Assessing Spectral Metrics to Estimate Aquatic Plant Diversity from Hyperspectral Imaging: Preliminary Results

For their morpho-physiological peculiarities and wide phenotypic plasticity, aquatic plants occupy the extremes of the global spectrum of vegetation forms; such heterogeneity results in them displaying contrasting patterns of diversity along ecological and geographical gradients. In the last decade, remote sensing has opened new ways of measuring biodiversity; in particular, high-throughput imaging spectroscopy is a feasible, efficient option for assessing plant diversity based on spectral proxies directly related to morphological and biochemical traits, which we define as spectro-functional traits. Linking spectral features to plant species diversity to characterize plant communities can further advance this topic. In this study, we explored the use of spectral features extracted from centimetric resolution hyperspectral images collected from a drone to estimate taxonomic diversity via generalized additive models (GAMs) within communities of floating hydrophytes and helophytes sampled across a trophic gradient. Hyperspectral images were acquired during drone flights over target aquatic plant communities during the summer of 2021 in five wetland sites in Italy, using a Nano-Hyperspec® (Headwall Photonics) sensor (400-1000 nm), with a nominal spatial resolution around 3 cm pixel size. Spectral diversity features were derived from hyperspectral data cubes as spectro-functional traits (SFTs) and spectral Principal Components (sPCs). Four SFTs were computed: three normalized difference proxies of aquatic plant leaf biochemical parameters and the Water Adjusted Vegetation Index (WAVI), a spectral index specifically developed for its sensitivity to aquatic vegetation greenness. As additional spectral diversity features, the first five components (PC1-5) were computed from all hyperspectral bands over green material pixels. As a last step, mean, standard deviation and distance from the multidimensional centroid were calculated over all green pixels falling within a circular area of 4 m radius centered on in situ surveyed plant communities. Alongside drone flights, in situ data were collected from boat-based surveys, covering species abundance and coverage across 89 target plant communities, where diversity was calculated as effective number of species (2D) and species evenness (J'). After the exclusion of mutually correlated ones within each group (SFTs and SPCs), spectral metrics were used as input to GAMs for predicting species diversity across all 89 target communities, as 2D or J' (log-transformed for reducing distribution skewness), without considering interactions and using site information as random effect. Our results show that spectral features better predict 2D (R2 up to 0.75) than J' (R2 up to 0.55), especially when SFTs are employed.