Article

Leaf Area Index Estimation of Grassland Based on UAV-Borne

Hyperspectral Data and Multiple Machine Learning Models in
Hulun Lake Basin
Dazhou Wu 1,2,† , Saru Bao 3,† , Yi Tong 1,2 , Yifan Fan 1,2 , Lu Lu 3 , Songtao Liu 3 , Wenjing Li 1,2 , Mengyong Xue 1,2,4 ,
Bingshuai Cao 1,2 , Quan Li 3 , Muha Cha 5 , Qian Zhang 6 and Nan Shan 1,2, *

1 Nanjing Institute of Environmental Sciences, Ministry of Ecology and Environment of the People’s Republic
of China, Nanjing 210042, China; [email protected] (D.W.); [email protected] (Y.T.);
[email protected] (Y.F.); [email protected] (W.L.); [email protected] (M.X.);
[email protected] (B.C.)
2 Inner Mongolia Hulun Lake (Wetland) Comprehensive Monitoring Station for Ecological Quality,
Hulunbuir 021000, China
3 Hulunbuir Academy of Inland Lakes in Northern Cold & Arid Areas, Hulunbuir 021008, China;
[email protected] (S.B.); [email protected] (L.L.); [email protected] (S.L.);
[email protected] (Q.L.)
4 School of Geographic Science, Nantong University, Nantong 226019, China
5 College of Wildlife and Protected Area, Northeast Forestry University, Harbin 150006, China;
[email protected]
6 College of Geodesy and Geomatics, Shandong University of Science and Technology, Qingdao 266590, China;
[email protected]
* Correspondence: [email protected]
† These authors contributed equally to this work.

Abstract
Leaf area index (LAI) is a crucial parameter reflecting the crown structure of the grassland.
Accurately obtaining LAI is of great significance for estimating carbon sinks in grassland
ecosystems. However, spectral noise interference and pronounced spatial heterogeneity
within vegetation canopies constitute significant impediments to achieving high-precision
LAI retrieval. This study used hyperspectral sensor mounted on an unmanned aerial vehicle
Academic Editor: Jochem Verrelst
(UAV) to estimate LAI in a typical grassland, Hulun Lake Basin. Multiple machine learning
Received: 5 June 2025
(ML) models were constructed to reveal a relationship between hyperspectral data and
Revised: 14 August 2025
Accepted: 15 August 2025 grassland LAI using two input datasets, namely spectral transformations and vegetation
Published: 21 August 2025 indices (VIs), while SHAP (SHapley Additive ExPlanation) interpretability analysis was
Citation: Wu, D.; Bao, S.; Tong, Y.; further employed to identify high-contribution features in the ML models. The analysis
Fan, Y.; Lu, L.; Liu, S.; Li, W.; Xue, M.; revealed that grassland LAI has good correlations with the original spectrum at 550 nm
Cao, B.; Li, Q.; et al. Leaf Area Index and 750 nm–1000 nm, first and second derivatives at 506 nm–574 nm, 649 nm–784 nm, and
Estimation of Grassland Based on vegetation indices including the triangular vegetation index (TVI), enhanced vegetation
UAV-Borne Hyperspectral Data and
index 2 (EVI2), and soil-adjusted vegetation index (SAVI). In the models using spectral
Multiple Machine Learning Models in
Hulun Lake Basin. Remote Sens. 2025,
transformations and VIs, the random forest (RF) models outperformed other models (testing
17, 2914. https://doi.org/10.3390/ R2 = 0.89/0.88, RMSE = 0.20/0.21, and RRMSE = 27.34%/28.98%). The prediction error of
rs17162914 the random forest model exhibited a positive correlation with measured LAI magnitude
Copyright: © 2025 by the authors.
but demonstrated an inverse relationship with quadrat-level species richness, quantified
Licensee MDPI, Basel, Switzerland. by Margalef’s richness index (MRI). We also found that at the quadrat level, the spectral
This article is an open access article response curve pattern is influenced by attributes within the quadrat, like dominant species
distributed under the terms and and vegetation cover, and that LAI has positive relationship with quadrat vegetation cover.
conditions of the Creative Commons The LAI inversion results in this study were also compared to main LAI products, showing
Attribution (CC BY) license
a good correlation (r = 0.71). This study successfully established a high-fidelity inversion
(https://creativecommons.org/
licenses/by/4.0/).
framework for hyperspectral-derived LAI estimation in mid-to-high latitude grasslands of

Remote Sens. 2025, 17, 2914 https://doi.org/10.3390/rs17162914

Remote Sens. 2025, 17, 2914 2 of 20

the Hulun Lake Basin, supporting the spatial refinement of continental-scale carbon sink
models at a regional scale.

Keywords: grassland; hyperspectral remote sensing; leaf area index; interpretable machine
learning; carbon stock

1. Introduction
Grasslands, which are among the most widely distributed terrestrial ecosystems glob-
ally, are critical components of global carbon sinks, with its carbon sequestration capacity
second only to forest ecosystems [1]. Grasslands play an irreplaceable ecological role in
carbon sequestration and oxygen exchange, soil and water conservation, and biodiversity
protection [2]. The leaf area index (LAI), first proposed in the mid-20th century, refers to
the sum of all single-sided leaf areas per unit surface area in an ecosystem, measured in
(m2 /m2 , dimensionless) [3], used to quantify canopy density. The leaf surface is a critical
site for energy exchange between plant matter, so LAI, as a critical parameter reflecting
vegetation canopy structure, has played a significant role in both small-scale studies (e.g.,
plant breeding and forest resource management [4]) and large-scale studies (e.g., simulating
terrestrial carbon cycling and biogeochemical processes [5]).
The direct measurement methods for LAI mainly include destructive harvesting,
leaf collection, and leaf area meter measurements. Traditional methodologies for direct
LAI quantification are inherently constrained by substantial labor requirements and low
operational efficiency [6]. In recent years, using remote sensing techniques to estimate
LAI has become mainstream, primarily through empirical formulas and canopy radiation
transfer models (RTMs) [7,8]. However, due to insufficient spectral resolution, insensi-
tivity of biochemical parameters, and complex parameters in canopy radiation transfer
models, the LAI products obtained from remote sensing inherently carry a certain degree
of uncertainty. These uncertainties are further amplified when such products are used to
simulate terrestrial carbon–water cycles [9,10]. Liu et al. [11] found that the relative root
mean square error (RRMSE) between mainstream LAI products and ground measured
LAI ranged from 47.4% to 48.9%, while Tian et al. [12] found that in field experiments, the
MODIS LAI product generally underestimated the LAI within the sample plot by about 5%.
Comparatively, hyperspectral data encompass rich spectral information, thereby providing
superior discriminatory power compared to conventional multispectral approaches [13].
Commonly, optical satellites equipped with multispectral sensors typically have no more
than 30 bands. To meet the needs of large-scale monitoring, band settings are usually
discrete, with spectral resolutions typically greater than 50 nm, whereas hyperspectral
data typically have spectral resolutions less than 3 nm. Higher spectral resolution helps
extract changes in the vegetation’s spectral response across continuous spectra to retrieve
the physiological status of vegetation and provides spatially adaptive processing for het-
erogeneous grasslands [14]. Recent hyperspectral studies on vegetation primarily focus on
the extraction of certain parameters, such as nitrogen content, chlorophyll content, and the
extraction of the growing season of crops [15–17]. Platforms equipped with hyperspectral
sensors typically include satellites, handheld devices, or UAVs. Compared to the other
two types of payloads, UAVs can cover a moderate range and spatial resolution, making
them suitable for regional research [18].
Current research on hyperspectral LAI inversion has mainly focused on crops, such
as winter wheat and rice [19], with relatively few studies on natural grasslands. Unlike
managed crop fields or forests, where the canopy composition is more uniform, natural
Remote Sens. 2025, 17, 2914 3 of 20

