Analysis of hyperspectral images for high-throughput plant phenotyping

Close-range hyperspectral imaging (HSI) has recently gained popularity for its application in high-throughput phenotyping platforms (HTPP), which allow for rapid and non-destructive monitoring and assessment of growth dynamics of plants throughout their vegetative stages. The integration of close-range HSI in HTPPs for the extraction of phenotypic traits is still under development and requires input from the biological as well as the technological perspective. The acquisition of hyperspectral data in close-range setups is challenging and requires a detailed inspection of factors related to the imaging setup, that influence the obtained spectra. Particularly, the reflectance spectra from leaves in close-range imaging are highly influenced by plant geometry and its specific alignment towards the imaging system. This induces high uninformative external variability in the recorded signals, whereas the spectral signature informing on plant biological traits remains undisclosed. In this thesis, a method to deal with these effects is developed. In the proposed method, the standard normal variate (SNV) normalization method was applied to remove linear effects and a clustering approach is employed to remove pixels that exhibit nonlinear multiple scattering effects. Once the uninformative variability due to the external factors has been normalized, the desired information was extracted from the spectral reflectance. In this thesis, two major strategies for the interpretation of reflectance spectra with respect to phenotypic traits were developed, unsupervised and supervised data-driven approaches. In the unsupervised data-driven approach, the plant characteristics were inferred entirely from the spectral fingerprint without any prior information about the plant’s parameters. For this, a spectral similarity measure and a band selection strategy were employed to quantify subtle biological information and characterize the plant growth dynamics. In the supervised data-driven approach, machine learning regression algorithms were employed to model the relationship between the spectral variables and the plant physiological parameters. The developed models were tested for their ability to estimate four plant physiological parameters which highly correlate with the water-deficit stress factors as the proxy for drought stress responses. The proposed methodologies were demonstrated by the study of water-stress and recovery of maize plants in a high-throughput plant phenotyping platform. The results showed that the analysis methods allow for an early detection of drought stress responses and of recovery effects shortly after re-watering. This demonstrates that close-range HSI has a high potential to become a rapid and non-destructive novel technology for high throughput phenotype studies.

Publication year:2019

Keywords:Doctoral thesis

Accessibility:Closed