Extracting Switchgrass Features Through Minirhizotron and Hyperspectral Image Processing
Authors:
Mira Saldanha1* ([email protected]), Jeffrey Drew1* ([email protected]), Justin Rossiter1, Jason Bonnette2, Felix Fritschi3, Alina Zare1, Thomas Juenger2
Institutions:
1University of Florida; 2University of Texas–Austin; 3University of Missouri
Goals
This team’s goal is to develop computer vision software pipelines for efficient analysis of minirhizotron and hyperspectral images of switchgrass.
Abstract
Both minirhizotrons and unmanned aerial vehicles (UAVs) can provide a massive amount of image data on plants like switchgrass (Panicum virgatum), a potential source of biofuel. However, manual analysis of these images is time-consuming. Researchers focus on developing computer vision software pipelines to segment the images, or classify pixels based on their respective imagery, and accurately quantify data for these segmentations and raw images.
Minirhizotrons allow researchers to track the growth of the same plant roots over time. For any automated analysis, it will be necessary to align the images so that calculations from pixel differences are accurate. Researchers use the Binary Robust Invariant Scalable Keypoints (BRISK) algorithm for feature detection (Leutenegger et al. 2011), and then random sample consensus (RANSAC) to calculate a homography between matched points, to align two images from different dates. Minirhizotrons provide eight images, in color, of the roots at different depths. Researchers experiment with aligning raw minirhizotron images versus segmented images, where roots have been identified, and with aligning each level separately versus all at once, when images have been stitched together. Additionally, the group demonstrates an analysis pipeline for UAV hyperspectral data of individual switchgrass plants. This pipeline produces both individual segmentations of switchgrass plants, and extractions of vegetation indices for the respective plants using the hyperspectral data and segmentations. The team first obtains segmentations of individual plants by using a combination of Sparsity Promoting Iterated Constrained Endmembers (SPICE) and manually inputting key-points for each plant (Zare and Gader 2007). Then, the team applies the watershed algorithm to assign boundary labels for each plant from the binary SPICE output. The team’s minirhizotron analysis pipeline is able to identify a change in biomass over time, and alignment results are promising. Researchers’ hyperspectral analysis pipeline calculates all vegetation indices after performing radiance, reflectance, and orthorectification processing and stitching together all data cubes. The group’s combined aim with this research is to expedite data collection and analysis for biologists.
References
Leutenegger, S., et al. 2011. “BRISK: Binary Robust Invariant Scalable Keypoints,” International Conference on Computer Vision, 2548–55. DOI:10.1109/ICCV.2011.6126542.
Zare, A., and P. Gader. 2007. “SPICE: A Sparsity Promoting Iterated Constrained Endmember Extraction Algorithm with Applications to Landmine Detection from Hyperspectral Imagery,” Proceedings of SPIE 6553. DOI:10.1117/12.722595.
Funding Information
This research was supported by the Office of Science BER program, U.S. DOE, Grant DE-SC0021126.