grasslands present challenges for hyperspectral LAI inversion due to their diverse veg-
etation species, strong spatial heterogeneity, and low plant height, which mean that the
data are easily influenced by the soil background [20]. RTMs are still considered as the
main method for LAI inversion on natural grassland, but previous studies could hardly
reach decent LAI inversion accuracy using hyperspectral data and RTMs [21–23]. The
spatial heterogeneity is considered as the key factors causing RTMs’ poor performance in
natural grasslands [24], where spatial heterogeneity can be represented by species richness.
To address these challenges, vegetation indices resistant to soil interference, such as the
optimized soil-adjusted vegetation index (OSAVI), or red-edge indices, like the chlorophyll
red-edge index (Clre), are often designed to enhance hyperspectral sensitivity to grassland
LAI, but the improvements are not significant [25–27]. To address the limitations of pre-
vious studies in grassland LAI inversion, approaches utilizing machine learning (ML) or
integrating ML with empirical methods have achieved certain results in recent years [28–30],
but few studies used unmanned aerial vehicle (UAV) hyperspectral data. Hence, how well
ML models can improve LAI estimation accuracy compared to RTMs and other empirical
models has still not been validated. ML can extract complex grassland characteristics
to mitigate the difficulties caused by the inherent properties of natural grasslands [31].
One limitation of hyperspectral data is band redundancy, which can negatively affect model
inversion accuracy [32]. Methods using unsupervised ML, local detrended fluctuation
analysis, etc., can help select input bands for ML models and eventually, increase model
performance to some extent [33], but few studies has been conducted in natural grasslands
using ML and band selection methods at the quadrat level (~1 m × 1 m). The crown
reflectance of grasslands is usually contributed by multiple grass species, while in other
vegetation types, such as crops or coniferous/broadleaved forests, it is usually dominated
by single species or even a single individual tree. Hence, the LAI inversion results at the
quadrat scale in natural grasslands are more suitable for extrapolation to satellite scales
compared to other vegetation types which may have scale mis-match problems, such as in
forests [34,35]. Darvishzadeh et al. [36] found that RTMs’ LAI inversion could not reach
decent accuracy when the species number in sample plots exceeded two species. However,
the relationship between species richness and LAI inversion accuracy has not yet been
well discussed.
The grasslands in the Hulun Lake Basin represent a typical semi-arid steppe ecosystem
in Northeast Asia. Accurate measurement of grassland LAI in this area is critical for
quantifying vegetation productivity and validation of terrestrial carbon sink models in
local scale. The conservation of grassland ecosystems in the Hulun Lake Basin is of great
significance, as it not only sustains regional biodiversity and carbon sequestration but
also mitigates soil erosion and maintains hydrological stability, thereby safeguarding the
ecological integrity of this critical watershed in northern China. The accurate regional
LAI inversion model of grasslands in the Hulun Lake Basin is not developed yet. This
study selected the typical grassland of the Hulun Lake Basin as the research area, using
UAV-borne hyperspectral reflectance data and direct measurements of LAI at the quadrat
scale, ML algorithms, such as random forests (RFs), support vector machines (SVMs),
K-nearest neighbors (KNNs), and partial least squares regression (PLSR) were employed
to build a relationship between the hyperspectral data and LAI. Two scenarios, namely
using hyperspectral reflectance data and typical vegetation indices (VIs), were used to
build regression models for grassland LAI separately to enhance feature interpretation
and optimize model performance, as well as to enhance a data screening method based
on the correlation between bands and LAI. The performance of each model was evaluated
and compared, and the factors unique to natural grasslands that affected the best model’s
performance (based on prediction error) were discussed. The objective was to develop
Remote Sens. 2025, 17, 2914 4 of 20

estimation models for hyperspectral data and LAI of grasslands in the Hulun Lake Basin at
the quadrat level, analyze factors affecting model accuracy, and then explore the model’s
upscaling ability, which can support the spatial refinement of carbon sink models at the
regional scale.

2. Materials and Methods

2.1. Overview of the Research Area and Data Collection
The study area was located in the grassland region of the Hulun Lake Basin (Figure 1),
covering a total area of 150,000 km2 , with a climate characterized as temperate continental
monsoon. The average annual temperature is 0.83 ◦ C, and the average annual precipitation
ranges from 200 to 300 mm. Ground survey and data collection were conducted from
25 July to 1 August 2022, and from 19 July to 30 July 2023, when the grassland growth in the
Hulun Lake Basin is at its peak and can best reflect the typical LAI of the grassland. A total
of 17 plots with a size of 100 m × 100 m were set up within the study area. Within each
plot, five 1 m × 1 m quadrats were established using the five-point method, bounded by
red cones and white nylon lines. The main dominant grass species in study area included
Carex duriuscula, Allium ramosum, Potentilla bifurca, and Artemisia adamsi, etc.

Figure 1. Location of the study area and plot setting. (a) Study area location. (b) UAV platform: DJI
M600 pro. (c) Plots and quadrats setting. (d) Calibration panel.

The grassland LAI at each individual quadrat was measured using the LAI-2200
Canopy Analyzer (LI-COR, Lincoln, NE, USA). The canopy analyzer employs a “fish-
eye” sensor to measure light intensity at five different zenith angles above and below
the canopy, and calculates parameters, such as LAI, using an RTM model. Five measure-
ments were taken per quadrat at the center and four corners, with LAI values calculated
as the average of the valid measurements. Hyperspectral data were collected by a DJI
Matrice 600 Pro drone (DJI, Shenzhen, China) equipped with Headwall Nano-Hyperspec
micro hyperspectral camera (Headwall, Bolton, USA). The hyperspectral sensor measures
76 × 76 × 87 mm and weighs 0.52 kg. The raw hyperspectral data include 270 independent
spectral bands covering wavelengths from 400 to 1000 nm, with a fixed spectral resolution
of 1.85 nm and a full width at half maxima (FWHM) of 6 nm, details of the sensor could
be found in Table S1. The hyperspectral platform of the drone scanned each plot, which
Remote Sens. 2025, 17, 2914 5 of 20

consisted of five randomly distributed quadrats and one calibration panel used to correct
surface reflectance, thereby obtaining hyperspectral images of the quadrats simultaneously
with the measured LAI.

2.2. Hyperspectral Data Processing and Quality Control

A total of approximately 1 TB of raw hyperspectral data were collected during the
field survey. Using SpectralView software, the original hyperspectral data underwent
absolute radiometric correction, reflectance calculation, and geometric correction in se-
quence. First, the digital number (DN) values of the original hyperspectral data were
converted into physical radiation quantities using pre-calibrated coefficients from the lab-
oratory. Before takeoff, the drone lens was aligned with a 3 × 3 m calibration plate, and
the sensor automatically adjusted the exposure time based on lighting conditions to avoid
overexposure or underexposure. During flight, the calibration plate served as a reference
for reflectance calculations, eliminating the impact of lighting conditions and converting
radiance into surface reflectance. Finally, combining the aircraft attitude parameters and
DEM data, geometric correction was performed on the hyperspectral image to eliminate
image distortion.
In ENVI 5.6, hyperspectral reflectance data of individual quadrats were extracted
using white nylon ropes and red cones at the four corners as boundaries within the plot.
The buffer pixels with three pixels inside the inner edge of the white nylon rope were
removed to avoid the impact of the nylon rope on the surface reflectance within the quadrat.
Additionally, to ensure data reliability, strict quality control was applied to the pixels within
the quadrat (Table 1). For the valid pixels that passed the quality control, S-G filtering was
performed on each pixel individually to obtain a smoothed spectral response curve. Based
on this, the average spectral response curve for the quadrat was derived.

Table 1. Quality control conditions for hyperspectral data in the sample area.

