mark before training. This is sufficient for the neural net to interpret the intent of marking the visually distinct root. The resulting root detections are often more accurate that the original human markings.As the neural networks are trained on probability clouds, their outputs are equally fuzzy. We find the center lines of these clouds by thresholding and applying an optimized Zhang skeletonization algorithm. We developed custom algorithms for fitting Bezier curves to these skeletons, finding the width of each root, and computing average color. As a test of the end to end performance of our algorithm, we find a tight correlation between the total root length in human and machine detections with a r value of 0.95.
Results/ConclusionsWe have trained artificial neural networks on six separate root imagery datasets provided by five researchers around the globe. These represent a range of soil types and plant species from trees to lentils. One of our key challenges was that roots are traditionally marked with lines, instead of the pixel masks required to train a net. Furthermore, they are subject to human error. We addressed this by computing a probability cloud around each root mark before training. This is sufficient for the neural net to interpret the intent of marking the visually distinct root. The resulting root detections are often more accurate that the original human markings.As the neural networks are trained on probability clouds, their outputs are equally fuzzy. We find the center lines of these clouds by thresholding and applying an optimized Zhang skeletonization algorithm. We developed custom algorithms for fitting Bezier curves to these skeletons, finding the width of each root, and computing average color. As a test of the end to end performance of our algorithm, we find a tight correlation between the total root length in human and machine detections with a r value of 0.95.