+ All Categories
Home > Documents > Phenotyping-v2 nathalie (1) - HSI | Hyperspectral …imaging&for&plant&phenotyping& Puneet Mishra1,...

Phenotyping-v2 nathalie (1) - HSI | Hyperspectral …imaging&for&plant&phenotyping& Puneet Mishra1,...

Date post: 08-May-2018
Category:
Upload: lamcong
View: 217 times
Download: 0 times
Share this document with a friend
2
Close range hyperspectral imaging for plant phenotyping Puneet Mishra 1 , Mohd Shahrimie M. A. 1 , Stien Mertens 2,3 , Stijn Dhondt 2,3 , Nathalie Wuyts 2,3 , Paul Scheunders 1 1 iMinds, Vision Lab, University of Antwerp, Antwerp, Belgium 2 Department of Plant Systems Biology,VIB, Gent, Belgium 3 Department of Plant Biotechnology and Bioinformatics, Gent University, Gent, Belgium INTRODUCTION Plant phenotyping signifies comprehensive assessment of quantitative and qualitative functional plant traits. It includes monitoring the interaction of a plant of particular genetic characteristics with its environment. Typically, during plant growth, the surrounding environment influences its structural and functional traits. Together, the structural and functional traits determine plants performance in terms of biomass and yield. Therefore, it is of crucial importance to monitor the structural and functional traits for understanding plant growth dynamics. Traditional methods used for phenotyping are still time consuming, labor intensive and destructive in nature (1). This fosters the need for high throughput and non-destructive technologies (2). Hyperspectral imaging (HSI) combines spectral and spatial information and is nowadays emerging as a potential tool for studying functional plant traits (3). Spanning a broad range of the electromagnetic spectrum in a fast and non-destructive way, HSI creates opportunities for high throughput phenotyping. Present work explores the potential of VNIR HSI of a high throughput phenotyping platform PHENOVISION (http://www.psb.ugent.be/phenotyping/phenovisi on) for early drought stress detection in maize (Zea mays) plants. Ten different maize plants belonging to a particular genotype were monitored with HSI for a period of nine days of drought induction. Out of ten, five were watered normally and the other five were put in drought. All experiments were performed with the fully automated PHENOVISION, located in the greenhouse infrastructure of the VIB Department of Plant Systems Biology, Belgium. MATERIALS AND METHOD All images were recorded with a line scan push broom VNIR HS camera (ImSpector V10E, Spectral Imaging, Oulu, Finland) installed in PHENOVISION. The camera is located in a dedicated enclosed cabin with a lift and rotating platform for sample presentation, and two lighting racks with nine equally spaced 35 W halogen lamps for illumination. While recording an image, the lighting racks move along with the camera in push broom sense to maintain a homogeneous light distribution. The acquired images have a spatial resolution of 510 × 328, and spectral range of 400-1000 nm. All images were radiometrically calibrated by subtracting a dark frame and calculating relative reflectance with respect to a white reference (Spectralon). In close range HSI, the spectra get affected due to the local leaf angles and curvature of plants. Therefore, to correct for these effects spectral smoothening and normalization (Standard Normal Variate) was performed for each individual image pixel (4). To capture the spectral differences emerging in plants during drought induction, clustering (6 cluster centroids) was performed on the normalized spectra. To provide a biophysical explanation to the cluster centroids, the Red Green Ratio Index (RGRI) was used. The obtained cluster centroids were in normalized scale, therefore, the index values for the centroids were predicted with the help of a linear regression model developed with an independent offline experiment with a leaf. To highlight the discrimination between the drought and control plants, weighted averages based on predicted RGRI values and corresponding cluster proportions were calculated for the complete duration of the monitoring. RESULTS The raw and pre-processed spectra are presented in Fig 1a,b. Fig1c, shows the standard deviation of the raw (blue) and pre-processed (red) spectra. It can be seen that after the pre-processing a major variability due to illumination effects got reduced and various peaks can be identified over the spectral range (Fig1c). The spectral regions
Transcript

Close  range  hyperspectral  imaging  for  plant  phenotyping  

Puneet Mishra1, Mohd Shahrimie M. A.1, Stien Mertens2,3, Stijn Dhondt2,3, Nathalie Wuyts2,3, Paul Scheunders1

1iMinds, Vision Lab, University of Antwerp, Antwerp, Belgium 2Department of Plant Systems Biology,VIB, Gent, Belgium

3Department of Plant Biotechnology and Bioinformatics, Gent University, Gent, Belgium INTRODUCTION Plant phenotyping signifies comprehensive assessment of quantitative and qualitative functional plant traits. It includes monitoring the interaction of a plant of particular genetic characteristics with its environment. Typically, during plant growth, the surrounding environment influences its structural and functional traits. Together, the structural and functional traits determine plants performance in terms of biomass and yield. Therefore, it is of crucial importance to monitor the structural and functional traits for understanding plant growth dynamics. Traditional methods used for phenotyping are still time consuming, labor intensive and destructive in nature (1). This fosters the need for high throughput and non-destructive technologies (2). Hyperspectral imaging (HSI) combines spectral and spatial information and is nowadays emerging as a potential tool for studying functional plant traits (3). Spanning a broad range of the electromagnetic spectrum in a fast and non-destructive way, HSI creates opportunities for high throughput phenotyping. Present work explores the potential of VNIR HSI of a high throughput phenotyping platform PHENOVISION (http://www.psb.ugent.be/phenotyping/phenovision) for early drought stress detection in maize (Zea mays) plants. Ten different maize plants belonging to a particular genotype were monitored with HSI for a period of nine days of drought induction. Out of ten, five were watered normally and the other five were put in drought. All experiments were performed with the fully automated PHENOVISION, located in the greenhouse infrastructure of the VIB Department of Plant Systems Biology, Belgium. MATERIALS AND METHOD All images were recorded with a line scan push broom VNIR HS camera (ImSpector V10E, Spectral Imaging, Oulu, Finland) installed in

