Recent graduate Kexin Guo from Lund University's Department of Physical Geography and Ecosystem Science has successfully demonstrated how SITES Spectral data can power innovative machine learning solutions for ecological monitoring.
In her Bachelor's thesis, "Enhancing PhenoCam Annotation Efficiency via Transfer Learning: Focus on Snow and Image Quality", Guo developed an automated system to detect snow presence in PhenoCam images from SITES Spectral stations at Lönnstorp and Röbäcksdalen.
The research addressed a critical challenge in phenological monitoring: the time-intensive manual annotation of thousands of images. Using over 5,000 PhenoCam images from the two contrasting Swedish sites, Guo trained a MobileNetV2 deep learning model that achieved 72% validation accuracy while reducing annotation time by 61% – from 44 to 17 seconds per image.
The thesis demonstrates that "The SITES Spectral PhenoCam data provided an excellent foundation for developing automated ecological monitoring tools". The model successfully generalized across sites with different vegetation types and climatic conditions, from the mild winters of Lönnstorp to the extended snow cover at Röbäcksdalen.
This work exemplifies how SITES infrastructure data enables next-generation research approaches, helping scientists track vegetation phenology and climate impacts more efficiently. The developed methods could benefit various ecological monitoring tasks, including phenological transition detection and vegetation damage assessment from extreme weather events.
The mast at the Lönnstorp Research Station with Phenocams (photo: José Beltran)