Counting Showy Lady’s Slipper Orchids with Machine Learning

Overview

Showy lady’s slipper orchids (Cypripedium reginae) are a critically endangered species in New Hampshire and much of the northeastern United States. They live in temperate regions and are highly adapted to fens, a type of wetland. The label of “critically endangered” indicates that there are less than five distinct populations in a given state. The New Hampshire Academy of Science seeks to monitor showy lady’s slipper populations in the wild for conservation purposes. We are particularly interested in being able to identify the distribution of showy lady’s slippers in a given fen. The goal of this project was to design an efficient and accurate automated method to assess wild lady’s slipper populations. Because of the potential damage caused by walking through a fen, we chose to use a drone to photograph the lady’s slippers. The drone, fitted with a digital camera, was used to take high-resolution images of all parts of a fen located in Strafford, Vermont. Images were taken at heights ranging from approximately 3 to 8 meters with an average height of about 6 meters. We used a version of YOLOv2, an image recognition system, that had previously been translated into Python, to create an object detection model that could count lady’s slippers in a photograph. The images from the drone were human-annotated to mark lady’s slipper flowers and used to train and test this model. When tested with images taken at about 6 meters, similar to the majority of the ones it was trained with, the model averaged 4.8 true positives, 0.2 false positives, and 90.4 false negatives after 100 epochs of training. An epoch is when the machine learning model looks at each image in the training set once in order to learn what constitutes a lady’s slipper. When tested with pictures taken at about 3 meters above the ground, the model averaged 1.4 true positives, 0 false positives, and 11.6 false negatives, after the same number of epochs. This could perhaps be improved by training the model with more and higher quality images. This model, although imperfect, is a substantial first step toward the goal of being able to precisely map the locations and distribution of individual showy lady’s slippers in a fen. The next steps will be to improve the accuracy of the image recognition and modify the program to perform the same computation on videos.