Functional Traits

Leaf Dry Matter Content

Over the Tibetan Plateau, Li et al. (2018b) attempted to measure plant dry matter content with satellite imagery and use this as a proxy measure for the community weighted mean (CWM) of LDM content. However, the correlation was weak, with an R2 of just 0.1. A different approach to mapping the CWM LDM content was conducted by Polley et al. (2020a) at a restored grasslands site in Texas. Four years of hyperspectral reflectance measurements from the ground and UAVs, as well as ground sampling were incorporated into a partial least squares regression (PLSR) analysis, with the model explaining 73% of the LDM content of their canopies. Three further studies used the same analysis and LDM content data as Polley et al. (2020a) for different goals, such as looking what regulates the temporal stability of grassland metacommunities (Polley et al., 2020b), biomass production (Polley, Collins and Fay, 2020) and the influence of community LDM content on plant production (Polley, Collins and Fay, 2022). Returning to the Tibetan Plateau, Zhang et al. (2022) used UAV based hyperspectral imagery and ground sampling, to map community LDM content through numerous different machine learning models. The generic algorithm integrated with the PLSR performed best for LDM content, explaining just 30% of the variance.

Crude Protein

UAV and satellite data combined with handheld hyperspectral and ground sampling were successfully used to assess grass quality, such as CP, under varying soil management conditions in a grassland research site in Ireland. It was found that UAV data with multi-linear regression models worked best and performed better than satellite data (Askari et al., 2019). At natural steppe grasslands in China, Gao et al. (2019) used a multispectral UAV and a variety of vegetation and band indices to map feed quality, achieving the best results using the MERIS terrestrial chlorophyll index. In a Colombian grazed grassland, Giraldo et al. (2023) achieved an R2 of 0.76 using multispectral UAV VIs and a generalised additive model. However, Hart et al. (2020) failed to achieve good accuracy with the multispectral UAV over commercial grasslands in Switzerland. The authors blamed the open access model they used, GrassQ, being calibrated on a different type of grassland. In south-east Germany, Raab et al., (2020) used both Sentinel-1 and -2 and RFs to predict CP. While a strong relationship was found, the authors found that the benefit of the additional Sentinel-1 data inclusion was minimal. Using MODIS derived NDVI values and RFs over Tibet, Han et al (2022) achieved R2 values of over 0.9 for CP. Using ground based hyperspectral measurements in combination with Sentinel-2 data, Zhao et al. (2023) mapped CP across Inner Mongolian grasslands with an R2 0.77 using RFs regression. Across three different farmlands in western Colombia, Zwick et al. (2024) used Planetscope imagery and ground sampling over three years to try and model nutrient quality with machine learning. No single machine learning model worked best overall, with their accuracy varying depending on the location, but the best results ranged between an R2value of 0.52 and 0.75, and RF model variants achieving the best CP accuracy for two of the three locations. With a UAV mounted hyperspectral camera, forage quality was assessed over grasslands in central Germany with several different statistical and machine learning models. Support vector regression predicted CP most accurately, with a high R2 of 0.81 under cross validation (Wijesingha et al., 2020). Forage quality, including CP, was also mapped using UAV hyperspectral imagery over grasslands in northeast Australia, in combination with SfM models for grass height and biomass (Barnetson et al., 2020). The authors found that the simple ratio, NIR/Red produced the strongest relationship. At experimental grassland sites in Norway, Geipel et al. (2021) assessed UAV mounted hyperspectral mapping for forage yield and quality with powered partial least squares regression modelling, achieving an R2 of 0.71 for CP.

Potassium, Phosphorous and Nitrogen

Of the 24 research (two were reviews) papers that made up the review of Potassium (K), Phosphorous (P) and Nitrogen (N):
Due to the overlap in papers measuring these plant nutrients, K, P and N have been grouped together, and then split in multispectral remote sensing methods, and hyperspectral methods.

