Abstract:
Plant phenotyping, the measurements of various important plant traits, is essential to the advancement of plant breeding and precision agricultural practices, for improving agricultural production. This is particularly useful for field crops such as maize, that include several varieties grown in multiple regions and have a wide range of usage in food and non-food products. However, field-based plant phenotyping can pose challenges towards acquiring adequate amounts of reliable data,especially image data and extraction of traits from images.
In this thesis, we address the problem of maize phenotyping using image analytics. In the first part of our work, the focus is on computing lengths of maizetassels, an important component of maize, from field image data. We have developed a pipeline that has tassel detection, localization of foreground tassel and length computation tasks performed sequentially. We have used faster region based convolutional network for detection of tassel objects from field images of maize crop.
We utilized the luminance and colour information of the extracted tassel patches using YCbCr model to locate the foreground tassel pixels. Piecewise linear computation of tassel lengths is then obtained using a Hough transform based approach.
We have experimentally demonstrated the efficiency of tassel detection and length computation in our pipeline. In the second part of the thesis we address the limited availability of field data for maize tassel phenotyping. We proposed a method, TasselGAN, for synthetically
creating maize tassels against sky background. In this method, we utilized deep convolutional generative adversarial network (DC-GAN) to generate maize tassels and sky patches separately. The generated tassels and sky patches are later merged to form field like maize tassel data. We modified the original design of DC-GAN by removing batch normalization layers and adding residual layers to enable generation
of maize tassels. We have also conducted a detailed ablation study related to these changes. Quantitative and perspective qualitative evaluation of generated field based maize tassel data is carried out to show that the generated data is fairly realistic, and that it can improve the quality of phenotyping.