Quality Control Standards Threshold

Surface reflectance (ρ) 0<ρ<1
Noise–signal ratio > 30
Solar zenith angle < 30◦
Light conditions Clear and cloudless

2.3. Spectral Information Extraction and Data Screening

The redundancy of bands in hyperspectral data and the multicollinearity between
bands with similar central wavelengths can affect modeling accuracy; thus, it is neces-
sary to remove redundant bands from hyperspectral data [37]. Surface reflectance bands
with central wavelength ranging from 400 nm to 1000 nm were selected due to the good
vegetation extraction responses of these bands. This study employed first-order deriva-
tives, second-order derivatives, and reciprocals of the original spectrum to transform
spectral characteristics, thereby further extracting useful information from the original
spectrum [38]. At the same time, hyperspectral vegetation indices (VIs) were utilized to
extract vegetation information. The calculation of VIs uses the built-in Vegetation Indices
Calculator tool in ENVI 5.6, selecting a total of 20 suitable VIs for this study, including
traditional Vis, like NDVI, EVI, TVI, and SAVI, and edge vegetation indices unique to
hyperspectral remote sensing data, such as mND705 and mSR705, which used red-edge
bands (i.e., a central wavelength at 705 nm) to enhance feature extraction ability [39,40].
All VIs were calculated by selecting bands with the nearest central wavelength of the
hyperspectral data according to the formulas in Table 2. These selected VIs have all under-
gone rigorous biological condition testing and are highly representative. Subsequently, the
Remote Sens. 2025, 17, 2914 6 of 20

correlation between all these data, including the original bands and their transformations
and vegetation indices, and LAI was calculated to obtain specific data sensitive to LAI.

Table 2. Vegetation indices names and their abbreviations used in this study.

Name Formula Name Formula

Normalized difference Normalized difference
(R800 − R670 )/(R800 + R670 ) (R790 − R720 )/(R790 + R720 )
vegetation index (NDVI) red-edge index (NDRE)
Soil-adjusted vegetation Chlorophyll red-edge index
1.5 × (R800 − R670 )/(R800 + R670 + 0.5) R750 /R720 − 1
index (SAVI) (Clre)
Optimized soil-adjusted Modified red-edge simple
1.16 × (R800 − R670 )/(R800 + R670 + 0.16) (R750 − R445 )/(R705 − R445 )
vegetation index (OSAVI) ratio index (mSR705 )
Modified

Hyperspectral-specific VIs
Green-normalized difference red-edge-normalized
(R750 − R550 )/(R750 + R550 ) (R750 − R705 )/(R750 + R705 − 2R445 )
vegetation index (GNDVI) difference vegetation index
Conventional VIs

(mND705 )
Chlorophyll green index Meris terrestrial chlorophyll
R800 /R560 −1 (R750 − R710 )/(R710 − R680 )
(Clgreen) index (MTCI)
Anthocyanin reflectance
Plant pigment ratio (PPR) (R550 − R450 )/(R550 + R450 ) (1/R559 )/(1/R721 )
index (ARI)
Two-band enhanced
2.5 × (R800 − R670 )/(R800 + 2.4 × R670 + 1) Greenness index (GI) R554 /R667
vegetation index (EVI2)
Red-edge normalized
(R750 − R705 )/(R750 + R705 ) Plant biochemical index (PBI) R810 /R560
vegetation index (NDVI705 )
Triangular vegetation index 0.5 × [120(R800 − R550 ) − 200(R670 − Spectral polygon vegetation
0.4 × [3.7(R800 − R670 ) − 1.2(R530 − R670 )]
(TVI) R550 )] index (SPVI)
Transformed chlorophyll
Enhanced vegetation index 2.5(R864 − R660 )/(R864 + 6R660 − 7.5R487 + 3 × [(R700 − R670 )− 0.2(R700 −
absorption in reflectance
(EVI) 1) R550 )(R700 /R670 )]
index (TCARI)

2.4. Model Construction and Accuracy Evaluation Method

Based on the characteristics of hyperspectral data, four ML regression algorithms,
namely random forest, support vector machine, K-nearest neighbor regression, and partial
least squares regression, were selected to establish models between hyperspectral data and
LAI using two sets of inputs, namely (1) original spectra and their transformations, and
(2) retrieved vegetation indices.
(1) Random Forest
A random forest (RF) is an ML algorithm that inherits the concept of ensemble learning.
By combining multiple decision trees, it enhances the model’s generalization ability and
predictive stability. In regression problems, the predictions from all decision trees are
averaged to derive the final prediction, which compensates for the bias or noise effects of
individual trees. The core advantage lies in their dual randomness of data and features, as
well as their ensemble strategy. This makes them highly robust against high-dimensional
data and noisy data, making them widely used in constructing hyperspectral models [41].
(2) Support Vector Machine
A support vector machine (SVM) is a supervised learning model whose core idea
is to find a hyperplane that separates different categories of data while maximizing the
margin between classes, thereby enhancing the model’s generalization ability. In regression
problems, SVM minimizes the error between predicted values and actual measurements
by finding a hyperplane. The advantage of SVM lies in its excellent performance on
datasets with a large number of features, and hyperspectral data falls precisely into this
category [42].
(3) K-Nearest Neighbor
K-nearest neighbor regression (KNNR) is a simple and effective instance based regres-
sion algorithm based on the principle of “similar to similar”, that is, similar data points
Remote Sens. 2025, 17, 2914 7 of 20

have similar output values. It calculates the distance between data points, finds the K-
nearest neighbors, and predicts the value of new data points according to the labels of these
neighbors [43].
(4) Partial Least Squares Regression
Partial least squares regression (PLSR) is a multivariate statistical method commonly
used to handle high-dimensional data, especially when there is a high degree of correlation
(multicollinearity) among independent variables, such as in hyperspectral data. The princi-
ple behind PLSR is to reduce the dimensionality of the data and project it simultaneously
into a new lower-dimensional space, extracting principal components that can explain
changes in the dependent variable, thereby constructing a regression model. In addition
to excelling in handling high-dimensional data, PLSR also offers deep insights into data
structure. It is widely applied in model building for hyperspectral remote sensing [44].
This paper selected the original spectra and their transformations, as well as typ-
ical vegetation indices, as input data to build LAI inversion models in two scenarios.
The measured LAI data from sample quadrats were used as ground truth data. These
two scenarios are two major methods for hyperspectral ML model inversion, and this paper
used both scenarios to evaluate the performance of two approaches.
A grid search method was employed to iterate over the hyperparameters of different
ML models to find the optimal combination of model hyperparameters, the detailed grid
seach parameters could be found in Table S2. At the same time, tenfold cross-validation
was implemented to mitigate evaluation variance and enhance result robustness [45].
In terms of model accuracy evaluation, this paper used three indices, namely coefficient
of determination, root mean square error (RMSE), and relative root mean square error
(RRMSE), to evaluate the prediction accuracy of the above models, defined as follows:

2
∑in = 1 (yi − ŷi )
R2 = 1 − 2
(1)
∑in = 1 (yi − y)
s
1 n
ni∑
RMSE = (yi − ŷi )2 (2)
=1

RRMSE = RMSE/y × 100% (3)

where yi is the measured value of the LAI of the sample, ŷi is the estimated value of the
LAI model of the sample, and y is the average value of the measured sample point.

2.5. Quadrat-Level Species Richness

This study used Margalef’s richness index (MRI) to represent species richness at the
quadrat level [46]. The species richness D was calculated using the number of species in
single quadrat S and the total grass plant count in a single quadrat N, as follows:

D = (S − 1)/ln( N ) (4)

2.6. SHAP Explains Models

The SHAP (Shapley additive explanation) method can quantify the impact of each
input feature value on the output within an artificial intelligence model, thereby providing
a quantitative explanation for the “black box” inside ML models [47]. For a model with
n feature values as inputs, the LAI value of the model output LAIpredict is contributed by
Remote Sens. 2025, 17, 2914 8 of 20

multiple feature values, and the contribution degree of a single feature value is its SHAP
value, as follows:
n
LAI predict = φ0 + ∑ φi xi (5)
i=1

where φi is the SHAP value representing individual feature importance. A positive SHAP
value indicates that the feature xi has a positive impact on the target LAIpredict , while a
negative value indicates a negative impact. The absolute value of the SHAP score reflects
the influence intensity of φi on the individual feature xi . The experimental environment
for this study was Python 3.8.18, and the model training used the ML library Scikit-learn
(version 1.3.0). For model evaluation calculations, the NumPy library in version 1.24.3 was
used. The SHAP interpretation model used the SHAP library in version 0.44.1.

