Toni Kazic

Research

A complex phenotype is formed by multiple genes interacting together nonlinearly and with the organism’s environment and human management practices. One example is the plant’s immune response to pathogen infection: the plant encloses the pathogen in a visible lesion to prevent the infection’s spread. These lesions are very similar to the lesions that form spontaneously in the disease lesion mimic mutants of maize we use in our field work. Deconvolving these nonlinear interactions requires both higher resolution data over many thousands of individuals and better approaches to modeling.

We develop computational methods to extract and understand high resolution phenotypic information on individual maize plants and their leaves captured on video by consumer-grade drones. Flying ‘low and slow’, we image individual plants and their lesions at high resolution. We stitch the video into an overall view of the field with our mosaicking pipeline that uses only the pixel information in the videos, without relying on telemetry, GPS signals, or ground control points, simplifying efficient data acquisition. Our pipeline uses our deep learning model that aligns pairs of frames; this model is faster and more accurate than the standard algorithms. We use this imagery to count seedlings and build 3D models of plants, rows, and fields. We are developing methods to segment lesions from leaves, leaves from plants, and plants from rows using deep learning. Once we have hundreds of thousands of lesions, we will map these phenotypes to their high-dimensional manifold. We also build machines to simplify field operations.