Automatic tree species classification from Sentinel-2 images using deficient inventory data

The paper presents a novel tree species classification approach based on an automated analysis of Sentine1-2 images, where deficient inventory data from an analysis area are available for reference. The procedure proposed involves clustering of one or several multispectral images of the area of interest taken in different seasons, with the subsequent assignment of a particular tree species class to each cluster by using sparse and imprecise inventory data. The analysis is performed only within the forest areas separated using the forest mask obtained from LiDAR scanning data. Classification results are tested against the ground truth prepared manually from remote sensing images or taken from the same kind of imprecise inventory data. An analysis of a sample area in Latvia for which the sparse inventory data was available was performed. Several clustering algorithms were compared, and it is shown that recently proposed Dynland clustering algorithm has advantages over the traditional k-means clustering. Identification of the most important 4 species was targeted, and it is shown that considerable classification accuracy, with a kappa coefficient up to 0.9 can be achieved in this case by the approach proposed. The proposed approach can be used within automatic workflows for estimation of forest microstand parameters from remote sensing data.