3. Results
3.1. Feature Band and Vegetation Index Selection
The correlation between the surface reflectance of hyperspectral bands and its transfor-
mation with LAI is shown (Figure 2). The derivatives can enhance the detection of subtle
biochemical features that might be ignored by original bands. The area where the original
spectral reflectance has a high correlation with LAI was concentrated in the 750 nm–1000 nm
range. The spectral intervals that are sensitive to first-order and second-order derivatives
of LAI mainly fall within the 520 nm–540 nm, 580 nm–650 nm, and 700 nm–750 nm ranges.
Among these, there is a significant negative correlation between 580 nm–650 nm, while
the other two intervals show a clear positive correlation. After the inverse transformation
of the bands, the data have a lower correlation with LAI, so no bands are selected from
this range. Bands with a correlation coefficient above 0.5 in the original spectra, various
transformations, and vegetation indices are chosen. After removing multicollinearity using
variance inflation factor analysis, the selected characteristic bands were used for model
training (Table 3).

Figure 2. Pearson correlation between quadrat original reflectance and its transformation and LAI.
Remote Sens. 2025, 17, 2914 9 of 20

Table 3. Results of band selection, X1–X3: original bands, Y1–Y3: first derivatives, Z1–Z3: second
derivatives.

Spectrum Central
Pearson’s r Vegetation Indices Pearson’s r
Variables Wavelength (nm)
X1 972.5 0.58 NDVI 0.55
X2 808 0.57 SAVI 0.63
X3 769.2 0.56 PPR 0.54
Y1 726.7 0.71 EVI2 0.65
Y2 573.3 -0.70 TCARI 0.62
Y3 649.1 -0.66 TVI 0.67
Z1 689.3 0.61 SPVI 0.65
Z2 506.8 0.60 EVI 0.59
Z3 784.0 0.58

The Pearson correlation between filtered bands, vegetation indices, and LAI (Figure 3)
revealed that filtered bands have good correlations with both vegetation indices and LAI,
with correlation coefficients exceeding 0.55. Bands Y2 and Y3, which show a negative
correlation with LAI, exhibit strong negative correlations with all other bands. Among
the bands and their transformations, those with better correlations to LAI are Y1 and Y2,
both with absolute values of Pearson’s r of above 0.7. In the vegetation indices, the best
correlations are SAVI, TVI, and SPVI, all with correlations above 0.65.

Figure 3. Pearson correlation heat map between selected variables and LAI. The numbers represent
Pearson’s r.
Remote Sens. 2025, 17, 2914 10 of 20

3.2. Model Performance and Validation Between Current LAI Products

Random forest (RF), support vector machine (SVM), K-nearest neighbor (KNN), and
partial least squares regression (PLSR) techniques were used to model two scenarios,
one based on original spectra and transformations (Scenario 1), and the other based on
hyperspectral vegetation indices (Scenario 2), the results were shown in Figure 4. Random
forest (RF) consistently and significantly outperformed other models across both scenarios
(testing R2 = 0.88–0.89, RMSE = 0.20–0.21, RRMSE ≈ 27–29%), with a slightly better R2 ,
RMSE, and RRMSE in Scenario 1. SVM, KNN, and PLSR exhibited comparable performance
with R2 , RMSE, and RRMSE values ranging from 0.42 to 0.52, 0.42 to 0.46, and 57% to
63%, respectively. These three models all performed marginally better in Scenario 1 than
in Scenario 2.

Figure 4. Comparison of LAI modeling results between Scenario 1 (original spectra and transforma-
tions) and Scenario 2 (vegetation indices).

A validation was made to address the relationship between the LAI predicted in this
study and a widely recognized LAI product (Figure 5). The average LAI of the whole plot
(~100 m × 100 m) was generated using RF models in both scenarios to spatially match
two LAI products, namely MODIS LAI (MOD15A2H) and GLASS LAI. LAI-VI consistently
performed better than LAI-SPEC in correlating with both products, achieving Pearson’s r
values of 0.78 (vs. MOD15A2H, Figure 5c) and 0.41 (vs. GLASS, Figure 5d), compared to
LAI-SPEC’s Pearson’s r values of 0.71 (MOD15A2H, Figure 5a) and 0.35 (GLASS, Figure 5b).
Moreover, both models exhibited stronger correlations with MOD15A2H than with GLASS
LAI, underscoring MOD15A2H’s superior alignment with our predictions.

3.3. Uncertainty and Mapping of LAI Estimation Model Predictions

The uncertainty values (derived as the absolute value of the deviation between pre-
dicted LAI and measured LAI) of random forest models in both scenarios are shown in
Figure 6, using LAI value and species richness at the quadrat level (represented by Mar-
galef’s richness index, MRI) as classification criteria, and the interval was set unevenly to
ensure that each interval had a similar amount of samples. The violin plot in Figure 6a
shows that when the MRI increases, the variation of LAI estimation uncertainty of both
Remote Sens. 2025, 17, 2914 11 of 20

methods decreases. The spectral data-based model (Scenario 1) showed lower uncertainty
when the MRI was <1 and 1–1.8 than the VI-based model (Scenario 2), but higher uncer-
tainty in the >1.8 interval. The box plot in Figure 6b shows the distribution of uncertainty
with the measured LAI. In both models, the low uncertainty was observed when the mea-
sured LAI was low (<0.4 and 0.4–0.6); when measured LAI increases (0.6–1 and >1), the
distribution of uncertainty samples tends to vary.

Figure 5. Comparison between LAI prediction in this study using RF models in two scenarios (LAI
SPEC and LAI LAI) and LAI products (MOD15A2H and GLASS); the LAI prediction is the average
LAI in a whole plot (~100 m × 100 m) to match the spatial resolution of the LAI products.

The predictions using RF models constructed using spectrum (LAI-SPEC) and VIs
(LAI-VI) over non-sample area are shown in Figure 6c. The spatial mapping ability of
two models were examined by comparing RGB image with two predictions. Notably,
both models can accurately predict LAI under different grassland structures, especially in
dense vegetation area. An overestimation of LAI value on bare soil was observed in both
scenarios (the LAI value is expected to be 0), and LAI-VI showed larger overestimation
than LAI-SPEC. In the S11 plot, the model provided a LAI prediction value of around 0.4
for bare soil pixels, whereas LAI-SPEC, on the other hand, provided a prediction of around
0.2 to 0.4 for bare soil pixels. The poor performance of both models on bare soil pixels may
be due to a lack of bare soil samples. In this study, all quadrats were set in vegetated area,
and pure bare soil quadrats were not included in sampling and training session. In the
complex scenario of S14, where vegetation is sparse, LAI-SPEC better captures high-LAI
pixels, showing an advantage over the VI-based model in terms of LAI mapping.
Remote Sens. 2025, 17, 2914 12 of 20

Figure 6. (a) Absolute LAI estimation deviation in different species richness samples; (b) absolute
LAI estimation deviation in different measured LAI samples; (c) spatial distribution of LAI modeling
results in the sample plots. From top to bottom are RGB composite image, LAI mapping visualization
using the random forest model in Scenario 1, and LAI mapping visualization using the random forest
model in Scenario 2.

