Deciphering Tropical Tree Communities Using Earth Observation Data and Machine Learning

Publicly available EO datasets offer new possibilities to generate biodiversity information at the community composition level, an essential biodiversity variable, beyond forest type. We demonstrated the potential of Sentinel-2, GEDI LiDAR canopy height and ALOSDEM in discriminating and classifying tropical tree communities in the Western Himalayas, India. For this, tree communities were first identified based on the ordination of field data and subsequently classified using satellite data applying machine learning, i.e. random forest (RF). From the three forest types in the study area, eight distinct tree communities were identified for which classification accuracy increased from single date (75.17%) to multi-date images (85.33%) and further by applying feature selection (88.17%). Whereas the best classification accuracy of 94.66% was achieved when canopy height and topographic variables were also considered. The findings suggest that RF is suitable for mapping tree communities by combining Sentinel-2 with GEDI and DEM parameters.