PHENOVISION. The camera is located in a dedicated enclosed cabin with a lift and rotating platform for sample presentation, and two lighting racks with nine equally spaced 35 W halogen lamps for illumination. While recording an image, the lighting racks move along with the camera in push broom sense to maintain a homogeneous light distribution. The acquired images have a spatial resolution of 510 × 328, and spectral range of 400-1000 nm. All images were radiometrically calibrated by subtracting a dark frame and calculating relative reflectance with respect to a white reference (Spectralon). In close range HSI, the spectra get affected due to the local leaf angles and curvature of plants. Therefore, to correct for these effects spectral smoothening and normalization (Standard Normal Variate) was performed for each individual image pixel (4). To capture the spectral differences emerging in plants during drought induction, clustering (6 cluster centroids) was performed on the normalized spectra. To provide a biophysical explanation to the cluster centroids, the Red Green Ratio Index (RGRI) was used. The obtained cluster centroids were in normalized scale, therefore, the index values for the centroids were predicted with the help of a linear regression model developed with an independent offline experiment with a leaf. To highlight the discrimination between the drought and control plants, weighted averages based on predicted RGRI values and corresponding cluster proportions were calculated for the complete duration of the monitoring. RESULTS The raw and pre-processed spectra are presented in Fig 1a,b. Fig1c, shows the standard deviation of the raw (blue) and pre-processed (red) spectra. It can be seen that after the pre-processing a major variability due to illumination effects got reduced and various peaks can be identified over the spectral range (Fig1c). The spectral regions

showing major variability were 550-600 nm (Green), 650-700 nm (Red) and 700-750 nm (Red edge). Fig 1d,e presents the cluster centroids obtained from raw and pre-processed spectra. Fig 1f shows the standard deviation of the cluster centroids.

Figure 1: (a) Raw spectra, (b) Normalized spectra, (c) Standard deviation of spectra, (d) Raw cluster centroids, (e) Normalized cluster centroids, and, (f) Standard deviation of centroids. It can be seen that the centroids obtained from raw spectra were only capturing the difference in illumination conditions, whereas, after normalization, the remaining differences in the centroids are assumed to be entirely coming from the biophysical changes in the plants due to the drought stress.

Figure 2: Cluster map before (b) and after normalization (c). The proportion maps obtained from the cluster centroids of raw and pre-processed spectra are presented in Fig 2. Without pre-processing, the cluster maps were only explaining the regions of different illumination over the plant. However, after pre-processing, the obtained maps (Fig 2c) identify the different biological parts of the plant such as a young leaf, old leaf, vein etc. The weighted average index values obtained from the predicted index value of cluster centroids and respective cluster proportion is presented in Fig3. The red colour represents the plants under drought. It can be seen in Fig 3, that

drought can be detected from the 4th day of induction.

Figure 3: Weighted average RGRI values for complete duration of experiment. CONCLUSION Spectral pre-processing helped in removing the illumination effects caused by leaf angle, curvature and distance to the source of illumination. Drought stress was detected on the 4th day of drought induction, confirming the potential of HSI for early drought stress detection studies and opening the doors for further application of HSI in plant phenotyping . REFERENCES 1. Busemeyer, L., Mentrup, D., Mo ̈ller, K., Wunder, E., Alheit, K., Hahn, V., Maurer, H. P., Reif, J. C., Wu ̈rschum, T., Mu ̈ller, J., Rahe, F., Ruckelshausen, A., 2013. BreedVision A Multi-Sensor Platform for Non-Destructive Field-Based Phenotyping in Plant Breeding. Sensors 13 (3), 2830.

2. Hawrylak-Nowak, B., 2008. Changes in Anthocyanin Content as Indicator of Maize Sensitivity to Selenium. Journal of Plant Nutri- tion 31 (7), 1232–1242.

3. Bergsträsser, S., Fanourakis, D., Schmittgen, S., Cendrero-Mateo, M. P., Jansen, M., Scharr, H., Rascher, U., 2015. HyperART: non- invasive quantification of leaf traits using hyperspectral absorption- reflectance-transmittance imaging. Plant Methods 11 (1), 1–17. 4. Mohd, S.M.A., Mishra P., Mertens S., Dhondt S., Wuyts N., Scheunders P., 2016. Modeling Effects of Illumination and Plant Geometry on Leaf Reflectance Spectra in Close-Range Hyperspectral Imaging. IEEE 8th Workshop on Hyperspectral Imaging and Signal Processing.


Recommended