Evergreen and summergreen classification with Sentinel-2 data, K-means clustering derived labels and Machine learning methods

Circumboreal forests represent close to 30% of all forested areas and are changing in response to climate, with potentially important feedback mechanisms to regional and global climate through altered carbon cycles and albedo dynamics (e.g., Loranty et al., 2018). A large portion of these boreal forests are located in Siberia, Russia. Here the forests are made up of mainly two types: evergreen (coniferous i.e., Pine, Picea) and summergreen (deciduous i.e., Larix) needle-leaf.  The evergreen–summergreen forest zone stretching across Western Central Yakutia is a dynamic vegetation transition zone with high disturbance due to forest fire and potential invasion of evergreen forest taxa to the east into the summergreen dominated forest zone that needs mapping and monitoring.

Sentinel-2 based Remote Sensing offers the opportunity to obtain forest type maps on a 10-20 m spatial scale. We provide in the SiDroForest (Siberian drone-mapped forest inventory) data collection (https://doi.org/10.1594/PANGAEA.933268), a Sentinel-2 data set containing Level-2 Bottom of Atmosphere labelled image patches for the early (April-May), peak (June-July) and late (August-September) summer seasons (van Geffen et al., 2022). This dataset contains 63 30 by 30-meter labelled patches with vegetation labels assigned derived from fieldwork measurements taken by the Alfred Wegener Institute in Siberia, Russia in 2018.

Building on the SiDroForest dataset, we used K-means clustering to perform an unsupervised classification of Sentinel-2 for five  locations from the SiDroForest set.  We then assigned two broad forest classes in the Sentinel-2 images, summergreen and evergreen. We used the SiDroForest Sentinel-2 patches as validation data for the K-means generated classes in addition to the fieldwork and expert knowledge.

The new dataset contains 100,000 labelled pixels, distributed over  the two classes. We created the dataset for three time stamps to include different forest phenophases in the classification. The phenophases make it easier to distinguish between the two types of forests as summergreen’s spectral signal changes significantly over the seasons.

We trained a Gaussian Naïve Bayes (GNB), a Random Forest (RF) and a Decision Tree (DT) classifier on three-time stamps separately. A combination of Sentinel-2 bands and the NDVI were evaluated with the different classifiers. The highest average accuracy score was achieved with a DT classifier and a balanced set for the two classes and the early summer time stamp and the NDVI band (82%). The peak summer also performed decently with 74%, but the accuracy dropped to 60% for the late summer time stamp. 

We used the trained DT to classify Sentinel-2 data at two locations in Siberia ; Lake Khamra and Nyurba. We masked out all non-forest data and created a forest map to measure the distribution of evergreen and summergreen over the larger areas. With our analyses we will improve the understanding of satellite data for monitoring remote places. The insights from the Siberian boreal forests are valuable in analyses of the boreal forests located in other parts of the world as well in these times or rapidly changing climate.