Using Vegetation Indices Developed for Sentinel-2 Multispectral Data to Track Spatiotemporal Changes in the Leaf Area Index of Temperate Deciduous Forests

Abstract

The leaf area index (LAI) in temperate forests is highly dynamic throughout the season, and lacking such dynamic information has limited our understanding of carbon and water flux patterns in these ecosystems. This study aims to explore the potential of using vegetation indices based on Sentinel-2 data, which includes three additional spectral bands in the red-edge region of its multispectral imager (MSI) sensor compared to previous satellite-borne imagery, to effectively track seasonal variations in LAI within typical cold–temperate deciduous forests originating in rugged terrain in Japan. We evaluated reported vegetation indices and developed an index specific to Sentinel-2 data to effectively monitor the spatiotemporal changes of LAI in mountainous deciduous forests, providing more accurate data for ecological monitoring. Results showed that the developed index (SR_B12,B7) was able to track LAI at both seasonal and spatial scales (R² = 0.576). Further analyses revealed that the index nevertheless performed relatively poorly during the leaf-maturing season when LAI peaks, suggesting that it still suffers from a “saturation” problem. For high-resolution tracking of LAI in temperate deciduous forests at both temporal and spatial scales, future research is needed to incorporate additional information.

1. Introduction

Leaf area index (LAI) is typically defined as the ratio of the total leaf surface area to the unit ground area and can be expressed as LAI = leaf area/ground area [1,2]. LAI is inextricably linked to various physiological and ecological processes, including photosynthesis, transpiration, energy exchange, and carbon sequestration [3], and has served as a crucial structural metric for vegetation and has been a key input for numerous ecological process models [4]. Variations in LAI are accompanied by changes in ecosystem productivity [5], making it an important component of the ecological system cycle. Its temporal and spatial variability is particularly pronounced in deciduous forests [6], making accurate estimation of LAI fundamental to understanding terrestrial biogeochemical cycles and the broader functioning of deciduous forest ecosystems [7].

LAI is primarily obtained by means of both direct and indirect methods of measurement in the field. Direct methods, such as destructive sampling and litter collection [8], provide accurate estimates but are time-consuming and labor-intensive, limiting their applicability for long-term and large-scale spatial and temporal monitoring [7,9], particularly in deciduous forest ecosystems with pronounced spatiotemporal variation. Furthermore, repeated measurements may lead to the accumulation of errors [10]. In contrast, rapid and non-destructive indirect methods, especially non-contact LAI measurement techniques, have gained increasing popularity [11]. Among them, digital hemispherical photography (DHP) has been widely adopted due to its cost-effectiveness and ability to provide LAI estimates closest to the true value under optimal exposure conditions [12,13].

LAI estimation using DHP involves several key steps, including segmentation of sky and canopy pixels, calculation of gap fractions, estimation of the clustering index, correction for clustering effects, and the final LAI estimation based on gap fractions [13]. A variety of commercial and free software packages have been developed for processing hemispherical images, such as Hemiview, Gap Light Analyzer (GLA), Winscanopy, and CAN_EYE [13], to name a few. However, these applications were mainly based on thresholding methods that relied on the visual grayscale of the image [14]. The interactive graphical user interfaces (GUIs) of these software packages depended on user-defined settings, which could introduce systematic errors [15]. In response to these limitations, Brown proposed an open-source Python-based software package designed for a range of DHP image formats [16]. This software offers increased automation and flexibility over existing tools. While seasonal variations in canopy color were readily observed from DHP images, the challenge of accurately relating the extracted information to specific stages of leaf development remained unresolved [17]. Furthermore, the acquisition of DHP images proved to be challenging in rugged mountainous terrain, presenting a limitation for dynamic ecosystem studies.

Alternatively, for the large-scale monitoring of forest LAI, remote sensing technology offers improved opportunities. However, satellite data acquisition is often limited by long revisit intervals, especially in cloudy conditions, making acquiring high-quality, continuous, and high-frequency satellite data a significant challenge, especially in temperate mountainous forest regions [18,19]. The Moderate Resolution Imaging Spectroradiometer (MODIS) system, with its moderate resolution, offers a short revisit period. However, its spatial resolution ranges from 250 to 1000 m [20]. This relatively coarse spatial resolution limited its applicability for fine-scale monitoring, particularly in regions with complex forest structures [20,21]. In addition, MODIS LAI products have generally been overestimates of LAI, especially at higher spatial resolutions [22].

In comparison, Sentinel-2 provides multispectral imagery with more frequent revisit cycles (5 days) and finer spatial resolution (10 m, 20 m, 60 m). In addition, its three-satellite constellation facilitated global coverage [23]. The 13 spectral bands of the Sentinel-2 Multispectral instrument (MSI) sensor spanned the visible, near-infrared, and shortwave infrared regions, with the three narrow bands (B5, B6, B7) in the red-edge region (RE) being particularly sensitive to chlorophyll content. Under low LAI conditions, the reflectance of the red band was higher than that of the near-infrared (NIR) band, whereas the opposite was observed under high LAI conditions [24]. Additionally, the reflectance of the NIR band was higher in boreal broadleaf forests than in coniferous forests [25]. These characteristics enhanced the potential of Sentinel-2 to improve the accuracy of LAI estimation in broadleaf forests [26]. Recent studies have further highlighted that three red-edge vegetation indices (VI_3RE), such as three red-edge normalized difference vegetation indices (NDVI_3RE), derived from the three red-edge bands of Sentinel-2 showed a robust correlation with LAI, thereby enhancing the accuracy of seasonal LAI estimation [27]. Even so, the accuracy of the LAI derived from these images is not yet fully developed, particularly in mountainous areas. Research showed that terrain played a significant role in influencing remotely sensed reflectance, which in turn affected the LAI derived from ground-based reflectance data. As the slope of the terrain increased, the error in LAI estimates also tended to increase, indicating that topographic factors contributed to a greater degree of uncertainty in LAI inversion [28]. Therefore, accurate estimation of LAI in mountainous areas using Sentinel-2 remains a challenging problem.

