Improvements of image processing for wheat ear counting

One of the most important activity of agricultural research insititutes concerns the
agronomical experiments done under different conditions needing many land observations
and valuations to quantify several variables. These observations, although generally
accurate, are visually done by the agriculturist technicians and present numerous drawbacks:
penibility, weak productivity, numerous labor force, limited sampling … Two feasibility studies
lead in our laboratory recently have shown that some of the previous observations, and
particularly the counting of the number of wheat ear per m², can be done by color and/or
texture image processing for images taken directly in the field with a specific acquisition
system. This paper describes the improvements of the previous studies concerning the
image acquisition system, and especially the illumination control, and the justification of
different hypothesis on the number of classes to detect in an image.
The use of a cluster validity index has allowed to prove that 3 classes to determine all the
objects in a wheat ear image are not sufficient. A correlation with a study based on the size
of the analysis window is currently under investigation to improve the ear detection, which is
now of 6%, compared to manual counting done by agriculturist technicians.