Drones, Deep Learning, and Endangered Plants: A Method for Population-Level Census Using Image Analysis

Abstract

A census of endangered plant populations is critical to determining their size, spatial distribution, and geographical extent. Traditional, on-the-ground methods for collecting census data are labor-intensive, time-consuming, and expensive. Use of drone imagery coupled with application of rapidly advancing deep learning technology could greatly reduce the effort and cost of collecting and analyzing population-level data across relatively large areas. We used a customization of the YOLOv5 object detection model to identify and count individual dwarf bear poppy (Arctomecon humilis) plants in drone imagery obtained at 40 m altitude. We compared human-based and model-based detection at 40 m on n = 11 test plots for two areas that differed in image quality. The model out-performed human visual poppy detection for precision and recall, and was 1100× faster at inference/evaluation on the test plots. Model inference precision was 0.83, and recall was 0.74, while human evaluation resulted in precision of 0.67, and recall of 0.71. Both model and human performance were better in the area with higher-quality imagery, suggesting that image quality is a primary factor limiting model performance. Evaluation of drone-based census imagery from the 255 ha Webb Hill population with our customized YOLOv5 model was completed in <3 h and provided a reasonable estimate of population size (7414 poppies) with minimal investment of on-the-ground resources.