SwissPhenoCam: Country-scale automated tree-phenology tracking from webcam imagery.

Large-scale, high-quality phenological observations of trees are key to a better understanding of the environmental factors that control phenological processes, as well as their responses to a changing climate. Over the last decades, phenocams (i.e., webcams capturing time-lapse images of individual plants,  canopies, or communities) have been shown to be a reliable compromise between ground based human observation and satellite remote sensing. Phenocams combine the advantages of automated, real-time data acquisition and a high resolution that allows for the monitoring of individual organisms. Here, we focus on tree species in Switzerland and lay the foundation for a country-scale phenocam network.

In comparison to the global spatial coverage of satellite data, phenocam coverage is bound by the local implantation of cameras. To mitigate this limitation, we integrated a diversity of sources into our data pipeline: weather cameras, private cameras (e.g., from hotel or ski resorts), as well as cameras specifically installed for phenological observation. Combining those sources, we identified over 150 potential sites across the Swiss territory with cameras installed by the same industrial provider. In our first iteration, we focused on 27 of those sites, prioritizing based on the amount of clearly visible trees. We collected the image time series for each location with up to 12 years of site-level history. Due to the diversity of image sources the temporal resolution varied between 1 and 144 images per day. For each of the sites, we annotated the polygon delineating the boundaries of each tree, or group of trees in image pixel coordinates. Next, we identified the species of each tree via on-site visual inspection. Our dataset contains over 1,700 polygons of individual trees, covering over 20 predominant tree species of Switzerland, and over 1,300 polygons of groups of trees categorized into 5 classes.

To obtain phenological observations from this dataset, we adopt two distinct approaches. First, to relate to on-site observations, we reprocess the data for easy visual inspection and developed an ontology of 16 phenophases (e.g., ‘start of leaf unfolding’, ‘leaf maturity’, ‘start flowering’) that can be readily observed by a human from webcam imagery. Phenophases were defined such that they are meaningful for phenological studies and can be matched with Swiss Phenology Network observations where possible. The visual analysis of the images by phenology experts yields over 13,000 different phenological observations.  Second, to relate to satellite-based phenology metrics, we identified changes in greenness over time for each polygon which correspond to leaf development. 

In this communication, we show our dataset preparation pipeline, as well as a comparative study of the phenological metrics obtained via different means on the same trees: visual analysis of the images, greenness extraction, and citizen network reports. In future works, we will explore how this dataset can be used to train machine learning methods to predict phenological phases from the image time series. We will explore if machine learning methods can allow for precise phenophase identification like in visual inspection, while  being fully automated like greenness extraction.