Discussion and Recommendations

Biodiversity

Recent years have seen significant advancements in access the free, moderate to high-resolution multispectral satellite imagery, alongside the rapid development of UAV technology allowing both multispectral and hyperspectral data to be captures across a huge range of spatial resolutions and scales. However, evidence from the papers covered in this review shows that there are still substantial uncertainties regarding how best to connect the spectral diversity to species diversity, with significant variability in the correlations achieved. This uncertainty appears to exist regardless of whether the measurements occur with hyperspectral or multispectral instruments, regardless of the spatial scale, the spatial resolution, location or type of spectral variation metric tested. This is supported by a 2023 metanalysis that found an average correlation of just 0.36 between spectral variation and species diversity in grasslands, with significantly variability occurring both within and between studies (Thornley et al., 2023). Furthermore, another systematic review of related papers between 2000 and 2022 suggested that more work needs to be done to identify factors that influence the SVH (Lyu et al., 2024). Machine learning algorithms have emerged as new and effective tools for mapping grassland biodiversity but, like the SVH, needs to begin better accounting for sources of error and uncertainty. As such, we present some recommendations regarding the remote sensing of biodiversity:
Finally, integrated long-term surveys in a range of different grassland environments, linking ground data collection, low-elevation aerial observations, high-elevation aerial observations and satellite observations, should be combined with SD analysis and machine learning. This can best account for temporal and spatial scale discrepancies, thereby allowing for the analysis to more effectively identify the most suitable spectral diversity metrics and regression/machine learning tools to model biodiversity.

Functional Traits

Leaf Dry Matter Content

Little progress has been made in the remote sensing of grassland LDM content, although the work of Polley et al. (2020a) demonstrated that a significant correlation with LDM content could be established with UAV based hyperspectral imagery using PLSR. This was backed up by Zhang et al. (2022) using a similar method but only achieved weaker, though still significant, relationship with LDM content. PLSR and hyperspectral remote sensing appear to show promise in mapping LDM content, but this approach is in its early stages and much more work is required. • Initial studies by Polley et al. (2020a) and Zhang et al. (2022) show that UAV based hyperspectral UAV data and PLSR have the potential to predict LDM content, but more work needs to be done to exploit this method.

Crude Protein

Of the 11 studies presented on the remote sensing of CP, eight used primarily multispectral data (from both UAV and satellites) and three used hyperspectral UAV data. Six of those 11 in total incorporated machine learning methods. However, barring one exception (Hart et al., 2020) all studies found strong and significant correlations with CP, often using just simple band ratios and/or VIs with a mix of different regression algorithms. Given the ease of use, low costs and effectiveness of consumer grade multispectral UAVs and free multispectral satellite imagery, the tools and data for mapping of grassland CP content from centimetre to decametre scales are becoming increasingly accessible to a growing range of researchers and land managers.
Crude protein can be effectively measured at from both multi- and hyper-spectral data, at large and small spatial scales and with a range of simple and complex modelling methods. The studies presented here have exclusively focused on local or regional sites. Therefore, the feasibility of scaling these surveys to national or global scales should be assessed in the near future.

