Systematic Sensitivity Analysis of Spectral Variation Hypothesis Based Spectral Diversity Metrics from Sentinel-2 Using BiodivMapR

The Spectral Variation Hypothesis (SVH) proposes that spectral heterogeneity derived from remote sensing data can serve as a proxy for biodiversity. While spectral diversity metrics are increasingly applied in ecological studies, their reproducibility remains limited by methodological choices that are rarely evaluated systematically. In particular, the sensitivity of α-spectral diversity to clustering, spatial scale, masking strategies, and spectral input configuration within commonly used workflows such as BiodivMapR is poorly understood.

This study presents a systematic sensitivity analysis of α-spectral diversity derived from Sentinel-2 imagery using the BiodivMapR framework over a selected region in Estonia. The primary objective is to identify cluster size thresholds at which diversity values stabilize and to assess whether this stabilization depends on image extent. K-means clustering was evaluated across cluster sizes of 20, 40, 60, 100, 200, 500, and 1000, combined with image extents of 10, 15, 20, and 30 km. For all configurations, the same number of samples was used to derive clusters, while BiodivMapR’s default random initialization was retained to assess robustness. 

The analysis was extended to examine ecosystem masking effects. Forest-only masking represents landscape-level diversity restricted to forest ecosystems, while an inward buffer (~15 m) was applied to exclude edge pixels influenced by roads and non-vegetated surfaces, isolating within-forest spectral heterogeneity. Additional experiments assessed the sensitivity of α-diversity to window size and spectral input choice (bands versus indices). 

Results demonstrate that α-spectral diversity is highly sensitive to methodological configuration, with stabilization thresholds varying across spatial extents and masking strategies. Even within forest-only analyses, edge effects significantly influence diversity estimates. These findings highlight that spectral diversity metrics are strongly parameter-dependent and cannot be directly compared across studies without methodological harmonization.