There have been several approaches to infer LAI from remotely sensed imagery. Among them, regression methods based on the relationship between vegetation indices (VIs) and ground truth data have been successfully implemented across different remote sensing platforms and sensors [29]. VIs, which are derived from the mathematical combination of spectral bands based on vegetation’s absorption and reflection characteristics [30], effectively minimize the interference from background noise, making them essential for dynamic vegetation monitoring [30]. Previous studies based on LAI-VIs have generally concluded that vegetation indices were particularly sensitive to low LAI values [31]. However, when LAI is relatively low, vegetation indices such as the normalized difference vegetation index (NDVI) were significantly influenced by soil reflectance, and their sensitivity decreased as LAI increased [30,32]. On the other hand, during the summer months when LAI reaches saturation levels, VIs are challenging to track LAI effectively [22]. As a result, a number of improved vegetative indices, such as enhanced vegetation index (EVI) and soil-adjusted vegetation index (SAVI), have been developed to address the saturation issue associated with traditional VIs under high LAI conditions [2,30]. Despite these improvements, accurately tracking highly saturated LAI values using VIs remains a significant challenge to date. In addition, multiple factors, including species composition, developmental stage, stand structure, and seasonality, may contribute to uncertainty in LAI estimation as well [8].

This study aims to (1) evaluate the feasibility of using reported VIs to assess the spatial–temporal variations in LAI in deciduous forests by integrating DHP image data with Sentinel-2 MSI sensor data; (2) explore the potential of developing spectral vegetation indices uniquely tailored to Sentinel-2 data, suitable for plot-scale studies; (3) apply the developed index to estimate the annual, seasonal, and spatial LAI, assessing the developed index‘s performance and its superiority over existing VIs. This evaluation is based on DHP-derived LAI time series data from 2021–2023, collected from a rugged temperate site dominated by deciduous forests. Special attention is given to comprehensive evaluations of reported and developed VIs, on annual, seasonal, and spatial contexts based on DHP-derived LAI time series from 2021 to 2023 in a rugged temperate site dominated by deciduous forests.

2. Materials and Methods

2.1. Study Area

The study area is located within the Shizuoka University Forest in Nakakawane, Japan (35°04′ N, 138°06′ E), with an elevation ranging from 390 to 1560 m [22]. The vegetation composition exhibits a distinct altitudinal gradient, transitioning from an evergreen broad-leaved forest dominated by red oak at lower elevations to a deciduous broad-leaved forest predominantly consisting of beech and oak at higher elevations. The region is characterized by a typical alpine temperate climate, with an average annual temperature of 17 °C and an average annual precipitation of approximately 2153 mm [22].

The size of the study site is 150 m × 90 m. Within this site, sixty experimental plots, each measuring 15 m × 15 m, were established (Figure 1), and field measurements were conducted within each plot. The center of each plot was marked by GPS data recorded using a real-time kinematic (RTK) GPS receiver, which facilitated accurate and continuous data collection and enabled precise spatial mapping for subsequent analyses [7].

2.2. Field Sampling and Digital Hemisphere Photography Processing

Fieldwork was conducted annually from April to November, during which digital hemispherical photographs were captured using a Nikon D5100 digital camera equipped with a Sigma 4.5 mm F2.8 EX fisheye lens, providing a 180° field of view. To minimize the influence of the zenith angle, images were taken at a height of approximately 1.3 to 1.5 m above the ground. Given that previous studies have indicated that the camera’s horizontal angle has minimal impact on the LAI estimation, a handheld leveling device was employed to compensate for the tripod’s leveling [21], and two to three hemispherical images were captured vertically at the zenith angle, directed upward from the center of each plot.

To facilitate batch processing of DHP images collected in the field using the HemiPy v0.1.2 package, the images were assigned to structured file directories with root directories named by date and subdirectories named by plot. Each root directory contained data for 60 research plots, each consisting of two to three images taken upwards and vertically to the zenith. To process these images, the clustering algorithm was employed to separate sky and canopy (tree branches and leaves) pixels within the image [15]. This algorithm has demonstrated superior stability and accuracy when compared to other methods [15]. In the processing of these images, only the blue band was utilized to maximize the contrast between the sky and canopy, which helps to minimize the effects of multiple scattering and chromatic aberration within the canopy, thereby ensuring a more accurate calculation of the gap fraction [33].

