Hyperspectral Modeling of Relative Water Content and Nitrogen Content in Sorghum and Maize

Valerie Cross, Purdue University

Abstract

Sorghum and maize are two of the most important cereal grains worldwide. They are important industrially, and also serve as staple crops for millions of people across the world. With climate change, increasing frequencies of droughts, and crops being planted on more marginal land, it is important to breed sorghum and maize cultivars that are tolerant to drought and low fertility soils. However, one of the largest constraints to the breeding process is the cycle time between cultivar development and release. Early evaluation of cultivars with increased the ability to maintain water status under drought and increases nitrogen contents under nitrogen stress could be the key to decreasing breeding cycle time. New tools for non-destructive, high throughput phenotyping are needed to evaluate new cultivars. These new tools can also be used for monitoring and management of crops to improve productivity. Hyperspectral imaging holds promise as one tool to improve the speed and accuracy of predicting numerous plant traits including abiotic stress tolerance characteristics. In this thesis, hyperspectral imaging projects were designed to develop and test prediction models for relative water content (RWC) and nitrogen (N) content of sorghum and maize. The first study utilized three different genotypes of sorghum in an automated hyperspectral imaging system in greenhouses at Purdue University. From this study, models were developed for relative water content and nitrogen content using the data from all three genotypes collectively as well as the data from each genotype individually. Models developed using the spectral and morphological features obtained from the hyperspectral images are predictive of both relative water content and nitrogen content. The coefficients of determination (R2) for all graphs comparing the predicted relative water content to the reference relative water content of sorghum averaged 0.90 while the same graphs for maize averaged 0.64. The coefficients of determination for all graphs comparing the predicted nitrogen content to the reference nitrogen content of sorghum averaged 0.85 while the same graphs for maize averaged 0.61. Models built only with the spectral features for sorghum were also predictive of both relative water content and nitrogen content. The coefficients of determination for all graphs comparing the predicted relative water content to the reference relative water content of sorghum averaged 0.91 while the same graphs for nitrogen content in sorghum averaged 0.85. The nitrogen content models developed using the data from the Tx7000 genotype are highly predictive of both Tx7000 and B35 but not highly predictive of Tx623. However, models developed using the data from Tx623 are highly predictive of all three genotypes. Another important finding from this study was that the water and nitrogen signals overlap and the most predictive models are developed from data where water and nitrogen vary continuously. Models to predict one factor that do not account for variation in the other factor are not very accurate. The second experiment utilized hyperspectral imaging to characterize RWC and N content of maize. Models for RWC and N content were developed using spectral and morphological features. The models developed for maize were not as predictive as the models for sorghum but they were still predictive of RWC and N content for the models developed using all six genotypes and the models developed using the data from the individual genotypes. Models built using the four half-sibling genotypes were not more predictive than the models based on all six genotypes. The final portion of this thesis explored predictions across species using both the sorghum and maize data. We found that models developed using only sorghum were not predictive of the maize reference measurements. However, when the sorghum and maize data were combined and used to generate models, both the RWC model and the N content model were highly predictive for both reference measurements.

Degree

M.S.

Advisors

Tuinstra, Purdue University.

Subject Area

Agronomy|Plant sciences|Agricultural engineering

COinS