Evaluation of cotton stand count using UAV-based hyperspectral imagery

Authors: FENG, AIJING - University Of Missouri; Sudduth, Kenneth - Ken; Vories, Earl; ZHOU, JIANFENG - University Of Missouri

Submitted to: Proceedings of the American Society of Agricultural and Biological Engineers International (ASABE)
Publication Type: Proceedings
Publication Acceptance Date: 5/1/2019
Publication Date: 7/7/2019
Citation: Feng, A., Sudduth, K.A., Vories, E.D., Zhou, J. 2019. Evaluation of cotton stand count using UAV-based hyperspectral imagery. Proceedings of the American Society of Agricultural and Biological Engineers International (ASABE). Paper No. 1900807. https://doi.org/10.13031/aim.201900807.
DOI: https://doi.org/10.13031/aim.201900807

Interpretive Summary: It is important to assess crop stand counts in the seedling stage to aid in making replanting decisions. While the conventional method of manually counting plants is time consuming and labor intensive, remote sensing can provide an efficient way to obtain information. Unmanned aerial vehicle (UAV) technology has created interest in developing new remote sensing applications for use in precision agriculture, including early determination of plant stand density. In this project, data collected two weeks after planting from a sensor mounted on a UAV were investigated as an estimator of cotton plant density in a research field in Southeast Missouri. Much of the effort concentrated on distinguishing seedlings from soil and weeds, and the results obtained were more accurate than commercial software. This study has demonstrated the potential of using UAV-based imaging to estimate cotton stand density early enough in the season for remedial action. This approach may be useful to researchers and to farmers who are interested in obtaining early-season plant population estimations at relatively low cost and high resolution.

Technical Abstract: It is important to assess the stand count of crops in the seedling stage for making proper field management decisions, such as replanting, to improve crop production. The conventional method of counting stand manually is time consuming and labor intensive, which makes it difficult to adequately cover a large field. Use of an unmanned aerial vehicle (UAV) as a high throughput tool could make this task more efficient. Cotton seedlings are small and not easily seen in UAV-based RGB images. Hyperspectral images could provide more information than common RGB images and help to distinguish cotton seedlings from soil. The goal of this study was to evaluate the potential of using a UAV-based pushbroom hyperspectral imaging system in cotton stand segmentation (i.e., separating the cotton seedling from the soil background in the images) and counting. A pushbroom hyperspectral camera covering the spectral range of 600 nm - 970 nm was integrated into a UAV platform to collect images of cotton seedlings at the altitude of 50 m above ground level. An image stitching and alignment algorithm including feature detection and matching, geometric transformation, dynamic panorama, and spectral band stitching, was developed to generate a panorama for each of the 103 bands. Vegetation indices were calculated using different bands and were used to remove soil background. A Hough transform was conducted for row identification and weed removal. The geometric characteristics of the identified objects in the images were used to estimate the cotton stand count. The number of cotton plants determined from the hyperspectral images was 2% less than the actual number (84.1% classification accuracy of stand count estimation in each segmented object in the binary images). Mean absolute percentage error (MAPE) = 9% was obtained for density and mean seedling space estimation, and MAPE = 6.8% was obtained for the seedling space standard deviation. The results showed that the UAV-based hyperspectral images had the potential to evaluate cotton stand count. The seedlings identified through the segmentation process will be used for evaluating plant vigor, uniformity and water stress detection in future studies.