Phenotyping studies of wheat by multispectral image analysis

Abstract

To meet an increasing demand for food production there is a need for faster genetic gains in Norwegian cereal breeding. Yield gains can be improved by use of High-Throughput Phenotyping (HTP) based on multispectral imaging and application of genomic selection. Several spectral indices have been tested to estimate grain yield, such as the Normalized Differential Vegetation Index (NDVI) and MERIS Terrestrial Chlorophyll Index (MTCI). For the present work, data was gathered from a field trial with 96 plots of 24 wheat cultivars laid out in an alpha-lattice split plot design. The design had two levels of nitrogen (N) fertilization, 75 and 150 kg N/ha, applied at sowing. Also, a larger field trial with 301 breeding lines with two reps of high N fertilization was used. Multispectral images where taken in the wavebands green (550 nm), red (660 nm), red edge (735 nm) and near infrared (NIR) (790 nm) with a Parrot Sequoia multispectral camera combined with a sunshine sensor. This allows vegetation indices to be calculated. In addition, 3D models and Digital Surface Models (DSM) are used to estimate plant height. All cameras and sensors were mounted on a light Unmanned Aerial Vehicle (UAV). Images were taken at regular intervals throughout the growth season. The time series of the vegetation indices showed high values during the period of grain filling for wheat plots that received higher dose of fertilization. The values reached their peak around the period of grain filling before declining when plants approached maturity. For site B, the historical cultivars showed significant differences in NDVI and MTCI, but the indices were weakly correlated with grain yield. On site B, however, the large field with breeding lines, both vegetation indices were associated with grain yield with MTCI showing the strongest correlation coefficient of 0.49. The plant heights computed from the DSM showed deviations of 0.1 to 0.2 meters from the manual measurements, indicating that more sophisticated models are needed for reliable prediction of plant height.