Automatic Plant Counting and Location Based on a Few-Shot Learning Technique
Azam Karami, Melba M. Crawford, Edward J. Delp
Abstract
Plant counting and location are essential for both plant breeding experiments and production agriculture. Stand count indicates the overall emergence of plants compared to the number of seeds that were planted, while location provides information on the associated variability within a plot or geographic area of a field. Deep learning has been successfully applied in various application domains, including plant phenotyping. This article proposes the use of deep learning techniques, more specifically, anchor-free detectors, to identify and count maize plants in RGB images acquired from unmanned aerial vehicles. The results were obtained using a modified CenterNet architecture, with validation performed against manual human annotation. Experimental results demonstrated an overall precision $>$95% for examples where training and testing were performed on the same field. Few-shot learning was also explored, where the trained network was 1) directly applied to the fields in other geographic areas and 2) updated using small quantities of training data from the other locations.