In addition to the MRI and LAI values, the uncertainty of the hyperspectral LAI
inversion results is also influenced by various factors, such as dominant species and species
diversity within the quadrat, and the overall canopy cover. The main dominant species
in this study included Carex duriuscula, Allium ramosum, Potentilla bifurca, and Artemisia
adamsi, etc. By extracting pure single-species vegetation pixels from quadrat images, we
obtained normalized spectral response curves for typical dominant species (Figure 7).
The spectral responses of pure vegetation pixels show strong consistency, with surface
reflectance forming local peaks around the green band (550 nm) and rapidly increasing
between the near-infrared bands (700–800 nm), differing from bare soil pixels. There are
differences in spectra among different vegetation types; the increase in the spectral response
of Artemisia adamsii at 550 nm is significantly stronger than that of other vegetation, while
the peak reflectance of Allium ramosum at 770 nm is slightly higher than that of other
vegetation, and it shows a slow downward trend after reaching the peak.
A statistically significant positive correlation was observed between total canopy
cover and measured LAI (Pearson’s r = 0.51, p < 0.01) (Figure 8). When the total canopy
cover exceeds 70%, the quadrats’ LAI is significantly higher than those with lower canopy
cover. The spectral response curves of different levels of total canopy cover show significant
differences between 600–700 nm; quadrats with high total canopy cover exhibit a downward
trend in this range, while quadrats with total canopy cover below 50% tend to be stable.
Quadrats with low total canopy cover are more influenced by bare soil pixels in the red-light
band compared to vegetation. In the 850–900 nm range, quadrats with high total canopy
cover show a “convex” distribution in their spectral curves, while quadrats with low total
canopy cover show a “concave” shape.
Remote Sens. 2025, 17, 2914 13 of 20

Figure 7. Quadrat level spectrum response curves. (a) Spectral response curves of typical dominant
species and bare soil, and pictures of grass species; (b) spectral response curves of pure vegetation,
sample average, and typical bare soil within a single dominant species quadrat.

Figure 8. Relationship between quadrat canopy cover and LAI. (a,b) Relationship between quadrats’
total canopy cover and LAI; (c) spectral response curves (normalized) for different quadrat canopy
cover levels; the dashed red squares represent region where the canopy cover shows differences.
Remote Sens. 2025, 17, 2914 14 of 20

3.4. Analysis of the Eigenvalue Contribution of SHAP

The SHAP analysis of the RF models in the two scenarios found that in the spectral
model (Scenario 1), the contribution of each feature value to the output LAI was ranked
as follows (Figure 9, detailed SHAP values of all features were listed in Table S3): Y2
(first derivative, 573.3 nm) > Z2 (second derivative, 506.8 nm) > Y1 (first derivative,
726.7 nm) > Y3 (first derivative, 649.1 nm). Among these, Y2 had a stronger explanatory
power than other bands and transformations. In the VI model (Scenario 2), the contribution
of the vegetation index to the model output was as follows: PPR > NDVI > TVI > SPVI.
When the model output LAI is extremely high (LAI > 1.5), PPR, TVI, and SPVI all exhibit
strong positive effects on the model output. When the output LAI is low, PPR and NDVI
show strong negative effects.

Figure 9. Schematic diagram of SHAP values for each feature. (a) Scenario 1; (b) Scenario 2. Black line
above represents the LAI value output by the model, the black bar on the right shows the absolute
size of the SHAP values for each feature, and the central heat map illustrates the SHAP values of
different features under a single input.

The high explanatory power of Y2 (first derivative, 573.3 nm) likely stems from its
sensitivity to subtle variations in chlorophyll absorption dynamics at the transition zone
between green reflectance and red absorption, where the first derivative enhances spectral
contrast related to leaf biochemical properties and canopy structure. In contrast, other
features primarily capture structural or red-edge features, which may exhibit less direct
coupling with LAI under varying environmental conditions, explaining Y2’s superior
predictive dominance. On the other hand, in the model constructed using vegetation
indices (Scenario 2), several vegetation indices exhibited considerable model explanatory
power in addition to the PPR because they integrate multiple spectral bands to enhance
sensitivity to vegetation biophysical properties, with their differential impacts at high and
low LAI levels reflecting distinct physiological and structural responses.

4. Discussion
4.1. The Influence of Spatial Heterogeneity on Model Performance
At the quadrat level, the spectral response curve integrates various species of vegeta-
tion and soil spectral characteristics within the quadrat, potentially introducing interference
in accurate LAI estimation. The discrimination of the spectral response curves between
different canopy covers (Figure 8c) at 600–700 nm is due to an increase in the response
curve of bare soil pixels in the red-light band, whereas vegetation pixels show the opposite
effect. The proportion of these two types of pixels in the quadrat determines the overall
Remote Sens. 2025, 17, 2914 15 of 20

trend of the average spectral response curve. Quadrat-level species composition and total
canopy cover introduce spectral variability that can bias quadrat-scale LAI inversion results,
but hyperspectral data provide compensatory advantages through pixel-level separation
of vegetation and non-vegetation components via continuous spectra and quadrat-level
canopy cover dependent spectral weighting using spectral curve attributes between certain
sensitive wavelength intervals. These techniques help to solve issues raised by the spatial
heterogeneity of grasslands. For instance, previous studies have discussed the relationship
between hyperspectral response and species richness [48,49], vegetation cover [50–52],
and dominant species [53,54], all of which supplied or showed good alignments with the
findings of this study.

4.2. Comparison of This Study with Existing LAI Products

LAI, as a key parameter in large-scale carbon sink estimation models, is widely used
in various process mechanism models, such as Biome-BGC and BEPS [55–57], and in
light energy utilization (LUE) models, like CASA or VPM, LAI is crucial to accurately
estimate the critical parameter of the photosynthetically active radiation absorption ra-
tio (fPAR ) [58,59]. Currently, most models’ LAI-driven data come from mature remote
sensing LAI products, like MOD15A2H and GLASS LAI [60], which are mostly based
on medium-to-low-resolution surface monitoring satellites serving all ecosystem types at
the global scale. However, these products have not been optimized for individual ecosys-
tems, such as grasslands. Studies have shown that MODIS LAI products overestimate
by 53% in grasslands [58]. The optimal models in this study achieved RRMSE values be-
tween 27.34% and 28.98%, substantially outperforming MODIS LAI products which exhibit
RRMSE > 47% in grassland validation performed by Liu et al. [11]. The RF model con-
structed using VIs in this study showed good correlation with MODIS LAI (Figure 5),
which may indicate a greater potential in correcting global LAI products at the region scale
compared to the RF model constructed using the spectrum in this study.

4.3. Potential and Limitations of Grassland UAV Hyperspectral LAI Estimation for Optimizing the
Accuracy of Regional-Scale Carbon Sink Models
Global carbon sink models require continuous LAI input data, so medium-to-low-
resolution satellites with shorter revisit cycles are typically used. These satellites do not
carry hyperspectral sensors, making it difficult for models based on hyperspectral bands
to be directly applied to multispectral satellite remote sensing fields [61]. Hyperspectral
satellites, such as PRISMA and GF-5, however, have excessively long revisit cycles and
cannot provide the high temporal resolution data required by carbon sink models. Cur-
rently, most mature LAI inversion algorithms are based on RTMs, establishing lookup
tables between red/near-infrared surface reflectance and LAI to find the matched LAI
value [62]. A study calculated VIs for winter wheat LAI inversion using Sentinel-2 and
multiple ML models, concluding that the random forest is the best model, which is similar
to the findings of our study [63]. From the SHAP results of this paper, the high contributed
features in the RF model constructed using VIs, such as PPR, NDVI, SPVI, and TVI, do not
involve hyperspectral-unique bands, which makes it possible to reproduce the RF model
constructed by VIs in this study on different satellite remote sensing platforms, such as
MODIS, Landsat, Sentinel-2, etc.
This study utilized hyperspectral data from UAVs to establish an inversion model of
hyperspectral data and grassland LAI at the quadrat scale. UAVs have two advantages
in grassland LAI inversion, namely scale-up and scale-down. First, UAVs can scale up,
which to some extent simulates the spatial heterogeneity within a pixel caused by the
scale effect in LAI data used in large-scale carbon sink models. The other advantage
of UAVs at the quadrat scale is that they can scale down, allowing for more accurate
Remote Sens. 2025, 17, 2914 16 of 20

spatiotemporal matching with direct measurements of LAI. Whether using destructive

sampling or handheld leaf area meters, the area that can be measured in a single session is
limited (usually meter-level), where satellite data with spatial resolution of 10–1000 m can
hardly match the direct measurement scale [64]. The UAV scale, on the other hand, generally
has spatial resolutions in the centimeter range, allowing for spatio-temporal alignment with
direct measurements of LAI, making model inversion results more convincing compared
to results from satellite remote sensing data. Due to these two advantages, UAV-borne LAI
inversion models have the potential to serve as a bridge between ground measurement
and large-scale satellite monitoring. The scaling up of modeling results from drone remote
sensing data is a significant direction for future research.
To reach the goal of regional refinement of LAI inputs for carbon sink models, future
efforts should be focused on reproducing the LAI inversion models mentioned in this
study with satellite sensors, such as Sentinel-2 and MODIS. The VI-based RF model is a
better choice because the SHAP results (Figure 7) show that important features have little
dependency on hyperspectral-unique VIs, which makes it possible to reproduce the model
on multi-spectral sensors on satellites, and the validation results in Figure 5 have already
shown good correlation between the VI-based RF model and current LAI products. The
future work will include quantifying the impact of band width and central wavelength
discrepancies on VI calculations, establishing sensor-specific correction coefficients, and
validating model robustness within the study region. Notably, the following current
limitations must be addressed: UAV data’s limited temporal coverage hinders capturing
seasonal dynamics; scale mismatches between UAV and satellite pixels may introduce
biases in upscaling; and model performance under extreme conditions (e.g., drought or
heavy grazing) remains untested, necessitating long-term, multi-context validation.