Due to the complexity of image classification, directly estimating LAI could lead to significant errors [34]. Thus, the Hemipy v0.1.2 package, available on GitHub, was used to estimate the Plant Area Index (PAI) (https://github.com/luke-a-brown/hemipy.git (accessed on 1 November 2024)). Image classification for upward-images is generally more sensitive to all canopy elements (such as branches, leaves, and stems) [34]. To account for the clumping and woody elements within the canopy, a clumping factor was introduced for correction. The corrected effective LAI (LAIe) is referred to as PAI (PAIe) [5]. Hemipy estimates PAIe with two methods, the Hinge method and the Miller method. The Hinge method considers only the gap fraction within a 57.5° angular range around the hinge, which makes the calculation of the gap fraction largely independent of the leaf angle distribution. This approach applies to a variety of canopy types. The Miller method is a generalized version of the Miller integral model that incorporates multi-angle observations [16]. Previous studies demonstrated that the LAI derived from the Hinge method more accurately reflected the true ground-based LAI [16]. Therefore, we adopted the PAIe calculated using the Hinge method as the dependent variable LAIe in the subsequent analyses. Since previous studies suggested LAIe exhibited a stronger correlation with remote sensing observations and field measurements than with true LAI, in our research, we utilized available LAI estimates as the dependent variable [22,35].

2.3. Sentinel-2 Data and Processing

The Sentinel-2 MSI Level 2A data were obtained from Google Earth Engine (GEE, https://developers.google.com (accessed on 14 November 2024)). These data have been processed through atmospheric correction and surface reflectance conversion to mitigate atmospheric effects and provide more accurate surface reflectance information [36], making them suitable for accurate ground vegetation analysis and other environmental studies. The images corresponding to the dates on which the DHP images were taken (31 times in total over the three years) were screened, and those with >20% cloud cover were excluded from further analysis. Finally, a total of 22 Sentinel-2 images were downloaded, including 7 images in 2021, 9 in 2022, and 6 in 2023. Each Sentinel-2 imagery includes 13 spectral bands covering aerosol, visible, red edge, water vapor, near-infrared (NIR), cirrus, and shortwave infrared (SWIR) regions of the electromagnetic spectrum. For this study, bands B2 through B12 were selected, excluding B9 (the water vapor band).

To optimize the output from GEE, the image was cropped to focus on the regions of interest (ROIs), which covered the 60 research plots. All spectral bands were resampled to a 10 m resolution using the nearest neighbor method to ensure a uniform pixel size. A 7.5 m buffer was generated around the center of each plot to account for the difference in spatial resolution between Sentinel-2 imagery (10 m) and the 15 m × 15 m ground-based LAI measurement plots, ensuring better alignment between the remotely sensed data and the field measurements. The average spectral values of all pixels within this buffer were then extracted [22].

2.4. Reported Vegetation Indices and New Spectral Index Development

Since single-band reflectance can be easily affected by background factors, VIs are generally more stable and provide a more reliable representation of land cover information [37]. VIs are advantageous because they account for the combined reflectance of multiple spectral bands, which helps minimize the effects of background variability [37]. A range of characteristic indices were selected for this study, including vegetation indices reported in previous studies and a set of developed spectral indices. The empirical indices and their formulas are summarized in

Table 1. Summary of reported empirical vegetation indices.

Index	Abbreviation	Formula	References
Normalized difference vegetation index	NDVI	(NIR − R)/(NIR + R)	[38]
Enhanced vegetation index	EVI	2.5 × (NIR − R)/(NIR + 6× R − 7.5B + 1)	[39]
Normalized green difference vegetation index	NGDVI	(NIR − G)/(NIR + G)	[40]
Atmospheric resistance vegetation index	ARVI	(NIR − 2R + B)/(NIR + 2R − B)	[41]
Ratio vegetation index I	RVI54	SWIR1/NIR	[42]
Ratio vegetation index II	RVI64	SWIR2/NIR	[42]
Excess green index	EXG	2G − R − B	[43]
Excess green minus red index	EXGR	2G − 2.4R	[43]
Excess red index	EXR	1.4R − B	[43]
Modified soil adjustment vegetation index	MSAVI	(2NIR + 1 − sqrt((2NIR + 1)2 − 8(NIR − R)))/2	[44]
Modified second ratio index	MSRI	((NIR/R) − 1)/sqrt((NIR/R) + 1)	[45]
Moisture vegetation index	MVI	sqrt((NIR − R)/(NIR + R) + 0.5)	[20]
Normalized difference index	NDI	(NIR − SWIR1)/(NIR + SWIR1)	[46]
Normalized green-red difference index	NGRDI	(G − R)/(G + R)	[47]
Chlorophyll normalized vegetation index	NPCI	(R − B)/(R + B)	[43]
Optimization of soil regulatory vegetation index	OSAVI	1.6(NIR − R)/(NIR + R + 0.16)	[48]
Redness index	RI	(R − G)/(R + G)	[49]
Ratio vegetation index	RVI	NIR/R	[50]
Source address validation improvement	SAVI	1.5(NIR − R)/(NIR + R + 0.5)	[39]
Structure-independent pigment index	SIPI	(NIR − B)/(NIR + B)	[48]
Simple ratio pigment index	SRPI	B/R	[48]
Transform chlorophyll absorption index	TCARI	3((RE1 − R) − 0.2(RE1 − G))/(RE1/R)	[48]
Transformed vegetation index	TVI	60(NIR − G) − 100(R − G)	[50]
Visible atmospherically resistant index	VARI	(G − R)/(G + R − B)	[40]
Visible difference vegetation index	VDVI	(2G − (R + B))/(2G + (R + B))	[51]
Wide dynamic range vegetation index	WDRVI	(NIR − R)/(NIR + R)	[52]
Red edge chlorophyll index	CIred-edge	(RE3/RE1) − 1	[53]
Inverted red edge chlorophyll index	IRECI	(RE3 − R)/(RE1/RE2)	[54]
Modified chlorophyll absorption ratio index	MCARI	((RE1 − R) − 0.2 * (RE1 − G)) × (RE1 − R)	[55]
MERIS terrestrial chlorophyll index	MTCI	(RE2-RE1)/(RE1-R)	[56]
Sentinel-2 red-edge position index	S2REP	705 + 35× ((((RE3 + R)/2) − RE1)/(RE2 − RE1))	[54]
Red-edge-based plant index	REPI	(0.5× (RE3 + R) − RE1)/(RE2 − RE1)	[38]

. All these indices were calculated in GEE.

