Developing an image processing pipeline to improve the position accuracy of single UAV images

Authors: FENG, AIJING - University Of Missouri; VONG, CHIN NEE - University Of Missouri; ZHOU, JING - University Of Missouri; CONWAY, LANCE - University Of Missouri; ZHOU, JIANFENG - University Of Missouri; Vories, Earl; Sudduth, Kenneth - Ken; Kitchen, Newell

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/10/2023
Publication Date: 1/25/2023
Citation: Feng, A., Vong, C., Zhou, J., Conway, L., Zhou, J., Vories, E.D., Sudduth, K.A., Kitchen, N.R. 2023. Developing an image processing pipeline to improve the position accuracy of single UAV images. Computers and Electronics in Agriculture. 206. Article 107650. https://doi.org/10.1016/j.compag.2023.107650
DOI: https://doi.org/10.1016/j.compag.2023.107650

Interpretive Summary: Unmanned aerial vehicle (UAV) based remote sensing has been extensively used in precision agriculture applications; however, conventional procedures for UAV data collection and processing require collecting highly overlapped images and stitching them to generate an orthomosaic. Ground control points (GCPs) in the field or UAV onboard real-time-kinematic (RTK) global navigation satellite system (GNSS) data are required to improve position accuracy. These steps to a final product are very time consuming (i.e., hours to days). A previous study developed a framework to process individual UAV images for mapping cotton emergence and the current study aimed to further improve UAV image processing efficiency by extracting geo-referencing information from individual UAV images. The method was tested for both cotton and corn and the results showed that the position accuracy compared favorably to ground truth data collected with an RTK-GNSS. This new method currently allows processing in near-real time and may possibly be implemented in real-time using an onboard computer. The new method was evaluated in two studies to generate field maps for cotton and corn fields, suggesting it could provide a low-cost near real-time tool for mapping emergence parameters at field-scale in both research and agricultural production.

Technical Abstract: Unmanned aerial vehicle (UAV) based remote sensing has been extensively used in precision agriculture applications such as vegetation growth and health monitoring, yield estimation, and irrigation management. Conventional procedures for UAV data collection and processing require collecting highly overlapped images, stitching images to generate an orthomosaic, and using ground control points (GCPs) in the field or UAV onboard real-time-kinematic (RTK) global navigation satellite system (GNSS) data to improve position accuracy. For improving efficiency, a previous study developed a framework to process individual UAV images for mapping cotton emergence. The current study aimed to further improve UAV image processing efficiency by extracting geo-referencing information from individual UAV images. The method was tested in field mapping for both cotton and corn. The improved geo-referencing algorithm steps comprised feature detection and matching, false matches removal, geometric transformation matrix calculation, crop row alignment, geo-referencing of each single image, and mapping. The results showed that the position accuracy when measuring the distance between GCPs using the new algorithm were 0.17 ± 0.13 m and 0.57 ± 0.28 m in a cotton and a corn field, respectively, when compared to ground truth data collected with an RTK-GNSS. This new method did not require GCPs in the field or image post-processing steps, such as image stitching and feature extraction, which allowed processing in near-real time and may possibly be implemented in real-time using an onboard edge computing system. The new method was evaluated in two studies to generate field maps for cotton and corn fields, including stand count, canopy area, mean of day after first emergence, and plant spacing standard deviation. These maps demonstrated the success of the developed methods in providing a low-cost near real-time tool (8.6 and 3.6 s/image for the cotton and corn fields, respectively) for mapping emergence parameters at field-scale use in both research and agricultural production.