Single Plant Detection and Isolation on High Resolution UAV Data

摘要

Rating the quality of the growth of plants in fields is a labor intensive task important for planning the application of pesticides and herbicides. In recent years more robotic focused approaches have emerged as a tool in agriculture to reduce the need for manual labor and increase the data coverage. Automatic rating of growth quality on gathered data comes with its own challenges, mainly identifying the relevant parts of images and assessing the relevant growth indicators correctly. Single plant detection and rating provides even bigger challenges as possible overlap between plants or occlusions due to growing weeds need to be correctly detected. In this work we propose a multi-stage training algorithm to single out plants from high resolution UAV image data. First we pre-annotate detailed images using a classical filtering approach. Afterwards the pre-annotated labels are checked by a human annotator and incorrect labels are removed before a siamese network is trained on the sparse data. Finally image tiles centered on single plant instances are generated for future plant assessment steps. The presented approach is evaluated on data of maize fields recorded as part of the BoniKI project. This work presents the first steps in creating an end-to-end field data analysis pipeline to extract growth quality data from large datasets while reducing human annotation effort to a minimum.