Remote Sensing of Grassland Functional
Traits
Hyperspectral Remote Sensing of Grassland Functional
Traits
On experimental grassland sites in Germany, Capolupo et al. (2015) used
a UAV mounted hyperspectral camera and ground-based surveys to examine a
range of physical and chemical traits (e.g., height, biomass, crude
protein, nitrogen, potassium, etc). To estimate these traits, the
authors used two methods, partial least squares regression (PLSR) and
vegetation indices. PLSR performed well for both physical and chemical
traits, while vegetation indices worked well only with physical traits.
In a separate experimental barley site in Germany, Aasen et al. (2015)
combined 3D with hyperspectral imagery for precision agriculture. Plant
height, chlorophyll, leaf area index (LAI) and biomass were measured
with r2 values of 0.7, 0.52, 0.32 and 0.29,
respectively.
Airborne hyperspectral surveys were used over Swiss alpine grasslands,
in combination with ground surveys, to examine links between community
traits, functional traits and spectra with PLSR. The authors noted some
inconsistent results with plant life and growth forms, but functional
type modelling produced more consistent results (Schweiger et al.,
2017). In a barley test site in Finland, Näsi et al. (2018) equipped a
UAV with hyperspectral and RGB cameras. The authors combined the digital
surface model and digital terrain model, generated with using SfM, with
the hyperspectral information, estimating dry and fresh matter, and
nitrogen content. The 3D data combined with RGB information was nearly
as accurate as hyperspectral data for biomass, but hyperspectral imaging
performed better for nitrogen content.
Wang et al. (2019) used ground sampling and airborne hyperspectral
surveys to map foliar functional traits across a grassland experimental
site in Minnesota. They used PLSR and gaussian processes regression
(GPR) for modelling, uncertainty and performance. Both regression
methods produced similar results, with leaf mass per area, soluble cell
contents, hemicellulose and cellulose all producing r2values > 0.8, while models for the contents of lignin,
nitrogen and some pigments performed more poorly.
Across eight grassland sites in Northern Germany, Wijesingha et al.
(2020) attempted to predict forage quality by mapping crude protein and
acid detergent fibre using hyperspectral airborne surveys. Five
predictive modelling methods were assessed PLSR, random forest
regression, GPR, support vector regression and cubist regression.
Support vector regression performed best for crude protein, while cubist
regression performed best for acid detergent fibre.
In an Inner Mongolian monoculture test site, Zhao et al. (2021a) used a
UAV mounted spectrometer to measure chemical traits (carbon, nitrogen,
phosphorus, lignin, cellulose, and chlorophyll a and b). They
successfully measured numerous functional traits and noted that
retrieval worked better on an area basis rather than mass basis. Certain
traits could also be used effectively as predictors of AGB across the
monoculture sites.
In mixed grasslands on the Tibetan Plateau, UAV hyperspectral surveys
were used by Zhang et al. (2022) to map functional traits. Algorithms
used were PLSR, the generic algorithm integrated with the PLSR, random
forest (RF) and extreme gradient boosting (XGBoost). Chlorophyll a,
chlorophyll b, carotenoid content, starch content, specific leaf area
and leaf thickness were estimated well (r2 values
between 0.64 and 0.8), while nitrogen content, phosphorus content, plant
height and leaf dry matter content were modelled with lower accuracy
(r2 values between 0.3 and 0.54). Finally, Gholizadeh
et al. (2022) used aerial spectroscopy (1 m resolution) to map 12
functional traits and use these to differentiate an invasive grassland
species, Lespedeza cuneata , from other species in a tallgrass
prairie site in Oklahoma. They achieved an accuracy on 94%, showing
that functional trait measurements can be used to identify invasive
species. However, the accuracy was lower in species-rich grasslands.
Multispectral Remote Sensing of Grassland Functional
Traits
On the Qinghai-Tibetan Steppe, Li et al. (2018) used Landsat 8 and
Sentinel 2 to map plant functional traits at the community level -
canopy chlorophyll, specific plant area and plant dry matter content.
The authors used field sampling in combination with vegetation indices
and statistical modelling, with google earth engine, and achieved
moderately good results (r2 from 0.22 to 0.53). Imran
et al. (2020) used ground-based hyperspectral measurements to simulate
Sentinel 2 and 3 imagery, utilising red edge and NIR vegetation indices
to map LAI in grasslands in northern Italy and Austria. They found that
LAI retrieval is strongly influenced by plant traits, physical and
chemical (e.g., AGB, leaf angle distribution, brown pigment content and
chlorophyll content).
Over an experimental site in northern Germany, Grüner at al. (2020) used
UAV multispectral data, with PLSR and RF regression, for predicting
aboveground biomass and nitrogen fixation in legume-grass mixtures.
While consistent estimates of biomass were achieved using RF and the
results for nitrogen fixation were also strong, the most effective
regression method depended on the specific legume/grass proportions in
the test site. In a subsequent study, at the same location, Grüner at
al. (2021) combined a terrestrial laser scan survey with a multispectral
UAV survey (including texture analysis) to measure fresh and dry matter
(biomass) and nitrogen fixation in legume grass mixtures. The fusion
approach proved to be a significantly better predictor, overcoming the
limitation of each separate sensor. For nitrogen fixation, the
multispectral sensor had a relative root mean squared error of
prediction (rRMSEP) of 17.64%, with a rRMSEP of 20.07% for canopy
surface height and 14.4% for the sensor fusion.
Rakotoarivony et al. (2023) used field sampling to identify unique
functional traits that distinguishesL. cuneata from the native grassland species. Next, airborne
hyperspectral data was used to identify the vegetation indices most
closely related to these traits, and PlanetScope multispectral satellite
imagery was then used to identify and map L. cuneata across a 47
km2 region of U.S., with an accuracy of over 80%.
Across three grassland sites in Inner Mongolia, Zhao et al. (2024) used
field sampling to identify 13 functional traits and associated these
traits with leaf spectra through PLSR. PLSR was also applied to twelve
bands, 30 vegetation indices and convex hull volume from Sentinel 2
imagery to map function diversity on a large scale. All but one of the
functional traits (carbon) were predicted reasonably well from the
Sentinel 2 imagery, with r2 values of between 0.32 and
0.82.