5. Conclusions
This study utilized ML models to construct a relationship model between UAV-borne
hyperspectral data and grassland LAI by screening original bands, their transformations,
and vegetation indices. Our results demonstrated that the random forest model performed
better than other models in both scenarios (R2 = 0.88), with model uncertainty increasing
with higher measured LAI and decreasing with higher quadrat species richness. The anal-
ysis of the impact of spatial heterogeneity within quadrats on model inversion accuracy
through various vegetation indicators measured within the quadrats was conducted. The
LAI model constructed from UAV hyperspectral data for grasslands has great potential
as an input for regional LAI data in large-scale carbon sink models. The UAV platform’s
principal advantage resides in its capacity for bidirectional scaling, effectively bridging
microscale ground observations and macroscale satellite applications. Additionally, the
spatial heterogeneity within quadrats helps to simulate the coarse resolution issues caused
by mixed pixels in ground observation satellite data to some extent. This study demon-
strated the modeling accuracy and inclusiveness of spatial consistency within plots using a
UAV hyperspectral data-based LAI inversion model for grasslands, laying the foundation
for future research in large-scale carbon sink models.

Supplementary Materials: The following supporting information can be downloaded at: https:
//www.mdpi.com/article/10.3390/rs17162914/s1, Table S1: Parameters of Nano-Hyperspec hyper-
spectral sensor; Table S2: Grid search hyperparameter settings and best parameters for 4 models;
Table S3: SHAP values for all features in both scenarios.

Author Contributions: Conceptualization, S.B., Y.F. and N.S.; data curation, Y.T., L.L., W.L., M.X.,
B.C. and Q.L.; formal analysis, D.W. and S.B.; funding acquisition, Q.Z. and N.S.; investigation,
Y.T., L.L., S.L., M.X., B.C., Q.L. and M.C.; methodology, D.W. and Y.F.; project administration, S.L.
Remote Sens. 2025, 17, 2914 17 of 20

and N.S.; Software, D.W. and S.B.; Supervision, N.S.; validation, Y.T. and W.L.; visualization, D.W.;
writing—original draft, D.W. and S.B.; writing—review & editing, Y.F. and N.S. All authors have read
and agreed to the published version of the manuscript.

Funding: This research was jointly funded by the First Phase of the Project of Strengthening the
Scientific and Technological Research Capacity of Hulun Lake Natrue Reserve (HSZCS-C-F-210094),
the Ecological Security Investigation and Assessment Project of Hulun Lake (HSZCS-G-F-210059),
the National Science Foundation of China (42071050), the Taishan Scholar Project (TSQN202306210),
and the Natural Science Foundation of Inner Mongolia Autonomous Region (2024ZD13).

Data Availability Statement: The original contributions presented in this study are included in the
article. Further inquiries can be directed to the corresponding author.

Conflicts of Interest: The authors declare that they have no known competing financial interests or
personal relationships that could have appeared to influence the work reported in this paper.

References
1. Lal, R. Carbon Cycling in Global Drylands. Curr. Clim. Chang. Rep. 2019, 5, 221–232. [CrossRef]
2. Bardgett, R.D.; Bullock, J.M.; Lavorel, S.; Manning, P.; Schaffner, U.; Ostle, N.; Chomel, M.; Durigan, G.; Fry, E.L.; Johnson, D.;
et al. Combatting global grassland degradation. Nat. Rev. Earth Environ. 2021, 2, 720–735. [CrossRef]
3. Chen, J.M.; Cihlar, J. Retrieving leaf area index of boreal conifer forests using landsat TM images. Remote Sens. Environ. 1996, 55,
153–162. [CrossRef]
4. Osem, Y.; O’Hara, K. An ecohydrological approach to managing dryland forests: Integration of leaf area metrics into assessment
and management. Forestry 2016, 89, 338–349. [CrossRef]
5. De Bock, A.; Belmans, B.; Vanlanduit, S.; Blom, J.; Alvarado-Alvarado, A.A.; Audenaert, A. A review on the leaf area index (LAI)
in vertical greening systems. Build. Environ. 2023, 229, 109926. [CrossRef]
6. Fang, H.L.; Baret, F.; Plummer, S.; Schaepman-Strub, G. An Overview of Global Leaf Area Index (LAI): Methods, Products,
Validation, and Applications. Rev. Geophys. 2019, 57, 739–799. [CrossRef]
7. Cho, M.A.; Ramoelo, A.; Math, R. Estimation of leaf area index (LAI) of South Africa from MODIS imagery by inversion of
PROSAIL radiative transfer model. In Proceedings of the IEEE Joint International Geoscience and Remote Sensing Symposium
(IGARSS)/35th Canadian Symposium on Remote Sensing, Quebec City, QC, Canada, 13–18 July 2014; pp. 2590–2593.
8. López-Lozano, R.; Casterad, M.A. LAI estimation by scaling up and inversion of radiative transfer models from Quickbird images.
Rev. De Teledetec. 2005, 24, 43–47.
9. Fang, H.L.; Wang, Y.; Zhang, Y.H.; Li, S.J. Long-Term Variation of Global GEOV2 and MODIS Leaf Area Index (LAI) and Their
Uncertainties: An Insight into the Product Stabilities. J. Remote Sens. 2021, 2021, 9842830. [CrossRef]
10. Fang, H.L.; Wei, S.S.; Jiang, C.Y.; Scipal, K. Theoretical uncertainty analysis of global MODIS, CYCLOPES, and GLOBCARBON
LAI products using a triple collocation method. Remote Sens. Environ. 2012, 124, 610–621. [CrossRef]
11. Liu, Y.B.; Xiao, J.F.; Ju, W.M.; Zhu, G.L.; Wu, X.C.; Fan, W.L.; Li, D.Q.; Zhou, Y.L. Satellite-derived LAI products exhibit large
discrepancies and can lead to substantial uncertainty in simulated carbon and water fluxes. Remote Sens. Environ. 2018, 206,
174–188. [CrossRef]
12. Tian, Y.H.; Woodcock, C.E.; Wang, Y.J.; Privette, J.L.; Shabanov, N.V.; Zhou, L.M.; Zhang, Y.; Buermann, W.; Dong, J.R.; Veikkanen,
B.; et al. Multiscale analysis and validation of the MODIS LAI product: I. Uncertainty assessment. Remote Sens. Environ. 2002, 83,
414–430. [CrossRef]
13. Mateen, M.; Wen, J.H.; Nasrullah; Akbar, M.A. The Role of Hyperspectral Imaging: A Literature Review. Int. J. Adv. Comput. Sci.
Appl. 2018, 9, 51–62. [CrossRef]
14. Bioucas-Dias, J.M.; Plaza, A.; Camps-Valls, G.; Scheunders, P.; Nasrabadi, N.M.; Chanussot, J. Hyperspectral Remote Sensing
Data Analysis and Future Challenges. Ieee Geosci. Remote Sens. Mag. 2013, 1, 6–36. [CrossRef]
15. Jain, N.; Ray, S.S.; Singh, J.P.; Panigrahy, S. Use of hyperspectral data to assess the effects of different nitrogen applications on a
potato crop. Precis. Agric. 2007, 8, 225–239. [CrossRef]
16. Yang, Y.C.; Nan, R.; Mi, T.X.; Song, Y.X.; Shi, F.H.; Liu, X.R.; Wang, Y.Q.; Sun, F.L.; Xi, Y.J.; Zhang, C. Rapid and Nondestructive
Evaluation of Wheat Chlorophyll under Drought Stress Using Hyperspectral Imaging. Int. J. Mol. Sci. 2023, 24, 5825. [CrossRef]
[PubMed]
17. Wang, L.; Chen, S.S.; Peng, Z.P.; Huang, J.C.A.; Wang, C.Y.; Jiang, H.; Zheng, Q.; Li, D. Phenology Effects on Physically Based
Estimation of Paddy Rice Canopy Traits from UAV Hyperspectral Imagery. Remote Sens. 2021, 13, 1792. [CrossRef]
18. Toth, C.; Józków, G. Remote sensing platforms and sensors: A survey. ISPRS J. Photogramm. Remote Sens. 2016, 115, 22–36.
[CrossRef]
Remote Sens. 2025, 17, 2914 18 of 20

