Automatic scoring of germination in the field , based on UAV acquired high resolution images

Organised by Laboratory of Geo-information Science and Remote Sensing

Wed 29 November 2017 09:30 to 10:00

Venue Gaia, gebouwnummer 101
Room 1

By Eric Verhoeff (the Netherlands)


Quantitative analysis of plant phenotypes in the field has become a major bottleneck due to the large amount of breeding trials that need to be scored and analysed.
Traditionally, phenotyping is done manually by plant breeders by visual scoring and manual measurements. It is labour intensive, non-systematic and susceptible to human error.
By improving the phenotyping efficiency, the number of crosses and environments that can be used for selection will increase leading to an increase in the rate of developing superior crop varieties.
This research presents a novel approach to determine the germination score of maize crops in the field two weeks after emergence based on VHR UAV borne imagery.
The study area consisted of a phenotyping experiment with 3838 plots planted with different maize (Zea Mays) cultivars. Two different image segmentation workflows were developed using open source software.
The image segmentation workflows were designed to be robust under the variable conditions experienced in the field during image acquisition and to be suitable for different crop types.
The effect of changing environmental conditions experienced in the field during data acquisition on the performance of the image segmentation workflows were reported to aid to the understanding of the specific challenges of analysing in field imagery.
The performance of both an edge detection based approach and a matched filtering based approach were applied to count the number of emerged maize plants in the field.
The edge detection based approach resulted in an average accuracy of 47%, largely underestimating the number of plants. The matched filtering based approach resulted in an average accuracy of 97%, only slightly underestimating the number of plants.
The main environmental factors influencing the results were changes in incoming radiation during image acquisition, the heterogeneous background colour and weed patches in the field. Specific steps in the image segmentation workflow were developed to deal with these conditions.
The main limiting factor for the performance of the edge detection based approach were the large number of neighbouring plants with overlapping leaves.
A possible improvement would be an addition to the workflow that can segment objects consisting of multiple neighbouring plants into individual plants.
The matched filtering based approach showed very good results and has great potential for precision agriculture applications. The main limiting factor is the variation in crop size and orientation on the imagery.
By improving the used selection criteria, the accuracy of the algorithm can be increased and the algorithm can be made more versatile.
The performance and potential applications of the developed image segmentation algorithms showed the potential of UAV borne remote sensing in combination with automated image analysis for precision agriculture and high throughput phenotyping.
The image segmentation algorithms have been developed using open source software and work perform on regular personal computers. This makes the developed workflow very suitable for upscaling and use in day to day practices.