Multispectral

In the alpine grasslands of the Tibetan Plateau Tang et al. (2021) used a UAV with a standard high-resolution camera and PLSR to map a number of different plant traits. Despite achieving strong correlations with most traits, they failed to establish a significant relationship with N content. In contrast, Oliveira et al. (2022) also used a standard drone mounted camera, in combination with four different convoluted neural network models in Finland, eventually achieving an R2of 0.82 with N concentration. In a legume-grass experimental site in Germany, Grüner et al., (2021) combined terrestrial laser scanning, multispectral UAV and RFs modelling to predict biomass and N fixation, achieving an R2 of 0.71 when combining UAV data with the laser scanner. Lussem et al. (2022) combined UAV mounted regular and multispectral cameras, with a variety of machine learning models for nitrogen uptake in west German grasslands, with RF and support vector machine learning achieving R2 values of 0.83. In southern Germany, a multispectral UAV and machine learning models were again used to map N concentration in alpine grasslands (Schucknecht et al., 2022). Most models produced poor correlations, with the maximum R2 achieved with RF (0.47). Several studies have employed Sentinel-2 imagery to map N and P across parts of China. Gao et al. (2020) achieved an R2 of 0.49 in July, and 0.59 in November by using RFs to map the N/P ratio over the Tibetan Plateau. In Inner Mongolia, Pang et al. (2022) enhanced Sentinel-2 imagery with ground based hyperspectral imaging before combining this with meteorological and geographic data. A fractional differential algorithm was used to extract the spectral information related to N and P, and a PLSR model used for estimating their contents This approach achieved an R2 of 0.85 for P, and 0.78 for N. In the mapping of N, P and K over the Tibetan Plateau, Zhang et al. (2023b) combined Sentinel-2 with Tiangong-2 imagery with SVM and RF models. While the results were strong for the individual satellite and modelling methods, combining both with RF modelling produced R2 values of 0.78, 0.74 and 0.84 for N, P and K respectively. In assessing grassland P in the Tibetan Plateau, Shi et al. (2024) used an approach based on graph theory to create hyperspectral data from Sentinel-2 bands, they then used a deep regression inversion model to map grassland P content across different phenological stages. The authors report R2 values above 0.8 and significant improvements over the original low spectral resolution data and other modelling approaches. Sentinel-2 has also been used in various other regions to successful map essential plant nutrients. Arogoundade et al. (2023) successfully mapped the C:N ratio in South African grasslands with Sentinel-2 and RFs, entirely within Google Earth Engine. Across a range of grassland sites in Portugal, Morais et al (2023) assessed the ability of Sentinel-2 to map N and P through machine learning methods. RF again worked best overall, with an R2 of 0.77 and 0.71 for N and P, respectively. Similarly, Cisneros et al. (2020) used the Three Band Index from Sentinel-2 to map foliar nitrogen content in an experimental plot in Brazil with 38% accuracy and Smith et al. (2023) applied Sentinel-2 and a range of machine learning methods to map nitrogen concentrations in a Bahia grass experimental site. However, here RFs produced a very strong R2 in the training dataset (0.99-1.00) but performed relatively poorly in the test data (0.20-0.57). Finally, Dehghan-Shoar et al. (2023) combined a radiative transfer model with a bidirectional reflectance distribution function into a single model to predict grassland N concentration from Landsat-7 and -8, and Sentinel-2, reaching an R2 of 0.50 with their validation dataset.

Hyperspectral

Over West African Savanna, Ferner et al. (2021) attempted to map phosphorous concentration from a ground-based spectrometer, Hyperion hyperspectral satellite imagery, and Sentinel-2. However, no significant correlations were established with any of the datasets. In experimental grassland sites in the US, both Wang et al. (2019) and Cavender-Bares et al. (2022) mapped foliar N content with similar accuracy from aircraft mounted hyperspectral cameras, with R2 values of 0.57 and 0.58 respectively. Gholizadeh et al. (2022a) produced similar predictive power for both N and K, with their own aircraft mounted hyperspectral imaging data. However, in tallgrass prairie sites, Pau et al. (2022) found the N concentration product of the National Ecological Observatory Network’s (NEON) surveys had an R2 of just 0.29 compared to ground sampling, Using UAV based hyperspectral surveys, Polley et al., (2023) achieved an R2 of 0.8 with a simple linear regression between the red-edge chlorophyll index and community N content in experimental grassland in Texas. In an experimental grassland in Finland, MLR and RF were used to combine UAV based hyperspectral images and photogrammetry for N concentration, with an R2 value of 0.90 (Oliveira et al., 2020). UAV based hyperspectral imagery had mixed predictive power for N when combined with PLSR on both a German grassland experimental site, R2 of 0.58 for content (Franceschini et al., 2022) and in an Inner Mongolian monoculture test site with an R2 for N and P of 0.87 and 0.54, respectively (Zhao et al., 2021a). Slightly more modest results were achieved with UAV hyperspectral data and a GA-PLSR model over natural grassland on the Tibetan Plateau, with an R2 of 0.50 and 0.54 for community level N and P (Zhang et al., 2022).

