Figure 3: The top panel shows aerial views depicting beta diversity in a diverse landscape (A) and a low-diversity landscape (B), and alpha diversity in and a highly diverse grassland (C) and a low-diversity grassland (D). The second panel shows the spectral signatures associated with the coloured boxes on the top panel. The bottom panel shows the corresponding spectral diversity in the green, red, red-edge and near-infrared wavelengths, from left to right, respectively.

Spectral Variation Hypothesis

The SVH has been employed across a broad range of spatial scales, yet there appears to be little consistency regarding the ideal spatial resolution at which it best operates. At an experimental prairie site, Wang et al. (2018) found the idea pixel size to be between 1 mm and 10 cm to establish a strong relationship between species richness and spectral diversity, with the relationship fading after 10 cm. Similarly, Polley et al. (2019) suggests that the sensitivity of SD to species diversity is scale dependent and that the necessary spectral details may be lost with greater spatial scales. Gholizadeh et al. (2019) found a strong relationship between SD and species richness at both 0.5 m and 1 m resolution, but not at 5 m. However, Gholizadeh et al. (2022a) failed to connect SD to species richness at both 1 m and 30 m, only achieving significant correlations with the Simpsons diversity index. Similarly, no strong correlations were found between species richness and a large range of standard SD metrics at 2 cm and 5 cm pixel sizes from multispectral UAV surveys (Perrone et al., 2024). The authors suggest that phenological stage plays a key role, and finer resolutions can be more noisy, due to factors like shadowing. Jackson et al., (2022) used a multispectral UAV to estimate biodiversity at 0.1-0.5 cm resolution. They could predict the Shannon–Weiner and Simpson’s biodiversity indices well, but not species richness. They note that the measure of SD decreased by 30% with every additional 1 m in elevation that the UAV was flown.
Confounding factors that influence how well SD is related to biodiversity have been explored on a variety of grasslands in recent years. Yang et al. (2023b) noted how an open-pit coal mine influenced α-, β-, and γ- diversity in an Inner Mongolian steppe, and that grazing increased the area that was most strongly affected. Using Landsat data for Himalayan grasslands, Chitale et al. (2019) explained that variance in species richness measured using vegetation indices increased from 54 to 85 % after the inclusion of physiographic indices. In more general terms, it has been found that accounting for the effects of bare soil on the spectral readings can significantly improve remote sensing estimations of both α- and β- diversity (Kamaraj et al., 2024; Xu et al., 2022). In contrast to many other studies, Conti et al. (2021) found a negative relationship between SD and taxonomic diversity in mesic meadows in Czechia, with the vertical complexity driving the relationship – the more vertically complex the grass structure, the more negative the relationships between SD and taxonomic diversity. In addition, the timing of flowering plants (Perrone et al., 2024), the presence of non-native species (Van Cleemput et al., 2023) and the proportion of live and dead biomass (Rossi et al., 2022), have all now been identified as confounding factors.
Several studies have explored the use of time series analysis to overcome uncertainties in the SD/biodiversity relationship. Rossi et al. (2021) introduced a spatio-temporal version of Rao’s quadratic entropy index (RaoQ) to examine changes in β-diversity over time with Sentinel-2 imagery and account for variations in grassland management and phenology. They suggest that, with higher resolution data, that this method can be applied to α-diversity too. In exploring how α- and β-diversity vary over two years in prairie grassland, Gholizadeh et al. (2020) found significant differences in species richness, due to factors such as fires and weather, and recommend multi-temporal surveys to account for these changes. When assessing grasslands in the USA and Europe, Rossi et al. (2024), with a Sentinel-2 timeseries, found a stronger and more consistent relationship with species diversity from temporal SD than spatial SD, suggesting that an analysis based on a single snapshot in time can be misleading.
Some new approaches have also attempted to tackle these uncertainties. Zhao et al. (2021b) used cluster analysis of hyperspectral data to identify distinct spectral species, allowing them to accurately predict plant species diversity (R2 of 0.73). Developing this idea further, Rossi and Gholizadeh (2023) used spectral unmixing. They determine number the distinct spectral entities, called endmember, within each image. Then calculate the number of endmembers and their abundance within each pixel and use that information to create endmember spectral diversity metrics. The authors claims that this approach is less sensitive to soil and can also be applied to multi-temporal datasets, which may help to overcome some of the previously identified confounding factors.

