Distinguishing natural and planted riparian forests by the Natural Numerical Network and the graph-Laplacian

Our study focuses on identifying and classifying Natura 2000 habitats using Sentinel-2 multispectral data. The Natural Numerical Network is a deep learning algorithm for the classification of complex structures such as plant communities. It is based on the optical information from Sentinel-2 satellite bands and the basic statistical characteristics calculated from that information. Using the Natural Numerical Network, desired areas are classified, and relevancy maps are created. The relevancy map tells us about the relevancy of the classification of the segmented area into the chosen habitat. Our research is putting emphasis on the riparian forests along the Danube River. We construct the mean graph-Laplacian and show its application in distinguishing the natural riparian forests of the Natura 2000 system with high biodiversity from the planted monodominant forests with a similar species composition. The basic idea is that the natural forests are represented by much higher variability of the optical data from satellites than the planted ones. Using the relevancy maps calculated by the Natural Numerical Network, we find the potential Natura 2000 habitat-riparian forest areas, and the mean graph-Laplacian eliminates the planted forests from the relevancy maps by assigning the low or zero values to the areas with low optical data variability.