The developed spectral indices encompassed five types: individual band (λ), simple ratio (SR), wavelength difference (D), normalized difference (ND), and inverse difference (ID), as outlined in

Table 2. Index types for estimating LAI, including λ: individual band, SR: simple ratio, D: wavelength difference, ND: normalized difference, and ID: inverse difference. λ1 and λ2 represent two distinct bands of Sentinel-2 that were randomly selected for the calculation.

Index Type	Formula of Index
λ (λ1)	λ
SR (λ1, λ2)	${\displaystyle \frac{{\lambda }_1}{{\lambda }_2}}$
D (λ1, λ2)	λ1− λ2
ND (λ1, λ2)	$\left({\displaystyle \frac{{\lambda }_1-{\lambda }_2}{{\lambda }_1+{\lambda }_2}}\right)$
ID (λ1, λ2)	${\displaystyle \frac{1}{{\lambda }_1}}-{\displaystyle \frac{1}{{\lambda }_2}}$

[22]. These indices were derived from all possible combinations of Sentinel-2 bands. In cases where the developed index utilized the same band combination as an existing vegetation index, the corresponding index was calculated only once to avoid redundancy [7,57]. The vegetation indices were calculated by R Studio v2024.09.0 (accessed on 20 November 2024).

2.5. Statistical Criteria

The primary model employed in this study was the linear regression type, and the key performance metrics used to assess its effectiveness include coefficient of determination (R²), root mean squared error (RMSE), ratio of performance to deviation (RPD), Akaike information criterion (AIC). After having performed the Kolmogorov–Smirnov test, we noted that the original data exhibited a markedly skewed distribution. To address this, a logarithmic transformation was applied to the data prior to conducting the regression analysis, as log-transformed data are often more likely to satisfy the assumptions of normality required for regression. R² quantifies the proportion of variance in the dependent variable explained by the independent variable. RMSE measures the disparity between predicted and actual observed values, with a lower RMSE indicating a better model fit. RPD (Equation (1)) is a standardized metric used to evaluate the relative prediction error of the model, offering a comparative measure of model performance. The calculation formula for RPD is as follows:

$$\mathrm{RPD}={\displaystyle \frac{\sigma }{\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}}$$

(1)

where σ represents the standard deviation of the observed data, reflecting the dispersion of the measured ground LAI values. A higher σ indicates the observed values were more dispersed from the mean.

The Akaike Information Criterion (AIC) (Equation (2)) has also been employed in the iterative process of selecting the optimal band combination [57]. This approach strikes a balance between the goodness of fit and the complexity of the model, thereby mitigating the risk of overfitting. The calculation formula is as follows:

$$\mathrm{AIC}=2k-2\ln \left(\widehat{L}\right)$$

(2)

where k denotes the number of parameters in the model, and $\widehat{L}$ represents the maximum likelihood estimate of the model. In this study, the valmetrics package in R studio was utilized to directly compute R², RMSE, and RPD. The AIC was calculated using the built-in linear model function, lm (), and the AIC values were iteratively compared using the AIC function in R studio.

3. Results

3.1. Annual and Spatial Variations of Ground LAI Derived from DHP

Figure 2 presents the average ground-measured LAI values within each leaf phenological stage derived from DHP photographs, illustrating the seasonal variation of ground-based LAI from 2021 to 2023. The leaf flushing stage corresponded to the months of April and May, the leaf maturity stage spanned from June to August, and the leaf senescence stage covered the months from September to November. However, some data were missing due to cloud screening by Sentinel-2 imagery, particularly during the summer months. Distinct seasonal patterns were observed in the distribution of color blocks. LAI began to rise sharply around DOY 150, reaching a peak near DOY 250, after which it gradually declined. Overall, the LAI in 2023 was the highest among the three leaf stages, with the annual average LAI reaching up to 6. The LAI during the maturity stage was generally higher than that during the flushing and senescence stages, particularly in 2023. The LAI in the flushing stage was generally slightly higher than that in the senescence stage. Although the LAI peak in 2022 during the flushing stage was higher than in 2021, the annual average LAI in 2022 was slightly higher than in 2021. In the senescence stage, the LAI performance across the three years was similar, but the LAI in 2023 was approximately 0.5 higher than in the other two years.

To further investigate the spatial variations in ground-based LAI, we conducted local spatial autocorrelation analysis for each phenological stage from 2021 to 2023 (

Table 3. The local spatial autocorrelation analyses of ground LAI. The evaluation metrics included I: Moran’s I statistic, SD: standard deviation, p: p-value, E: expected index, and Var: variance.