Machine Learning

To assess plant species diversity over part of the Tibetan Plateau, Zhao et al. (2022) used high-accuracy surface modelling (HASM), Landsat-8 data and a range of machine learning models, namely least absolute shrinkage and selection operator, ridge regression, eXtreme Gradient Boosting and Random Forest (RF). The authors found that the models combined with HASM performed better than the machine learning models alone, with the best combination being eXtreme Gradient Boosting and HASM, followed closely by RF and HASM. Fauvel et al. (2020) experimented with combining the multispectral data of Sentinel-2 with the radar from Sentinel-1 to map biodiversity in grasslands in southern France with multiple regressions methods - Linear regression, K-Nearest Neighbours, Kernel Ridge Regression, RF and Gaussian Process. They found that RF worked best overall, with R2 values above 0.4 for the Simpson and Shannon indices, and the addition of Sentinel-1 data provided no significant improvements. Another attempt to combine Sentinel-1 and -2 came from Muro et al. (2022) with RF and deep neural networks employed to predict biodiversity. The deep neural networks model performed slightly better than RF, though both performed poorly under cross validation, and the addition of Sentinel-1 again provided little benefit. Several other studies found success with mapping plant species diversity using different forms of neural networks. In semi natural meadows in Germany, convoluted neural networks were used to classify multispectral UAV data, mapping vegetation units with accuracies of up to 88% (Pöttker et al., 2023). In three distinct German grasslands, a residual neural network model was used with a time series of Sentinel-2 data to map a range of plant biodiversity metrics, achieving R2 values of up to 0.68 and showed significant improvements in accuracy compared other machine learning methods assessed (Dieste et al., 2024). Employing convoluted neural networks in a different way, Gallman et al. (2022) managed to identify and count individual flowers from images taken by drone mounted standard high-resolution camera, performing as well or better than manual counting for most flower species.
However, RF tended to produce the most accurate results for the majority of studies. Using weather data and MODIS based Normalised Difference Vegetation Index (NDVI) over Tibet, Tian and Fu (2022) found RF to produce the most accurate measures of plant species diversity compared to numerous other machine learning methods. Again, over the Tibetan Plateau and using MODIS data (and weather, soil and topographic variables) Yang et al. (2023a) achieved an R2 of 0.6 for plant species diversity with RF after using stepwise regression for variable selection. In mountainous grasslands in South Africa, Mashiane, Ramoelo and Adelabu (2024) used vegetation indices (VIs) from Sentinel-2 and Landsat-8 and RFs to model species richness and the Shannon–Wiener index, achieving r2 values above 0.85 for both. RF modelling was also most accurate compared to other ML methods in the Three Rivers Headwater Region of China, where Yang et al. (2024) used stepwise regression to select among variables from Landsat, climate, soil and topographic data. Indeed, using VIs, canopy height and textural data derived from multispectral UAV surveys over two summers in a wet grassland near Berlin, Bazzo et al. (2024) also found RF modelling to produce the most accurate and consistent measures of species richness. Finally, in an assessment of plant diversity in numerous ecosystems across the world (including 315 grassland plots), Xin et al (2024) again found that RF models produced the most consistent and accurate results compared to other regression and machine learning models.

Other methods

Some researchers have taken slightly different approaches to remotely mapping grassland biodiversity. Löfgren et al. (2018) attempted to use both satellite and UAV based NDVI values to map the richness of specialist species in grasslands on Baltic island in southern Sweden, but achieved only weak, negative correlations. In the alpine grasslands of Tibert, Qin et al. (2020) achieved significant correlations between richness, Shannon, Simpson and Pielou’s indices derived from the manual counting of species from UAV imagery versus traditional quadrat surveys, and identified more species (71) from UAV imagery than from quadrat surveys (63). Another study on the Tibetan Plateau found a significant relationship between UAV measured bare patches in grasslands with decreases in richness and increased species turnover (Hua et al., 2023). In mountainous grasslands of northern Portugal, Monteiro et al. (2021) found that the NIR/Green ratio values from Sentinel-2 and their seasonal amplitude correlated well with species richness, producing an R2 of 0.44. In tallgrass prairies in the USA, Hall and Lara (2022) compared combinations of hyperspectral UAV and LiDAR, then multispectral UAV, phenometric data and Structure from Motion (SfM), and finally RGB-SfM for the mapping of 10 different species, achieving accuracies of 78%, 52%, 45%, respectively. Using a uniquely cross disciplinary approach, Janišová et al. (2024) combined a time series of satellite based NDVI going back to 1984 with ground surveys, history and ethnology for land use change and cultural practises in two villages in the Serbian Carpathian grasslands. By gaining an understanding of the history and culture of the regions, the authors were better informed regarding the historical land management practises, how they’ve changed and the influence this has had on current biodiversity levels. This further enhanced their enhanced their interpretation of the historical NDVI record and allowed authors to make specific recommendation on land use, such as a partial return to historical land management practises to at least partially restore some of the lost species richness.