Phenological phenomena of plants are deeply affected by climate change, therefore studies on phenology are rapidly increasing. To collect plant phenology data, digital repeat photography is widely used worldwide, generating massive image data. In response to demands for automatic processing of image big data, several analysis techniques have been developed. For several decades, it had been mainly focused in quantifying leaf phenology. Recently, however, there is growing interest in detecting the changes of flowering phenology. In this study, we utilized machine learning techniques to detect and quantify flowering phenology. For our training dataset preparation, we used downloaded images from ImageNet and Google Images. Due to the fact that white color is more sensitive to illumination changes and possess greater difficulty in processing the images, we focused on the species that only have white colored flowers. We manually annotated every image to the patches that contain flowers or only leaves. For testing the performance of the model, we also prepared several time-lapse videos that show discrete changes in phenological stages. Then, we implemented and trained a Mask Region-Convolutional Neural Network(Mask R-CNN) model. The model, pre-trained with MS COCO dataset, was re-trained with our own dataset and tested with the time-lapse videos. We performed quantification of changing flowering phenology by calculating the polygon area that model predicted as flowered white.
Results/Conclusions
The results were consistently successful in quantifying the changing flower phenology as the calculated area showed progressive growth and decline along with the phenological changes, while showing highest value at the peak flowering moment. The results and insights gained from our study will broaden the understanding of automatic quantification and detection of changes in flowering phenology.