Quantifying physiological trait variation with automated hyperspectral imaging in rice

<p>Understanding the genetic underpinnings of plant physiological responses to the environment is critical to accelerate crop breeding under increasingly variable climates. Advancements in hyperspectral imaging (HSI) have enabled researchers to collect spectral images of plants nondestructively. These spectral images can then be used to infer phenotypic variation using models. To account for the large variability across experiments, developing robust HSI designs for calibration datasets is essential. The data in this publication can be used to develop methods for predicting rice physiological traits from HSI data collected at the Ag Alumni Phenotyping Facility (AAPF) at Purdue University.</p>

<p>Under high and low nitrogen conditions, 15 <em>indica</em> and eight <em>tropical japonica</em> rice genotypes were raised at the AAPF up to 13 weeks old. Fourteen physiological traits, which reflected<b> </b>aspects of<b> </b>plant<b> </b>growth, photosynthesis and water transport, were<b> </b>directly measured. Concurrently, canopy-level HSI (side-view) were employed two to three times per week. The reflectance of these images was averaged across a period of approximately one week to account for the variability of HSI data. The four averaged HSI datasets represented the spectral signals for the physiological trait data collected during the same period.  </p>

<p>The multidimensional physiological and HSI data could be used to examine treatment and subpopulation effects. Subpopulation and treatment effects were found in physiological trait data by using principal component analysis (PCA). On the other hand, HSI data showed clear treatment effect throughout the experimental period. Reflectances consistently centered around 715 nm, close to the red edge part of the spectrum, were important for classifying treatment groupings. Treatment prediction accuracy of HSI data was 80% or greater when the rice plants were six to 10 weeks old, estimated by support vector machines with wavelengths selected from PCA.</p>

<p>HSI data could be used to quantify leaf-level nitrogen content (<em>N</em>, %) and carbon to nitrogen ratio (<em>C:N</em>) by building partial least squares regression (PLSR) models with wavelengths selected from RReliefF algorithm. Relevant code of the above mentioned method could be found at: <a href="https://github.com/To-Chia/rice_imaging_ms.git">https://github.com/To-Chia/rice_imaging_ms.git</a>.</p>

<p>The data demonstrated that HSI data could serve as surrogates for plant physiological traits. The key to the discovery is intimately related to methods of data collection and analysis. Future research could gain insight unto experimental designs based on the findings in this publication.</p>