Unsupervised Approach to Forest Tree Species Diversity Assessment With Satellite Observations

Forest tree species diversity plays a critical role in maintaining ecosystem resilience and function. However, large-scale assessments remain challenging due to the limitations of field-based and supervised remote sensing methods, which require costly training data and species-level labeling. In this study, we propose an unsupervised approach to estimating tree species diversity based solely on satellite imagery (Sentinel-2 or Landsat-8) acquired during the 2019 growing season. The method integrates vegetation indices (GNDVI, EVI, NDMI), Self-Organizing Maps, and spectral clustering to derive the Evenness Index without the need for species classification. Validation against field data from over 10,000 hexagonal grid cells (10 square kilometers each) across Poland shows strong agreement, with Pearson's r = 0.87 (Sentinel-2,$R^{2}$= 0.75) and r = 0.81 (Landsat-8,$R^{2}$= 0.66). Because the approach does not require ground-based training data, it can be directly integrated into operational forest monitoring frameworks, including national forest inventory programs. This scalable, label-free method enables the repeatable monitoring of tree species diversity at national and continental scales.