Leaf Area Index

The LAI papers total 35, excluding reviews. As only eight deal primarily with hyperspectral sensors, it makes more sense to divide these by the spatial resolution of the sensors:
  1. High Resolution e.g., UAV and Aircraft (0.01 to 1.0 m).
  2. Medium Resolution e.g., Goafen-2, Landsat, Senitnel-2 (3 to 30 m).
  3. Low Resolution e.g., MODIS and Senitnel-3 (>30 m).

High Resolution

UAVs equipped with hyperspectral cameras have been used to measure LAI across 4 studies in Inner Mongolia since 2019. The best results were achieved on a grassland monoculture site with an R2 of 0.87 between UAV level canopy measurements and ground sampling through PLSR (Zhao et al., 2021a). Using linear regression to relate UAV derived VIs to LAI, Sha et al. (2019) produced an R2 of 0.45 between the Generalized soil-adjusted vegetation index and LAI, with more of the errors coming from regions with low LAI values. However, Zhu et al., (2023) used the PROSAIL model to determine the optimum VIs, then used a two-layer VI matrix to calculate LAI, with an R2 of 0.73. Zhu et al. (2024) developed this further over a species rich grassland by using the PROSAIL model and two simple vegetation indices, the optimized soil-adjusted vegetation index (OSAVI) and NDVI, achieving an R2 of 0.84. Two additional studies were carried out focusing on aircraft mounted spectrometers, with contrasting results. In a Tallgrass site in the USA, the NEON LAI was not significantly related to ground-based LAI measurements (Pau et al., 2022). With a different approach, Bandopadhyay et al. (2019) found higher rates of sun-induced fluorescence at 687 and 760 nm was associated with greater LAI (R2 of 0.80 and 0.86 respectively) in their natural test sites in Poland, which included many species rich grasslands. The sun-induced fluorescence measures also correlated well with greenness related Vis, such as NDVI.

Medium Resolution

A range of methods have been employed under the medium resolution remote sensing of LAI. Xu et al. (2018) and Qin et al. (2021) compared different VIs to ground based LAI measurements, with the perpendicular vegetation index from Landsat and the normalized difference phenology index from Sentinel-2 performing best, respectively. In two studies using the Copernicus Land Monitoring Service (CLMS) LAI products and Sentinel-2 in Poland, Dabrowska-Zielinska et al. (2024) and (Panek-Chwastyk et al., 2024) found strong agreements with ground-based LAI, with R2 values of between 0.62 and 0.93. Machine learning was also the focus of several LAI studies. In South Africa, RF has been used to successfully map LAI with both Landsat and Sentinel-2, but with slightly stronger results in the dry season vs the wet season (Dube, et al., 2019; Masenyama et al., 2023). In more mountainous South African grasslands, Tsele, Ramoelo and Mcebsi (2023) found that the optimal regression choice, RF or stepwise multiple linear regression, varied depending on the location. Shen et al. (2022) assessed a range of different machine learning approaches (RF, neural networks and support vector regression) on Landsat-8 data to model LAI, with RFs again tending to produce the most accurate results. Three studies have attempted to use machine learning methods to integrate SAR data with multispectral for mapping LAI. Lu and He (2019) found the improvements from including SAR in their RF over the southern Canadian Prairies was marginal. However, Wang at al. (2019) found that SAR data improved LAI estimates in their MLR model over areas of dense tallgrass vegetation where typical VIs tend to become saturated, a finding supported by a subsequent study of Alpine grasslands in northern Italy (Castelli et al., 2023). Five studies have used radiative transfer models (RTMs) with medium resolution satellite imagery to aid in mapping LAI. In test farms in southern England, the PROSAIL model was used with Sentinel-2 for LAI mapping, achieving strong correlations and offering an improvement over LAI calculated from NDVI (Punalekar et al., 2018). In Brazil, the Automated Radiative Transfer Model Operator (ARTMO) was used with Sentinel-2 also. The authors found that the Normalized Area Over Reflectance Curve (NAOC) index produced the strongest results (Cisneros et al., 2020). In Austria, Sentinel-2 was used with two RTMs for the growing seasons of 2018 and 2019, achieving an R2 of 0.87 with direct ground measurements of LAI (Klingler et al., 2020). Similar success was achieved in grassland of northern China using the PROSAIL model again, with an R2 of 0.82 between the newly developed Chlorophyll-Insensitive VI (CIVI) and LAI (Zhang et al., 2023A). In northeastern Germany, Schwieder et al. (2020) tested the accuracy of two methods for assessing LAI using Sentinel-2 – RF regression and a soil-leaf-canopy (SLC) RTM, with both models demonstrating strong predictive power. Brown et al. (2021) compared a novel Level 2 processor for Sentinel-2 data (SLP2-D), with updated artificial neural networks retrieval methodology. The updated method was close to or better than the old over many vegetation types, but slightly worse over grasslands. Jiang et al. (2024) developed a new Bi-directional Reflectance Distribution Function (BRDF) for the Gaofen-1 satellite to improve vegetation parameter accuracy with tests over grasslands in northeast China. The new BRDF produced an R2 of 0.58, 0.14 higher than the previous method. In central China, Peng et al. (2024) applied topographic corrections to a large range of LAI models, and compared them to LAI products, such as from MODIS and GLASS, and ground sampling. Topographic corrections, when combined with RTMs, improves the correlations (R2improvements of 0.18 to 0.04) and reduces the errors more than ML combined with RTMs. They also produced an R2improvement of >0.2 compared to MODIS and GLASS LAI products. The research is focused on mountainous terrain and so may not be as applicable in flatter grasslands.