19. Kanning, M.; Kühling, I.; Trautz, D.; Jarmer, T. High-Resolution UAV-Based Hyperspectral Imagery for LAI and Chlorophyll
Estimations from Wheat for Yield Prediction. Remote Sens. 2018, 10, 2000. [CrossRef]
20. Yu, K.Q.; Zhao, Y.R.; Zhu, F.L.; Li, X.L.; He, Y. Mapping of Chlorophyll and SPAD Distribution in Pepper Leaves During Leaf
Senescence Using Visible and Near-Infrared Hyperspectral Imaging. Trans. Asabe 2016, 59, 13–24.
21. Bayat, B.; Van der Tol, C.; Verhoef, W. Remote Sensing of Grass Response to Drought Stress Using Spectroscopic Techniques and
Canopy Reflectance Model Inversion. Remote Sens. 2016, 8, 557. [CrossRef]
22. Ding, Z.; Zhu, X.; Ma, L.; Zhao, Y. Estimating LAI and uncertainty in grassland using UAV hyperspectral data and PROSAIL.
Adv. Comput. Signals Syst. 2024, 8, 13–22. [CrossRef]
23. Si, Y.L.; Schlerf, M.; Zurita-Milla, R.; Skidmore, A.; Wang, T.J. Mapping spatio-temporal variation of grassland quantity and
quality using MERIS data and the PROSAIL model. Remote Sens. Environ. 2012, 121, 415–425. [CrossRef]
24. Lu, B.; Proctor, C.; He, Y.H. Investigating different versions of PROSPECT and PROSAIL for estimating spectral and biophysical
properties of photosynthetic and non-photosynthetic vegetation in mixed grasslands. Gisci. Remote Sens. 2021, 58, 354–371.
[CrossRef]
25. Fern, R.R.; Foxley, E.A.; Bruno, A.; Morrison, M.L. Suitability of NDVI and OSAVI as estimators of green biomass and coverage in
a semi-arid rangeland. Ecol. Indic. 2018, 94, 16–21. [CrossRef]
26. He, Y.H.; Guo, X.L.; Wilmshurst, J.F. Comparison of different methods for measuring leaf area index in a mixed grassland. Can. J.
Plant Sci. 2007, 87, 803–813. [CrossRef]
27. Imran, H.A.; Gianelle, D.; Rocchini, D.; Dalponte, M.; Martín, M.P.; Sakowska, K.; Wohlfahrt, G.; Vescovo, L. VIS-NIR, Red-
Edge and NIR-Shoulder Based Normalized Vegetation Indices Response to Co-Varying Leaf and Canopy Structural Traits in
Heterogeneous Grasslands. Remote Sens. 2020, 12, 2254. [CrossRef]
28. Shen, B.B.; Ding, L.; Ma, L.C.; Li, Z.W.; Pulatov, A.; Kulenbekov, Z.; Chen, J.Q.; Mambetova, S.; Hou, L.L.; Xu, D.W.; et al.
Modeling the Leaf Area Index of Inner Mongolia Grassland Based on Machine Learning Regression Algorithms Incorporating
Empirical Knowledge. Remote Sens. 2022, 14, 4196. [CrossRef]
29. Qin, G.X.; Wu, J.; Li, C.B.; Meng, Z.Y. Comparison of the hybrid of radiative transfer model and machine learning methods in leaf
area index of grassland mapping. Theor. Appl. Climatol. 2024, 155, 2757–2773. [CrossRef]
30. Tsele, P.; Ramoelo, A. Integrating Active Learning and Regression Methods for Estimation of Grass Lai Over a Mountainous
Region using Sentinel-2 Satellite Data. In Proceedings of the IGARSS 2024—2024 IEEE International Geoscience and Remote
Sensing Symposium, Athens, Greece, 7–12 July 2024; pp. 10516–10519.
31. Doepper, V.; Rocha, A.D.; Berger, K.; Graenzig, T.; Verrelst, J.; Kleinschmit, B.; Foerster, M. Estimating soil moisture content under
grassland with hyperspectral data using radiative transfer modelling and machine learning. Int. J. Appl. Earth Obs. Geoinf. 2022,
110, 102817. [CrossRef]
32. Liu, Z.W.; Jiang, J.B.; Du, Y.; Xu, Z.F. A Band Influence Algorithm for Hyperspectral Band Selection to Classify Moldy Peanuts.
IEEE Access 2021, 9, 147527–147536. [CrossRef]
33. Li, J.H.; Li, Q.Z.; Wang, F.; Liu, F. Hyperspectral redundancy detection and modeling with local Hurst exponent. Phys. A-Stat.
Mech. Its Appl. 2022, 592, 126830. [CrossRef]
34. Zhang, H.; Li, X.W.; Cao, C.X.; Yang, H.; Gao, M.X.; Zheng, S.; Xu, M.; Xie, D.H.; Jia, H.C.; Ji, W.; et al. Scale effects of leaf area
index inversion based on environmental and disaster monitoring satellite data. Sci. China-Earth Sci. 2010, 53, 92–98. [CrossRef]
35. Fan, W.J.; Gai, Y.Y.; Xu, X.R.; Yan, B.Y. The spatial scaling effect of the discrete-canopy effective leaf area index retrieved by remote
sensing. Sci. China-Earth Sci. 2013, 56, 1548–1554. [CrossRef]
36. Darvishzadeh, R.; Skidmore, A.; Schlerf, M.; Atzberger, C. Inversion of a radiative transfer model for estimating vegetation LAI
and chlorophyll in a heterogeneous grassland. Remote Sens. Environ. 2008, 112, 2592–2604. [CrossRef]
37. Zhang, M.Y.; Gong, M.G.; Chan, Y.Q. Hyperspectral band selection based on multi-objective optimization with high information
and low redundancy. Appl. Soft Comput. 2018, 70, 604–621. [CrossRef]
38. Yang, C.B.; Feng, M.C.; Song, L.F.; Jing, B.H.; Xie, Y.K.; Wang, C.; Yang, W.D.; Xiao, L.J.; Zhang, M.J.; Song, X.Y. Study on
hyperspectral monitoring model of soil total nitrogen content based on fractional-order derivative. Comput. Electron. Agric. 2022,
201, 107307. [CrossRef]
39. Ceccato, P.; Flasse, S.; Tarantola, S.; Jacquemoud, S.; Grégoire, J.M. Detecting vegetation leaf water content using reflectance in the
optical domain. Remote Sens. Environ. 2001, 77, 22–33. [CrossRef]
40. Lin, Y.H.; Shen, H.F.; Tian, Q.J.; Gu, X.F.; Yang, R.R.; Qiao, B.J. Mechanisms underlying diurnal variations in the canopy spectral
reflectance of winter wheat in the jointing stage. Curr. Sci. 2020, 118, 1401–1406. [CrossRef]
41. Shebl, A.; Abriha, D.; Fahil, A.S.; El-Dokouny, H.A.; Elrasheed, A.A.; Csámer, A. PRISMA hyperspectral data for lithological
mapping in the Egyptian Eastern Desert: Evaluating the support vector machine, random forest, and XG boost machine learning
algorithms. Ore Geol. Rev. 2023, 161, 105652. [CrossRef]
Remote Sens. 2025, 17, 2914 19 of 20