Year	Leaf Stage	I	E(I)	Var	SD	p
2021	Flushing	0.100	6.374	0.084	−0.017	0.007
	Maturity	0.291	6.374	0.000	−0.006	0.002
	Senescence	−0.199	6.374	1.000	−0.006	0.002
2022	Flushing	−0.042	6.374	0.824	−0.004	0.002
	Maturity	−0.261	6.374	0.998	−0.017	0.007
	Senescence	−0.032	6.374	0.672	−0.007	0.003
2023	Flushing	0.271	6.374	0.000	−0.017	0.007
	Maturity	0.302	6.374	0.000	−0.008	0.004
	Senescence	−0.115	6.374	0.989	−0.006	0.002

). By combining Moran’s I value and the p-value, we found that during the Flushing and Maturity stages of 2021 and 2023, the Moran’s I value was higher, and the p-value was smaller, indicating a significant spatial agglomeration effect. In other words, there was strong spatial autocorrelation in LAI during these stages. However, the spatial autocorrelation was not significant during most of 2022, particularly in the Flushing and Maturity stages, but the Senescence phase exhibited weak spatial autocorrelation.

3.2. Evaluations of Reported Spectral Indices

Table 4. Comparison of the 32 empirical vegetation indices. The r values for the correlation analysis between each index and the log-transformed ground LAI are shown in the table, with statistically significant results (p < 0.01) highlighted in bold.

VIs	Total	2021			2022			2023
		Flushing	Maturity	Senescence	Flushing	Maturity	Senescence	Flushing	Maturity	Senescence
NDVI	0.407	−0.028	−0.070	0.949	0.607	−0.082	0.305	0.170	0.136	0.627
EVI	0.450	−0.051	0.035	0.945	0.633	−0.036	0.602	0.008	−0.136	0.460
GNDVI	0.185	−0.014	−0.070	0.943	0.169	−0.100	0.038	0.209	0.142	0.387
ARVI	0.539	−0.038	−0.032	0.949	0.727	−0.059	0.480	0.166	0.142	0.759
RVI54	−0.720	0.074	−0.067	−0.922	−0.708	0.048	−0.729	0.011	−0.063	−0.895
RVI64	−0.717	0.208	0.040	−0.932	−0.722	0.029	−0.360	−0.039	−0.123	−0.906
EXG	0.340	−0.114	−0.067	0.940	0.693	−0.046	0.561	−0.130	−0.143	0.617
EXGR	0.151	−0.012	−0.024	0.938	0.688	−0.042	0.355	−0.091	0.095	0.648
EXR	−0.052	−0.005	−0.023	−0.903	−0.573	0.033	−0.165	−0.195	−0.138	−0.587
MSAVI	0.468	−0.044	−0.023	0.945	0.709	−0.092	0.536	0.051	−0.038	0.373
MSRI	0.523	−0.057	−0.100	0.952	0.650	−0.051	0.538	0.158	0.135	0.743
MVI	0.363	−0.022	−0.068	0.947	0.598	−0.084	0.255	0.170	0.136	0.579
NDI	0.689	−0.063	0.070	0.936	0.704	−0.048	0.725	−0.017	0.063	0.886
NGRDI	0.623	−0.091	−0.101	0.940	0.735	−0.006	0.769	0.068	0.124	0.769
NPCI	−0.601	0.082	−0.099	−0.882	−0.521	−0.125	−0.743	−0.032	−0.083	−0.451
OSAVI	0.446	−0.033	−0.042	0.949	0.691	−0.092	0.426	0.081	0.046	0.456
RI	−0.623	0.091	0.101	−0.940	−0.735	0.006	−0.769	−0.068	−0.124	−0.769
RVI	0.528	−0.082	−0.114	0.944	0.651	−0.030	0.607	0.152	0.132	0.751
SAVI	0.456	−0.037	−0.014	0.946	0.710	−0.100	0.489	0.034	−0.058	0.372
SIPI	0.148	−0.017	−0.100	0.932	0.281	−0.097	−0.019	0.181	0.127	0.342
SRPI	0.581	−0.086	0.089	0.895	0.497	0.123	0.656	0.033	0.087	0.504
TCARI	0.089	0.017	0.067	0.508	0.587	−0.113	0.133	−0.227	−0.149	−0.166
TVI	0.435	−0.045	0.016	0.943	0.713	−0.103	0.558	−0.022	−0.136	0.416
VARI	0.639	−0.091	−0.053	0.942	0.727	0.007	0.741	0.071	0.125	0.780
VDVI	0.522	−0.097	−0.130	0.951	0.706	−0.060	0.662	0.068	0.121	0.798
WDRVI	0.407	−0.028	−0.070	0.949	0.607	−0.082	0.305	0.170	0.136	0.627
CIred-edge	0.510	−0.055	−0.016	0.950	0.734	−0.058	0.702	−0.011	−0.062	0.513
MTCI	0.455	−0.073	0.063	0.930	0.210	0.165	0.616	0.201	0.048	0.475
MCARI	0.276	0.006	−0.013	0.925	0.647	−0.187	0.662	−0.128	−0.165	0.187
IRECI	0.510	−0.084	−0.040	0.950	0.719	−0.043	0.667	0.111	0.113	0.669
S2REP	0.130	−0.040	0.027	0.438	0.074	0.099	0.166	−0.004	0.029	−0.052
REPI	0.130	−0.040	0.027	0.438	0.074	0.099	0.166	−0.004	0.029	−0.052

summarizes the evaluation results of the correlation between 32 VIs and ground LAI. Among them, RVI54 exhibited the highest absolute value of the correlation coefficient with LAI (r = −0.720), while EXR showed the lowest absolute value of correlation with LAI (r = −0.052). Except for EXR, all indices demonstrated a significant correlation with LAI (p < 0.01).

