Tree species classification using multi-temporal Sentinel-2 data

Abstract: The Sentinel-2 program provides the opportunity to monitor terrestrial ecosystems with a high temporal- and spectral resolution. In this study, the utilization of multi-temporal Sentinel-2 imagery and it’s spectral variation due to phenology for classification of common tree species is evaluated at the forest estate Remningstorp in central Sweden. The tree species classes to be classified were: Norway Spruce (Picea abies), Scots Pine (Pinus silvestris), Hybrid Larch (Larix × marschlinsii), Silver Birch (Betula pendula) and Pedunculate Oak (Quercus rubur). The Random Forest classifier (RF) was fitted to four Sentinel-2 images taken during the vegetation period of 2017. The RF classifier was also coupled with the feature selection algorithm Recursive Feature Elimination to form a model with an optimal subset of bands. In addition to the classification, spectral profile plots were constructed for each species to visualize the possibility for identifying the less represented tree species. The use of four satellite images from April 7th, May 27th, July 9th and October 19th resulted in a higher overall accuracy (86.4 %) compared to using single images (71.5 % – 79.4 %). The late spring image (May 27th) was found to be important since it always was included in the most accurate classifications, independently of the number of images. The best combination of bands resulted in a model with 87.6 % in overall accuracy and included 37 of 40 bands. The highest ranked bands were all May bands except the red band, the SWIR 1-2 and red bands from April, July and October. The 5 tree species classes were classified with satisfying results and the Producer’s Accuracy ranged from 73.7 % to 97.4 %.