Potassium, Phosphorous and Nitrogen

The two studies dealing with Potassium, Zhang et al. (2023) and Gholizadeh et al. (2022a), used different observation platforms and analysis methods but both with strong results. While there has not been enough research on remote sensing of K in grasslands, the two results shown suggest that it is feasible, even with two very different approaches. As such, more work needs to be done to assess the range, consistency and applicability of these measurement tools. For the seven studies that provided estimates of P, the four that used multispectral satellite surveys achieved an average R2value of 0.78, while the three that used hyperspectral imagery (two UAV based, one satellite-based) averaged just 0.4. It’s difficult to infer anything significant given the sparse number of studies, but this result stands in contrast to the review by Van Cleemput et al. (2018), that found an average R2 of 0.75 for the hyperspectral remote sensing of Phosphorous in grass- and shrublands This suggests there are still substantial uncertainties that need to be addressed regarding the remote sensing of P, especially using hyperspectral imagery. The studies measuring Nitrogen produce more consistent results than P, with an average R2 0.63 and 0.62 for multispectral and hyperspectral measurements, respectively. This is more in line with Cleemput et al. (2018), that found an average R2 of 0.74, but that included proximal measurements that are likely to be more accurate. No significant difference arises from the platform used (UAV, Aircraft or Satellite), the regression or modelling approach nor the study location. Some studies did fail establish a significant relationship, such as Tang et al. (2021) using a standard camera on a UAV, or Pau et al. (2022) when assessing the NEON aircraft-based hyperspectral products. Furthermore, no studies made use of hyperspectral satellite data, with all hyperspectral nitrogen remote sensing being based on either UAV or aircraft surveys.

Leaf Area Index

Linking high-resolution, hyperspectral data from aircrafts and UAVs to LAI is a growing area of research. The data coming from these platforms appear to be capable of modelling LAI with high degrees of accuracy, with three of the five studies reviewed having R2values of 0.73 or higher. These studies have achieved success using analysis as basic as linear regression up to RTMs and machine learning, and over both experimental and natural sites. The NEON LAI product tested by Pau et al. (2023) are again the worst performing. However, all the successful studies were performed at just local scales, versus NEON which is more generalised. A wide variety of approaches have been used with medium-resolution, multispectral satellite data. The most successful appears to be those focused on radiative transfer models (n=5), with an average R2 of 0.75, and a range from 0.57 to 0.87. Studies primarily relying on machine learning models (n=8) have generally proven effective too, with an average R2 of 0.67, ranging from 0.46 to 0.87. However, those high and low values come from two studies, Masenyama et al. (2023) and Tsele, Ramoelo and Mcebsi (2023), and both using Sentinel-2 and both based in mountainous regions of South Africa. This hints at the possibility of additional uncertainties being added to surveys in mountainous terrain. Indeed, the work of Peng et al. (2024) showed that applying a topographic correction to Landsat-8 improved the correlations and reduced the errors from both RTM derived and machine learning derived LAI data. Despite three studies performing intercomparisons between global LAI products since 2018, no product performs consistently better than any other, and R2 values vary significantly from one comparison to the next (Li et al., 2018a; Liu et al 2018; Shen et al., 2023). Even products with greater spatial resolution often require a broader temporal window for complete daily coverage, meaning that transient or sharp changes in LAI can be missed (Yin et al., 2020). Attempts by Munier et al (2020) to extract sub-pixel LAI values failed to provide an accuracy improvement over grasslands, despite working for other vegetation types. Recent efforts to fuse finer resolution data from Landsat and Sentinel-2 with global LAI products such as those from MODIS, have produced mixed results thus far over grasslands (Li et al., 2018c; Zhou et al., 2020; Zhou et al., 2023). As such, further development of these fusion models will be necessary before a reliable, global LAI model with both high-spatial and -temporal resolution can be distributed. • LAI mapping from high-resolution hyperspectral surveys can be successful using a variety of regression and modelling approaches, but parameters need to be tuned to specific regions to ensure accuracy. Research integrating hyperspectral satellite measurements should aid in this task. • New research demonstrated that sun induced fluorescence at 687 and 760 nm has a strong association with LAI, potentially opening the door so a new form of high-resolution LAI mapping. • With moderate resolution multispectral data, both RTMs, especially PROSAIL, and machine learning methods, particularly random forests, have produced consistent and robust estimations of LAI. • Moderate resolution LAI estimates may also benefit from topographic corrections in more rugged terrain. • Comparisons of global LAI products have failed to identify a single best option and attempts to fuse high- and low-resolution data are still in development and lack consistency. It is therefore necessary to test a range of products to find the one most suited to the study area in question and with the necessary spatial and temporal resolution.