Evaluating remote sensing data as a tool to minimize spatial autocorrelation in in-situ vegetation sampling

The distinction between geographical patterns caused by underlying environmental factors and inherent spatial autocorrelation is a general challenge for field research. The quality and validity of phytogeographical studies is strongly dependent on disentangling spatial and ecological proximity. This is also crucial for applied studies in nature conservation. One key assumption for many statistical analyses is the independence of observations. In this study we first identify the range of spatial autocorrelation in managed grasslands based on field data. Along a gradient in a valley bottom, we set up five 60 m x 60 m squares, segmented in 36 10 m x 10 m square grid cells. In 20 of the 10 m x 10 m grid cells, we sampled vegetation along a 10 m line with a buffer of one meter resulting in a 20 m² sampling plot. In a second step, we matched Sentinel-2 images for the same locations and calculated the normalized difference vegetation index NDVI and the normalized difference red edge index NDRE. For both, field data and satellite data, Mantel correlograms for floristic distances and spectral indices were used to analyse the spatial autocorrelation. We found the vegetation in the studied grasslands to be spatially correlated up to 25 m. At none of the studied sites the positive spatial autocorrelation reaches beyond. The spatial autocorrelation of spectral indices correlates well with the correlations observed field data. The correlograms of NDVI resembled the ones of the field data slightly better compared to the correlograms of NDRE and RGB. We conclude that employing remote sensing to assess the role of spatial autocorrelation for grasslands is a valid approach. We show that it reflects similar patterns as the field data. The spatial resolution of freely available satellite data proved sufficient to test for the minimum distance between vegetation samples to avoid spatial autocorrelation.