Investigating image-based species identification for birds in the wildlife trade

In Southeast Asia, songbirds, parrots, owls, woodpeckers, and eagles are in high demand. The depletion of wild bird populations in this region led to the declaration of an Asian Songbird Extinction Crisis in 2017. This trade is megadiverse, and there is a lack of technology to identify the many bird species that appear in wildlife markets. This thesis focuses on bird trade in Indonesia and China, where demand for songbirds is particularly intense. To test if machine learning is viable for this problem, Chapter 2 sets a baseline for human identification error with a matching task requiring same/different decisions for pairings of 19 species. Chapter 3 trained, tested, and compared the capabilities of five deep learning computer vision networks and an ensemble of all networks, using a custom-built data collection, processing, and training pipeline. The best model (using the DenseNet-201 network) achieved a high test prediction accuracy (94.4%), using cross-validation for 37 classes on unseen data. We also demonstrate the effects of the visual dominance of cages (25%, 50% and 75% of images occluded) on test accuracy, precision, recall and the F1-score by artificially placing cage bars in the foreground of images for a subset of 26 species. Chapter 4 builds on Chapters 2 and 3 by comparing the human baseline to the top-performing computer model trained to identify the same species. The computer model performed better than the average human accuracy but worse than the best human score. These results suggest that computers can reliably outperform the average, non-expert human in bird identification tasks. It is hoped that this work will help demystify previous roadblocks for using machine learning to identify birds from pictures in wildlife markets. Image-based machine learning approaches hold great promise for identifying birds and other taxa in highly occluded environments.