The performance of each index varied across different leaf growth stages. After categorizing all the data into three leaf stages, it was found that during the Senescence stage, the correlation between indices and ground LAI was above 0.5, with the exception of TCARI (r = 0.055), S2REP (r = 0.135), REPI (r = 0.135), GNDVI (r = 0.382), and SIPI (r = 0.339). Among these, NDI exhibited the highest correlation with LAI (r = −0.8642). In the Flushing stage, the absolute values of the correlation coefficients between the indices and LAI varied considerably, with RVI64 having the highest absolute value (r = −0.730) and TCARI having the lowest (r = −0.008). In comparison with the previous two stages, the Maturity stage exhibited lower absolute values for the correlation coefficients between the vegetation indices and LAI, with EXG showing the highest absolute value (r = −0.504), and S2REP and REPI presenting the lowest absolute values (r = −0.026).

To gain a more detailed understanding of the performance of each index in each leaf stage in different years, we separated the total leaf stages by year. From 2021 to 2023, these indices performed best in the Senescence stage, especially in the Senescence stage of 2021. Except for S2REP, REPI, and TCARI, the absolute values of the correlation coefficients between the remaining indices and LAI exceeded 0.8. In the Flushing stage, only most of the indices in 2022 performed well. In 2021 and 2023, the absolute value of the correlation coefficient was only 0.227 (TCARI: r= −0.227). All indices performed poorly in the Maturity stage (r= −0.227). It is worth noting that S2REP and REPI performed poorly in all periods, with very low R² values in all leaf stages.

3.3. Performance of the Developed Indices

The developed spectral index is based on all possible band combinations of Sentinel-2. To evaluate the predictive ability of each band combination for the LAI, a series of regression analysis methods were used to identify the optimal band combination, and the effect of each combination was quantitatively evaluated (AIC, R², RMSE, RPD) through the linear fitting of ground LAI and the spectral characteristic index of different band combinations (

Table 5. Summary of the best indices identified for each type of vegetation index, including λ: individual band, SR: simple ratio, D: wavelength difference, ND: normalized difference, and ID: inverse difference. λ1 and λ2 represent the two distinct bands of Sentinel-2 that were used for the calculation.

Index Type	λ1	λ2	AIC	R2	RMSE	RPD
λ	B8A	-	1421.414	0.113	0.427	1.063
SR	B12	B7	501.701	0.576	0.295	1.537
D	B5	B7	1203.041	0.256	0.391	1.160
ND	B7	B12	504.850	0.575	0.296	1.535
ID	B7	B11	590.454	0.545	0.306	1.483

).

SR, ND, and ID provided a better fit for linear regression models, with higher R². In particular, the ground LAI and SR index exhibited the best-fitting results. The optimal combination of bands for this index was identified as B12 and B7, with an AIC of 501.701, R² of 0.576, RMSE of 0.295, and RPD of 1.537. To assess the performance of the developed SR_B12,B7, we further analyzed the relationship between the index and the logarithmically transformed ground LAI (ln (LAI)) over the years 2021 to 2023 (Figure 3). The results indicated that the interannual performance of the SR_B12,B7 index in tracking ground LAI was notably better in 2021 and 2023 compared to 2022. Specifically, the SR_B12,B7 index demonstrated strong performance in tracking LAI in 2021 and 2023 (2021: R² = 0.78, 2023: R² = 0.64). In contrast, its ability to capture LAI variations was limited in 2022, with an R² value of 0.34.

To evaluate the performance of the SR_B12,B7 index in capturing seasonal variations in LAI, we analyzed its performance across different growth stages. The results revealed that, compared to previously reported indices, the index SR_B12,B7 exhibited the highest performance in the Flushing and Senescence stages, with R² values of 0.575 and 0.725, respectively, demonstrating particularly strong performance in the Senescence stage. However, in the Maturity stage, the SR_B12,B7 index did not perform as well as several empirical indices (EXG, MSAVI, SAVI, TVI, CI_red-edge, MCARI, and IRECI etc.). Although the total R² of the SR_B12,B7 index over the three years was the highest, its performance was not the best when considering the leaf stages of each year. The SR_B12,B7 index was almost unable to effectively track the LAI during the Maturity stage, which was characterized by high LAI values.

4. Discussion

4.1. Annual, Seasonal, and Spatial Variability of LAI in Temperate Forests

In this study, the maximum value of LAI was 6, which is slightly higher than previous studies that reported the highest average LAI values for temperate evergreen forests to be between 5.1 and 6.7 [58] and for deciduous forests to be 2 to 5 [59]. Affected by temperature and precipitation, temperate forest ecosystems were highly sensitive to climate change, resulting in substantial annual variation. From a phenological perspective, leaf development in temperate regions exhibited a strong temperature regulation effect [60]. By examining the data reported by the Japan Meteorological Agency, it was observed that the annual precipitation in Nakakawane has been decreasing year by year (2021: 3502 mm, 2022: 3331 mm, 2023: 3126 mm). Additionally, the average daily temperature was the lowest in 2022, followed by 2021, while the highest in 2023. This variation in temperature and precipitation could have contributed to the relatively low value of LAI in 2022.

