PREDICTION OF LEAF RELATIVE WATER CONTENT USING PATTERNS OF HYPERSPECTRAL INTENSITY
thesisposted on 18.04.2022, 14:50 authored by Mark T Gee JrMark T Gee Jr
Drought is the leading cause of crop loss globally. Breeding for drought tolerance is difficult due to the polygenetic nature of the trait and low heritability of yield under drought. Plant relative water content is a secondary trait that may advance drought breeding programs.
The LeafSpec, a newly developed hyperspectral leaf scanner, was used to test the hypothesis that distribution of hyperspectral information across the leaf can be used to improve prediction of leaf relative water content. Data was collected across two experiments from five different maize genotypes representing temperate and tropical hybrids with varying levels of drought tolerance and inbreds with varying stomatal densities. The hyperspectral intensity averaged across the entire leaf was used to predict relative water content with an R2Prediction of 0.7989. Model performance was tested using additional predictors that quantify:
• Spectral information from multiple regions in the leaf (e.g. base, middle, tip).
• Spectral information from regions segmented by tissue type.
• The distribution of hyperspectral intensity in a cross section parallel to the midrib or in a cross-section perpendicular to the midrib.
• A contour pattern of hyperspectral intensity from the outside edge of the leaf to the midrib.
• Texture features extracted from each wavelength.
The mean spectrum model outperformed previously reported results, potentially due to the elimination of sources of noise and higher quality data produced by the LeafSpec. None of the models with expanded feature sets outperformed the mean spectrum model at a statistically significant level. The hyperspectral signal from the green tissue a third of the way from the base of the leaf and half way between the midrib and edge was the most correlated with relative water content. Models without midrib and vein tissue signals had increased performance. Distribution of the Water Index visually showed improved ability to discriminate leaf RWC as compared to individual wavelengths but this did not translate to improved model performance.
For future work, more data should be collected to improve model robustness and hyperspectral imaging should include SWIR wavelengths that have previously been found useful for predicting relative water content. Exploring indices composed from current spectral bands may lead to improved prediction performance.