Ecological data are costly to collect, and budget limitations or other logistical constraints often preclude sampling the breadth of conditions necessary to, e.g, model ecological phenomena with high accuracy. Advances in computer vision offer the potential to expand our sampling capacity via rapid and accurate data collection with conventional visible light cameras. Convolutional neural networks are employed in industry for computer vision tasks and may achieve high classification accuracies given sufficient training data. One means to circumvent large data requirements for these models is to pretrain a network on a larger more general dataset and then update parameters via additional training on a smaller more specific dataset of interest in a process known as transfer learning. I explored a transfer learning approach to classify lichens, an understudied and important component of the vegetation community in many systems. I retrained a ResNet101 convolutional neural network to classify lichen genera within imagery, using only freely available photo data drawn from collections and field specimens in online herbaria. From these sources I assembled a dataset containing 26 genera, with 400 photos per genus. I utilized 300 photos per genus for model training and validation, and retained 100 of each genus for model testing.
Results/Conclusions
I achieved a test accuracy of 83% across the 26 target lichen genera. Sources of error were in part due to photo quality, and accuracy improvements may be possible with additional quality control. For example, some herbaria photos were labeled with only a single species though multiple species were present in a scene. Automated cropping processes also excluded relevant information in certain instances. The photo data in this project were not collected in a systematic way, and I suggest standardized protocols to further improve outcomes. Lichens exhibit a diversity of pigments and are visually similar throughout life history stages, and thus may be more amenable to computer vision classification techniques compared to other groups. Vascular plants may prove a more challenging target for this task as they can vary greatly in form from seedling to senescence. Early agricultural applications of computer vision models for weed control show promise, however. I will discuss ongoing work exploring methodologies for classifying vascular plant species in natural systems using a computer vision approach. These early findings demonstrate the potential for rapidly detecting target groups of vegetation with visible light cameras from ground-based imagery in systems with less complex canopies, such as rangelands and tundra.