LAI exhibited a distinct seasonal trend. It began to rise rapidly in May (DOY = 150), reaching its peak between June and August (DOY from 200 to 250) in this research site. These observations were consistent with previous studies, which reported that most broad-leaved tree species experienced a burst of leaf growth in early May, leading to a rapid increase in LAI, with the maximum value typically occurring in mid-July [61]. For tree species such as birch, approximately 47% of their leaves were shed by early August [61], after which LAI gradually declined. However, according to data from the Japan Meteorological Agency, the annual sunshine hours in Nakakawane increased (from 1559.1 h in 2021 to 2177.9 h in 2023), which extended the growing season of temperate forests and caused LAI to exhibit stronger seasonal variations. Although the extension of the growing season likely promoted plant photosynthesis and increased LAI accumulation, the decrease in precipitation and the rise in summer temperatures intensified water stress, which affected plant growth and adaptability [35]. Plants with higher water requirements were likely restricted, causing some species to be at a disadvantage in competition, which in turn affected species abundance and the changes in LAI [62].

Furthermore, our study revealed that LAI exhibited significant spatial autocorrelation across different plots, which was consistent with previous research indicating that LAI showed strong spatial autocorrelation in small-scale, short-distance study areas (92 m) [63]. This phenomenon could be attributed to the spatial heterogeneity of LAI, which was highly sensitive to various biological factors [63]. For example, plant functional types were found to influence this spatial variation [64]. Additionally, other studies indicated that the maximum LAI in temperate forests increased with tree density and was strongly positively correlated with diameter at DBH (R² = 0.94) [35]. Therefore, the factors contributing to the spatial variability of LAI were multifaceted, and a more comprehensive analysis would require integrating additional data.

4.2. DHP-Based LAI Can Be Deviated with Different Processing Approaches

The correlation between the ground LAI estimates obtained using the Hinge and Miller methods showed a high degree of consistency (r = 0.9) (Figure 4). Both methods exhibited strong robustness when applied to different leaf angle distributions and canopy structures. However, the ground-level LAI obtained using the Hinge and Miller methods, as calculated by HemiPy, tended to overestimate the LAI to some extent, as both methods’ PAIe values incorporated the contribution of wood materials [5].

To evaluate the predictive effect of the SR_B12,B7 values obtained by the two methods on LAI, we performed regression analysis on the logarithmic-transformed LAI and predicted LAI (Figure 5), respectively. The results showed that the Hinge method exhibited slightly lower R² and RMSE than the Miller method (Hinge: R² = 0.576, RMSE = 0.295; Miller: R² = 0.651, RMSE = 0.302). This finding aligned with previous studies, which also reported that the Hinge method produced more accurate predictions compared to the Miller method [16], although it had a lower R². When converting PAI derived from DHP into LAI, the performance of the Hinge method showed the highest consistency with litter-based observations, as demonstrated in the comparison of litter and optical LAI [16]. The LAI estimates derived using the Hinge method may therefore be more applicable than those obtained using the Miller method.

4.3. Developed VIs vs. Reported VIs

Among all combinations, the combination of B12 (shortwave infrared band, SWIR band) and B7 (red-edge band, RE band) was identified as the most effective for estimating the LAI in our study. Previously reported VIs mainly used visible light bands (especially red, green, and blue bands) as well as near-infrared bands, such as traditional NDVI, EVI, EXG, EXR, EXGR, and other indices [38,65]. The new spectral index developed in this study was based on the B7 band, one of the three red-edge bands unique to Sentinel-2, along with the SWIR band. The RE band was sensitive to chlorophyll, LAI, and canopy structure, thus capturing the seasonal dynamics of vegetation [27]. Certain leaf components, such as cellulose and lignin, absorbed and scattered light, resulting in changes in the reflectance spectra of the vegetation. On the other hand, the SWIR band was found to be effective in discriminating between green vegetation, senescing vegetation, and bare soil [66]. Early research had shown that the NIR and RE regions were sensitive to LAI, and in the context of temperate forest ecosystems, the RE index performed better than the NIR index for LAI estimation [6]. Additionally, the SWIR band had proven effective for classifying temperate tree species [67], and thus, reflectance data from these regions had been integrated into various VIs [68]. However, it should be noted that these indices exhibited greater sensitivity to lower LAI values. As LAI increased beyond 2–3, reflectance became less responsive to changes in LAI, particularly in the visible and RE bands [69]. This aligns with findings indicating significant physiological changes in vegetation during the Flushing and Senescence stages, while tracking LAI in the Maturity stage became increasingly challenging, as evidenced by a low R² value (R² ≤ 0.03). During the maturity phase, chlorophyll absorption became saturated at higher LAI values [70], and plants’ capacity to absorb light became maximal, resulting in minimal changes in canopy reflectance [71]. Consequently, it becomes difficult to accurately track LAI during the leaf maturation period using vegetation spectral indices alone.

The developed SR_B12,B7 index can track the overall LAI annually and seasonally. However, the performance of the SR_B12,B7 index in tracking LAI over different years was found to have significant year-to-year variability. This indicates that the robustness and consistency of the index may be limited, potentially reducing its reliability for long-term monitoring applications. To further investigate the performance of the developed vegetation index in tracking the temporal dynamics of LAI, we separated the data by leaf phenological stage. Our analysis revealed that this novel index showed superior performance in tracking LAI during the leaf budding and senescence stages compared to the maturity stage. This observation is consistent with the performance of the other reported empirical indices (Table 4).

