Exploring Sentinel-2-Based Spectral Variability for Enhancing Grassland Diversity Assessments Across Germany

Abstract

Questions: Can remote sensing data support the assessment of High Nature Value (HNV) conservation categories in the German HNV monitoring scheme? Specifically, does spectral pixel-to-pixel variability improve classification accuracy of HNV categories based on Sentinel-2 data?. Location: Germany. Methods: We used multispectral Sentinel-2 imagery (10 m resolution) from 5 years (2017–2021) to classify HNV categories. Random Forest models were trained using different predictor combinations, including spectral data, phenology, and geographical location. We applied various cross-validation strategies to assess classification accuracy. Results: Classification accuracy was generally low (≈44%) when using target-oriented cross-validation, suggesting limited agreement between predictions and actual HNV categories. Spectral variability alone did not clearly correspond to HNV diversity categories. Instead, geographic location and management emerged as the most important predictors for classification. Conclusions: Our findings highlight the challenges of linking ecological field data with remote sensing information for biodiversity assessments. Improved integration of ecological and remote sensing data is necessary to enhance the effectiveness of biodiversity monitoring schemes.