Bildklassifikation zur Selektion von Genotypen in der Getreidezüchtung

Abstract The area of application for AI is steadily growing thanks to constant improvements. An area with little attention so far is agriculture. A task that could be done by machine learning lies in grain breeding for example: the rating of grain after the harvest of third-generation grain species. Here species are selected purely based on visual criteria. With the goal of training a model to predict and later practice this task, two datasets for the grain types triticale and wheat with a total of 3000 entries were created. Each entry contains two images to increase the amount of information per entry. Pictures were taken inside a photobox containing a color chart, with automatic color correction being applied as well. A residual neural network and a vision transformer were trained on the datasets. The aim is to perform this work without qualified personnel but with ordinary hardware. Unfortunately all selected entries were captured before all not selected entries, which served as a confounder during training and made models useless beyond their own dataset. Therefore less than 400 entries of the triticale dataset were recorded again in random order. Training the residual neural network on this dataset achieved accuracy slightly but significantly above the baseline on both the confounded and the non-confounded dataset. Attempts with a mixed dataset as well as deconfounding achieved worse results. There is potential for real life use. It is recommended to perform a similar experiment with data again being captured in random order but at a larger dataset size. The full source code can be found at https://gitlab.hochschule-stralsund.de/jonathan. c.a.reinbrecht/bachelorarbeit.