Leaf water potential of field crops estimated using NDVI in ground-based remote sensing—opportunities to increase prediction precision

PeerJ

article

Xuejun Dong¹  Bin Peng²  Shane Sieckenius¹  Rahul Raman¹  Matthew M. Conley⁴  Daniel I. Leskovar¹ 

[1] Texas A&M AgriLife Research and Extension Center at Uvalde;Yancheng Institute of Technology;Department of Soil and Crop Sciences, Texas A&M University, College Station;USDA-ARS, U.S. Arid-Land Agricultural Research Center

关键词: Drought stress;    Leaf water status;    Statistical resampling;    Multispectral sensor;    Ordinary least-squares;    Proximal sensing;    Weighted least-squares;    Measurement errors;    Diurnal change;    Model coefficients;   

DOI  :  10.7717/peerj.12005

学科分类：社会科学、人文和艺术（综合）

来源: Inra

【 摘 要 】

Remote-sensing using normalized difference vegetation index (NDVI) has the potential of rapidly detecting the effect of water stress on field crops. However, this detection has typically been accomplished only after the stress effect led to significant changes in crop green biomass, leaf area index, angle and position, and few studies have attempted to estimate the uncertainties of the regression models. These have limited the informed interpretation of NDVI data in agricultural applications. We built a ground-based sensing cart and used it to calibrate the relationships between NDVI and leaf water potential (LWP) for wheat, corn, and cotton growing under field conditions. Both the methods of ordinary least-squares (OLS) and weighted least-squares (WLS) were employed in data analysis, and measurement errors in both LWP and NDVI were considered. We also used statistical resampling to test the effect of measurement errors of LWP on the uncertainties of model coefficients. Our data showed that obtaining a high value of the coefficient of determination did not guarantee a high prediction precision in the obtained regression models. Large prediction uncertainties were estimated for all three crops, and the regressions obtained were not always significant. The best models were obtained for cotton with a prediction uncertainty of 27%. We found that considering measurement errors for both LWP and NDVI led to reduced uncertainties in model coefficients. Also, reducing the sample size of LWP measurement led to significantly increased uncertainties in the coefficients of the linear models describing the LWP-NDVI relationship. Finally, potential strategies for reducing the uncertainty relative to the range of NDVI measurement are discussed.

【 授权许可】

CC BY