Low Resolution

Many low-resolution global LAI products currently exist and have been used in a large range of studies over recent years. These products include, for example, the MODIS derived MOD15A2 and MOD15A2h, Geoland2 Version 1 (GEOV1) and Global Land Surface Satellite (GLASS), each with different development methods, temporal and spatial resolutions. Several recent studies have compared these LAI products with ground measurements and high-resolution satellite data across different ecosystems and countries. Li et al. (2018a) compared MOD15A2, GLASS, Global LAI Product of Beijing Normal University (GLOBALBNU) and Global LAI Map of Chinese Academy of Sciences (GLOBMAP), with ground measurements both across the globe and, with special emphasis, over China. Overall, GLASS performed best in both situations, with R2 values of 0.70 and 0.94 respectively. Even though grasslands made up 43.1% of the assessment area in China, specific correlations for grasslands are not provided. Another comparison was carried out by Liu et al. (2018) between the MOD15A2, GLASS and the Four-Scale Geometric Optical Model (FSGOM), over a mixture of land cover types in China. FSGOM was found to perform slightly better in grasslands, with and R2 of 0.5, 0.09 and 0.22 better than MOD15A2 and GLASS respectively. A specific grassland comparison of GEOV2, GLASS, GLOBMAP, and MOD15A2h was carried out in Inner Mongolia, with GLOBMAP performing best in meadows, GLASS best in typical steppe and GEOV2 best in desert steppe, but all with R2values below 0.4 (Shen et al., 2023). Yin et al. (2020) demonstrated that the temporal resolution is also important to consider. Comparing MOD15A2, MOD15A2h, GEOV1 and GLASS, the authors found that the MODIS based LAI products had lower R2 compared to the other datasets, but the shorter temporal window allowed for sudden changes to be detected, while GEOV1 and GLASS had high R2 values, but missed grazing induced sudden changes due to their broad temporal windows. Munier et al. (2018) used Kalman filtering to disaggregate global GEOV1 data, allowing them to assign different LAI values to different vegetation types within a single pixel. While producing improvements over most vegetation types, this method reduced the accuracy over grasslands, with the R2 dropping from 0.89 to 0.82 compared to the original data. Several attempts have been made to fuse high resolution, but temporally sparse, LAI data with low resolution global products, but with mixed results. Li et al. (2018v) used Landsat-7, -8 and Sentinel-2 to generate 30 m resolution LAI maps in northern China using PROSAIL. These were combined with the MODIS data using a spatial and temporal adaptive reflectance fusion model (STARFM). The authors note reductions in errors and noise in their new fused datasets, with the R2 of 0.62 vs 0.53 for the original MODIS LAI product. Zhou et al. (2020) took a different approach, using a timeseries of MOD15A2H as a long-term background signal, ground measurements and Landsat-7 and -8 were fused using a back propagating neural network to create 30 m LAI maps for the study regions in Ukraine and China. A modified ensemble Kalman filter model (MEnKF) using both the Landsat and MODIS data, allowed for the 30 m LAI to be spread over the space and time of the MODIS data, achieving an R2of 0.88 over grasslands. Across mixed test sites in the USA, another approach to combining field data, MODIS and Landsat through a deep transfer learning framework failed to produce substantial improvements over grasslands but was successful over croplands and forests (Zhou et al., 2023). Finally, in northern China, Li et al. (2024) compared the ability of Sentinel-3 to retrieve vegetation parameters such as LAI, with MODIS and PROBA-V, using a range of prediction methods. Sentinel-3 had better accuracy than other platforms, likely due to the red edge bands, but only slightly compare to PROBA-V (R2 of 0.63 each)