42. Rodríguez-Pérez, R.; Bajorath, J. Evolution of Support Vector Machine and Regression Modeling in Chemoinformatics and Drug
Discovery. J. Comput. Aided Mol. Des. 2022, 36, 355–362. [CrossRef]
43. Sumayli, M. Development of advanced machine learning models for optimization of methyl ester biofuel production from papaya
oil: Gaussian process regression (GPR), multilayer perceptron (MLP), and K-nearest neighbor (KNN) regression models. Arab. J.
Chem. 2023, 16, 104833. [CrossRef]
44. Burnett, A.C.; Anderson, J.; Davidson, K.J.; Ely, K.S.; Lamour, J.; Li, Q.Y.; Morrison, B.D.; Yang, D.D.; Rogers, A.; Serbin, S.P. A
best-practice guide to predicting plant traits from leaf-level hyperspectral data using partial least squares regression. J. Exp. Bot.
2021, 72, 6175–6189. [CrossRef] [PubMed]
45. Xiong, Z.; Cui, Y.X.; Liu, Z.H.; Zhao, Y.; Hu, M.; Hu, J.J. Evaluating explorative prediction power of machine learning algorithms
for materials discovery using k-fold forward cross-validation. Comput. Mater. Sci. 2020, 171, 109203. [CrossRef]
46. Qureshi, H.; Anwar, T.; Mohibullah, M.; Fatima, S.; Younas, R.; Habiba, U.; Malik, L.; Hanif, A.; Iqbal, M. Paired plot experiments
to assess impact of invasive species on native floral diversity in Pakistan. Front. Environ. Sci. 2023, 10, 1037319. [CrossRef]
47. Zhang, J.Y.; Ma, X.L.; Zhang, J.L.; Sun, D.L.; Zhou, X.Z.; Mi, C.L.; Wen, H.J. Insights into geospatial heterogeneity of landslide
susceptibility based on the SHAP-XGBoost model. J. Environ. Manag. 2023, 332, 117357. [CrossRef]
48. Psomas, A.; Kneubühler, M.; Huber, S.; Itten, K.; Zimmermann, N.E. Hyperspectral remote sensing for estimating aboveground
biomass and for exploring species richness pat-terns of grassland habitats. Int. J. Remote Sens. 2011, 32, 9007–9031. [CrossRef]
49. Gholizadeh, H.; Gamon, J.A.; Townsend, P.A.; Zygielbaum, A.I.; Helzer, C.J.; Hmimina, G.Y.; Yu, R.; Moore, R.M.; Schweiger, A.K.;
Cavender-Bares, J. Detecting prairie biodiversity with airborne remote sensing. Remote Sens. Environ. 2019, 221, 38–49. [CrossRef]
50. Zhang, F.; Wang, C.; Pan, K.; Guo, Z.; Liu, J.; Xu, A.; Ma, H.; Pan, X. The Simultaneous Prediction of Soil Properties and Vegetation
Coverage from Vis-NIR Hyperspectral Data with a One-Dimensional Convolutional Neural Network: A Laboratory Simulation
Study. Remote Sens. 2022, 14, 397. [CrossRef]
51. Pervin, R.; Robeson, S.M.; MacBean, N. Fusion of airborne hyperspectral and LiDAR canopy-height data for estimating fractional
cover of tall woody plants, herbaceous vegetation, and other soil cover types in a semi-arid savanna ecosystem. Int. J. Remote
Sens. 2022, 43, 3890–3926. [CrossRef]
52. Pu, Y.; Wilmshurst, F.J.; Guo, X. Separating Shrub Cover From Green Vegetation in Grasslands Using Hyperspectral Vegetation
Indices. Can. J. Remote Sens. 2024, 50, 2347630. [CrossRef]
53. Liu, W.; Han, W.; Jin, G.; Gong, K.; Ma, J. Classification of major species in the sericite–Artemisia desert grassland using
hyperspectral images and spectral feature identification. PeerJ 2024, 12, e17663. [CrossRef]
54. Zhu, X.; Bi, Y.; Du, J.; Gao, X.; Zhang, T.; Pi, W.; Zhang, Y.; Wang, Y.; Zhang, H. Research on deep learning method recognition
and a classification model of grassland grass species based on unmanned aerial vehicle hyperspectral remote sensing. Grassl. Sci.
2023, 69, 3–11. [CrossRef]
55. Srinet, R.; Nandy, S.; Patel, N.R.; Padalia, H.; Watham, T.; Singh, S.K.; Chauhan, P. Simulation of forest carbon fluxes by integrating
remote sensing data into biome-BGC model. Ecol. Model. 2023, 475, 110185. [CrossRef]
56. Zhang, S.; Zhang, J.H.; Bai, Y.; Koju, U.A.; Igbawua, T.; Chang, Q.; Zhang, D.; Yao, F.M. Evaluation and improvement of the daily
boreal ecosystem productivity simulator in simulating gross primary productivity at 41 flux sites across Europe. Ecol. Model.
2018, 368, 205–232. [CrossRef]
57. Amiro, B.D.; Chen, J.M.; Liu, J. Net primary productivity following forest fire for Canadian ecoregions. Can. J. For. Res. 2000, 30,
939–947. [CrossRef]
58. Fu, G.; Wu, J.S. Validation of MODIS collection 6 FPAR/LAI in the alpine grassland of the Northern Tibetan Plateau. Remote Sens.
Lett. 2017, 8, 831–838. [CrossRef]
59. Yan, K.; Park, T.; Chen, C.; Xu, B.D.; Song, W.J.; Yang, B.; Zeng, Y.L.; Liu, Z.; Yan, G.J.; Knyazikhin, Y.; et al. Generating Global
Products of LAI and FPAR From SNPP-VIIRS Data: Theoretical Background and Implementation. IEEE Trans. Geosci. Remote Sens.
2018, 56, 2119–2137. [CrossRef]
60. Ma, H.; Liang, S.L. Development of the GLASS 250-m leaf area index product (version 6) from MODIS data using the bidirectional
LSTM deep learning model. Remote Sens. Environ. 2022, 273, 112985. [CrossRef]
61. Sara, D.; Mandava, A.K.; Kumar, A.; Duela, S.; Jude, A. Hyperspectral and multispectral image fusion techniques for high
resolution applications: A review. Earth Sci. Inform. 2021, 14, 1685–1705. [CrossRef]
62. Cheng, J.P.; Yang, H.; Qi, J.B.; Sun, Z.D.; Han, S.Y.; Feng, H.K.; Jiang, J.Y.; Xu, W.M.; Li, Z.H.; Yang, G.J.; et al. Estimating
canopy-scale chlorophyll content in apple orchards using a 3D radiative transfer model and UAV multispectral imagery.
Comput. Electron. Agric. 2022, 202, 107401. [CrossRef]
Remote Sens. 2025, 17, 2914 20 of 20

63. Mudi, S.; Paramanik, S.; Behera, M.D.; Prakash, A.J.; Deep, N.R.; Kale, M.P.; Kumar, S.; Sharma, N.; Pradhan, P.; Chavan, M.;
et al. Moderate resolution LAI prediction using Sentinel-2 satellite data and indirect field measurements in Sikkim Himalaya.
Environ. Monit. Assess. 2022, 194, 897. [CrossRef]
64. Tsele, P.; Ramoelo, A.; Qabaqaba, M.; Mafanya, M.; Chirima, G. Validation of LAI, chlorophyll and FVC biophysical esti-
mates from sentinel-2 level 2 prototype processor over a heterogeneous savanna and grassland environment in South Africa.
Geocarto Int. 2022, 37, 14355–14378. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.