Image-based Estimation of Oat Panicle Development Using Local Texture Patterns

Flowering time varies between and within species, profoundly influencing reproductive fitness in wild plants and productivity in crop plants. The time of flowering, therefore, is an important statistic that is regularly collected as part of breeding programs and phenotyping experiments to facilitate comparison of genotypes and treatments. Its automatic detection would be highly desirable. We present significant progress on an approach to this problem in oats (Avena sativa L.), an underdeveloped cereal crop of increasing importance. Making use of the many thousands of images of oat plants we have available, spanning different genotypes and treatments, we observe that during flowering, panicles (the flowering structures) betray particular intensity patterns that give an identifiable texture that is distinctive and discriminatory with respect to the main plant body and can be used to determine the time of flowering. This texture can be located by a filter, trained as a form of local pattern. This training phase identifies the best parameters of such a filter, which usefully discovers the scale of the panicle spikelets. The results demonstrate the success of the filter. We proceed to suggest and evaluate an approach to using the filter as a growth stage detector. Preliminary results show very good correspondence with hand-measured ground truth, and are amenable to improvement in several ways. Future work will build on this initial success and will go on to locate fully mature panicles, which have a different appearance, and assess whether this approach can be extended to a broader range of plants.