Disentangling vegetation cover and structure influence on spectral diversity: insights from a study in North-Eastern Italian forests.

The estimation of plant taxonomic diversity (TD) based on spectral diversity (SD) in reflectance can be achieved through the so-called Spectral Variation Hypothesis (SVH). However, studies conducted in various habitats and using diverse sensors, methodologies and diversity indices have presented conflicting relationships between SD and TD. Several factors, including spatial and spectral scales, environmental characteristics of the study area, vegetation cover and structure, and computation methods can influence the SD. In this study, we tested the SVH hypothesis within a heterogeneous forest area in the north-east of Italy, employing Sentinel- 2 data. Our aim was to identify the factors that impact the strength and direction of the relationship between SD and TD. We determined the SD using the frameworks provided by the "biodivMapR" (BD) and "rasterdiv" (RD) R packages, considering 38 combinations of SD indices, at both the α (within plots) and β (among plots) levels, and computational parameters. These last accounted for chosen pixels extraction area, window size and spectral dimension mode. Information on vegetation structure was obtained from ground-based and LiDAR data. Random Forests were employed to analyse the relationships between SD, TD and vegetation structure, and to determine the optimal combination of SD computational parameters. At the α level, we observed negative associations between TD and SD indices derived from RD. The variation in RD SD was mainly influenced by gaps in the forest canopy, leading to higher values in plots dominated by conifers with low canopy cover and low TD. Conversely, the BD algorithm reduced the influence of background on SD and was able to differentiate between major forest types (broadleaves vs conifers). However, α level SD indices were not correlated with TD. At the β level, we found a significant positive correlation between β SD and TD, with a maximum correlation coefficient of 0.24. Additionally, we discovered that smaller computation windows and larger areas of pixel extraction for spectral species classification resulted in stronger correlations and R2 values. Our findings suggest that vegetation cover and structure have a greater influence than inter-species spectral differences in determining forest α SD. In this sense, SD might be better suited for capturing differences in species composition at the landscape level rather than the richness of individual communities.