1Department of Computer Science, University of Saskatchewan, Saskatoon, Canada
2Department ofPlant Sciences, University of Saskatchewan, Saskatoon, Canada
| Received 23 May 2024 |
Accepted 06 Oct 2024 |
Published 08 Nov 2024 |
Plant population counts are highly valued by crop producers as important early-season indicators of field health. Traditionally, emergence rate estimates have been acquired through manual counting, an approach that is labor-intensive and relies heavily on sampling techniques. By applying deep learning-based object detection models to aerial field imagery, accurate plant population counts can be obtained for much larger areas of a field. Unfortunately, current detection models often perform poorly when they are faced with image conditions that do not closely resemble the data found in their training sets. In this paper, we explore how specific facets of a plant detector’s training set can affect its ability to generalize to unseen image sets. In particular, we examine how a plant detection model’s generalizability is influenced by the size, diversity, and quality of its training data. Our experiments show that the gap between in-distribution and out-of-distribution performance cannot be closed by merely increasing the size of a model’s training set. We also demonstrate the importance of training set diversity in producing generalizable models, and show how different types of annotation noise can elicit different model behaviors in out-of-distribution test sets. We conduct our investigations with a large and diverse dataset of canola field imagery that we assembled over several years. We also present a new web tool, Canola Counter, which is specifically designed for remote-sensed aerial plant detection tasks. We use the Canola Counter tool to prepare our annotated canola seedling dataset and conduct our experiments. Both our dataset and web tool are publicly available.
