Plant community analysis with convolutional neural networks

Plant community composition and phenology are key environmental indicators, but their manual assessment is laborious, subjective, and error-prone. This dissertation establishes the field of automatic plant community analysis, proposing a deep learning framework using Convolutional Neural Networks (CNNs) to estimate species composition and phenology from images. This work tackles challenges like small/imbalanced datasets and occlusion through methods including pre-training, data augmentation, and a zero-shot approach. Evaluated on three novel datasets, the trained system outperformed independent ecologists by 29% in accuracy and showed greater consistency. The zero-shot system also matched expert performance without prior community data. The work further indicates dataset parameters crucial for future applications.