After comparing the obtained VI with those reported in previous studies using the SR_B12,B7 index, it was found that the SR_B12,B7 index performed better than the same type of index reported in previous studies [7,8,22,57]. During the period from leaf bud to canopy closure, the SR_B12,B7 index demonstrated superior performance compared to the previously reported EVIre2 and NDVIre3, both of which also utilized the red-edge band. However, in comparison to the NDVI-RE reported by Tillack, the SR_B12,B7 index exhibited a slightly lower performance (R² = 0.62) [8,21]. During the period of Maturity, both our study and previous research obtained very low R² values and high RMSE values, consistent with previous observations that there was almost no correlation between NDVI and LAI during this period, as saturation occurred when LAI ≥ 5 [8]. During the leaf senescence period, the performance of the SR_B12,B7 index in 2021 and 2023 exceeded that of the mSR-RE (R² = 0.408) and other indices [8]. Over the entire year, the performance of the SR_B12,B7 index (R² max = 0.821) also outperformed other previously reported indices (NDVI: r = 0.88, SAVI: r =0.82, EVI: r = 0.82, ARVI: r = 0.87) [72].

In the context of spatial scale, temperate deciduous forests typically encompass diverse tree species, canopy densities, and understory vegetation types. This heterogeneous environment complicated the tracking of LAI [73]. The spectral reflectance characteristics of the forest floor, which included shrubs and other low vegetation, often resembled those of the deeper canopy leaves in certain spectral bands [73]. This similarity in reflectance could explain the differences in spatial heterogeneity in LAI tracking by the SR_B12,B7. In addition to the underlying surface vegetation, the tracking effect of the SR_B12,B7 index on LAI was also influenced by the slope (Figure 6). In this study, the SR_B12,B7 index demonstrated the most effective tracking of LAI in areas with slopes between 20° and 25°, achieving an R² of 0.598 and an RMSE of 0.219. Previous studies also reported vegetation and soil conditions were generally better in the medium slope range [74]. Steep slopes led to a higher risk of soil and water loss, while areas with shallow slopes were prone to waterlogging. Moreover, topography introduced certain errors in LAI estimation, with the magnitude of these errors increasing with increasing slope within a certain range [28]. For example, in areas with slopes, greater than 15°, seasonal deviations in NDVI and EVI increased by 7.7% and 4.7%, respectively, while annual deviations increased by 5.1% and 8.4%, respectively [75]. Furthermore, reflectance in the SWIR band was highly sensitive to terrain variations, which may explain why the SR_B12,B7 index had a limited ability to track ground LAI in Nakakawane [28].

In addition to topographic factors, other site-specific conditions, such as understory vegetation and dominant tree species, can significantly influence LAI estimation. For instance, in plot 28 and plot 11, the presence of understory vegetation, in conjunction with the contribution of woody materials, influenced the reflectance of both the RE and NIR bands. As a result, the reflectance values of these bands may have been contaminated by the understory vegetation, preventing an accurate differentiation of the contributions from the canopy and ground vegetation, and potentially leading to an overestimation of the LAI [6]. Additionally, the dominant tree species in the given plot may have influenced the monitoring of LAI. A previous study reported that the SWIR band was more effective in discriminating between vegetation types [67], such as plots that are dominated by Quercus spp.

4.4. Limitation of VIs and Future Studies

While the SR_B12,B7 index performs better than other indices in certain years, it, nevertheless, is not sufficient to provide consistent inter-annual tracking when applied to time-series analysis. In addition, there are significant spatial differences across different plots. Although the SR_B12,B7 index performs slightly better than other indices in monitoring LAI, this trend is not consistent across all plots. Furthermore, this study used a simple linear regression model for LAI prediction, with R² values of 0.576 (Hinge) and 0.651 (Miller), and RMSE values of 0.295 (Hinge) and 0.382 (Miller). Previous studies have shown that combining color and texture indices with machine learning techniques can improve model prediction accuracy applied to estimate LAI during the crop growing season [37,76]. Adding texture features into the relationship between LAI and NDVI can improve the accuracy of LAI estimation by about 20% [77], but incorporating both color index and texture index into the random forest regression model was found to be more accurate than using only the spectral index to estimate LAI (R² = 0.85, RMSE = 0.56, RPD = 2.52) [7,22]. In addition, previous studies have demonstrated that the LAI back propagation neural network (BPNN) model based on CIs and texture features (R²  = 0.730, RMSE = 0.691, RPD = 1.927) outperforms other models [77]. These provide a direction for improving the accuracy of LAI estimation. Consequently, future research should explore integrating drone imagery with other methods, incorporating color and texture indices into the model, and applying machine learning techniques to LAI inversion. This approach may allow for more accurate, high-resolution tracking of LAI in temperate deciduous forests across both temporal and spatial scales.

5. Conclusions

In this study, currently reported vegetation indices were evaluated for their ability to effectively monitor the spatial-temporal changes of the LAI in temperate deciduous forests by taking advantage of the Sentinel-2 data. In parallel, a spectral index unique to Sentinel-2 data was developed by combining both SWIR and RE bands. Although the reported indices generally performed poorly, the newly developed SR_B12,B7 index can successfully track LAI at both seasonal and spatial scales. It performed particularly well during the Flushing and Senescence stages, which showed a high correlation with the ground-based LAI. Nevertheless, during maturity, when LAI peaks, the index showed some limitations in capturing the temporal and spatial dynamics of LAI in temperate deciduous forests, requiring further research in this direction.