Remote Sensing of Grassland Biodiversity

Hyperspectral Remote Sensing of Grassland Biodiversity

Hyperspectral instruments are most typically mounted on crewed aircraft for most grassland studies. Some satellites, such as Hyperion, have carried hyperspectral sensors, while in recent years a growing number of studies have made use of UAVs for carrying hyperspectral instruments. Each approach has drawbacks in terms of spatial resolution, spatial coverage, positioning and accuracy. A good agreement was found between airborne (crewed aircraft) hyperspectral data, spatial resolution and species turnover (β-diversity) in a highland savanna (Rocchini et al., 2010). However, spectral diversity in the smaller sampling areas (100 m2) produced a less robust correlation with β-diversity than in the larger sampling area, 1,000 m2, probably due to increased noise when using smaller grain sizes. Focusing on α-diversity (richness and Shannon’s Diversity index), Wang et al. (2016) found a strong positive relationship between biodiversity and productivity, and between optical diversity and species diversity for a Canadian prairie using airborne hyperspectral data, ground sampling and eddy covariance measurements. Comparing spectral diversity, recorded via a tram system, with a range of biodiversity metrics at an experimental prairie test site indicated that a resolution of 1 to 10 cm is best, and spectral diversity correlated differently with different biodiversity metrics (Wang et al., 2018). Spatial resolution also affects the confounding impact of bare soil on the correlation between remote sensing measures of species richness, and, depending on the resolution, different methods to account for bare soil need to be applied (Gholizadeh et al., 2018). Gholizadeh et al. (2019) assessed α-diversity in restored grassland plots in Nebraska that had been seeded with native prairie grasses, some of which were old and contained invasive species while others were younger and mainly contained grasses of the original study design. Ground based (quadrats) and airborne hyperspectral surveys were used. In young plots, spectral diversity was strongly related to α-diversity, but in old plots the relationship was not significant. The relationship between hyperspectral measures of biodiversity and ground surveys varied from one year to the next and weakened over the growing season – emphasising the need for multitemporal measures of grassland biodiversity (Gholizadeh et al., 2020). Lyu et al. (2020) combined a handheld spectrometer, Hyperion and Landsat data for species mapping to assess grassland degradation in Mongolia. This was achieved by comparing the relative proportions of different grass species, extracted from Hyperion data, with ground surveys and metrics derived from Landsat imagery, such as above ground biomass (AGB), net primary productivity and vegetation coverage. By focusing on the identification of typical and indicator species only, classification accuracies of over 70% were achieved. Utilising a UAV mounted hyperspectral sensor, Yang and Du (2021) classified plant species in a desert steppe ecosystem in inner Mongolia. They used a large variety of vegetation indices and decision tree classification to detect plant species with 87% accuracy. In an Alpine steppe nature reserve, Xu et al. (2022) compared four spectral metrics from a UAV mounted spectrometer with two species diversity indices (species richness and the Shannon–Wiener index). The authors found that the relationships between spectral diversity and species diversity were significantly strengthened when bare soil was filtered from the survey data. Finally, a meta-analysis of grassland biodiversity predictions from spectral diversity metrics found an overall correlation coefficient of r = 0.36 across studies (Thornley et al., 2023). The authors noted high levels of variability both within and between studies, with leaf spectra producing a stronger relationship than overall canopy spectra. Surveys of arid, tropical and southern hemisphere sites were lacking, and more scalable and multitemporal studies are required to reduce the uncertainty in the SV/biodiversity correlations.

Multispectral Remote Sensing of Grassland Biodiversity

RGB cameras capture light across the visible wavelengths, while multispectral instruments typically capture a few additional discrete wavelengths of light in the near and shortwave infrared range, meaning that less information can be derived regarding the surface under observation when compared with hyperspectral instruments. However, they are in use on many more satellites resulting in a multi-decadal record of observations on platforms such as the Landsat series. Furthermore, they are cheaper and smaller than hyperspectral cameras, making them more accessible and suitable to consumer grade UAV use. While examining the degradation of grasslands in South Africa, Mansour et al. (2015) employed field sampling, SPOT 5 data and random forest machine learning classification. Indicator species were identified and used to assess the level of degradation. The identification of indicator species was improved from 75 % to 89% when ground-based edaphic measurements were integrated with SPOT 5 data. Lu and He (2017) used a UAV with a near infrared (NIR)-GB camera to map tall grassland species in southern Canada at 5 cm resolution. The authors achieved an accuracy of 85% overall (averaged across all dates and species surveyed) using object-based classification, but suggest that this accuracy can be improved with more precise instruments and a greater number of spectral bands. Across 200 sites in southwest France, Lopes et al. (2017) used a timeseries of SPOT 5 derived NDVI over 18 dates, and a combination of the eight Sentinel 2 bands across eight dates to predict biodiversity. Results, compared to the Shannon and Simpson indices were poor, suggesting that high temporal resolution, moderate-high spatial resolution and multispectral data are not suited to mapping biodiversity at the grassland scale. In Alpine meadows on the Tibetan Plateau, Sun et al. (2018) flew a UAV at just 2 m elevation for highly detailed imagery. Plant species in each image were manually identified and compared to ground-based quadrat surveys and with a range of species composition indices. The UAV surveys proved highly effective compared with indices derived from traditional methods (r2 values between 0.726 and 0.872), while also covering a larger, more representative area and measuring a greater number of species. Shoko et al. (2020) attempted to differentiate between C3 and C4 species (Festuca costata and Themeda triandra ) in South Africa using multidate Sentinel 2 data. The authors achieved greater accuracy in winter (between 91.8% and 95.3%), than summer (between 81.4% and 90.3%). Using Sentinel 1 (Synthetic Aperture Radar) and Sentinel 2, in combination with ground surveys, Fauvel et al. (2020) measured and predicted plant diversity metrics in terms of richness indices, diversity indices and some functional indices in grasslands in southwest France. The methods used worked better for Simpson and Shannon indices than richness indices, and moderately well for functional indices. Incorporating Sentinel 1 data did not significantly improve predictions. In Conti et al. (2021) the authors assessed the links between spectral and taxonomic diversity, and vertical complexity, using a UAV mounted multispectral sensor in a mesic meadow in South Bohemia, Czech Republic. It was found that the relationship between spectral and taxonomic diversity was mediated by grassland vertical complexity - the more pronounced the vertical complexity, the more negative the relationship between taxonomic and spectral diversity. On Alpine grasslands, Rossi et al. (2022) flew a UAV with a consumer grade camera and combined the imagery with airborne hyperspectral surveys. The fused data set tested the effects of spatial resolution and a variety of spectral metrics on measuring species diversity. The authors found the fused dataset worked well but also produced a surprising finding – that spectral metrics centred on spectral complexity was negatively correlated with species richness. The authors, note that the presence of live and dead biomass acted as significant confounding variables in their correlations. On a semi-natural meadow in Saxony, Germany, Pöttker et al. (2023) achieved accuracies of up to 88% by using ground surveys and a multispectral UAV in combination with convoluted neural networks to map plant communities. Yang et al. (2023) combined ground sampling, environmental data and MODIS imagery to predict biodiversity and AGB in the Qinghai–Tibet Plateau grasslands using a random forest model, achieving and r2 of 0.60 for their plant species diversity model.