1. Introduction
Leaf area index (LAI), or the total one-sided leaf area per unit of ground area (m
2 leaf per m
2 surface or dimensionless), can be distinguished in two types. On the one hand, there is the green leaf area index (LAI
green), representing the leaves which are photosynthetically active, being the most common type of LAI [
1], and, on the other hand, there is the brown leaf area index (LAI
brown), representing the leaf area normalized which is senescent and losing photosynthetic function [
2]. The Sentinel-2 mission from the European Space Agency (ESA) has, with the improved optical sensor bands in the red-edge, an increased sensitivity towards LAI
green [
2], while the shortwave infrared bands are sensitive to cellulose and lignin (dry matter) absorption [
2]. Such improved capabilities to obtain more accurate quantifications of LAI
green over large areas provides an important aspect in climatic [
3], ecological [
4] and biogeochemical [
5] cycles models, as well as for estimating crop vegetation status [
6], developing soil maps [
7] and estimating light-use efficiency [
8]. Its determination is crucial for the understanding of biophysical processes of crop canopies, being the main morphological parameter used for determining crop growth through the correlation with crop productivity [
1,
9,
10]. In the context of agricultural monitoring, there is a strong interest in estimate LAI
green parameter. Near real-time LAI
green estimates provides the tool for farmers to obtain the crop health and growth status, further improving the effective technical support in farming practices such as fertilizer application and water management. In this way, increased crop yields and reduced costs and input resources for the agricultural sector are envisaged [
11,
12]. Remote sensing from satellite, aerial and unmanned aerial vehicle platforms has become a popular technique in monitoring crop LAI
green because of its ability to acquire synoptic information at different times and spatial scales [
13,
14,
15]. For agricultural monitoring by remote sensing, the spatial resolution should be at least 20 m and, preferably, 10 m in order to make site-specific management possible [
16]. A temporal resolution of less than a week would be required to follow-up acute changes in crop condition and provide timely response in management practices. These requirements are fulfilled by the ESA’s Sentinel-2 mission, providing 10 m pixel size products with a 10-day temporal resolution. Sentinel-2 is a polar-orbiting, superspectral high-resolution imaging mission with twin polar-orbiting satellites, Sentinel-2A and 2B. The mission’s main objective is providing quality information for agricultural and forestry practices and, hence, helping manage food security [
17]. With Sentinel-2A in orbit (launched 23 June 2015), the temporal resolution was not yet sufficient for real applications at the individual farmer’s level. But with the additional availability of data from Sentinel-2B (launched 7 March 2017) the revisit period goes down to five days under cloud-free conditions.
LAI
green is functionally linked to the canopy spectral reflectance, so its retrieval from optical remote-sensing data has prompted many studies using various techniques [
18,
19]. Essentially, these retrieval techniques can be classified into two groups, i.e., (1) empirical retrieval methods, which typically consist of relating the biophysical parameter of interest against spectral data through linear (e.g., vegetation indices) or nonlinear (e.g., machine learning approaches) regression techniques [
20,
21,
22,
23] and (2) physically-based retrieval methods, which refers to inversion of radiative transfer models (RTMs) against remote sensing observations [
24,
25,
26]. Concerning physical models, experimental studies using RTMs have shown great flexibility in retrieving plant cover variables, because of being able to parameterize these models to a wide range of land cover situations and sensor configurations [
27,
28]. However, two main drawbacks limit the use of the inversion of RTMs for operational applications. First, RTM approaches typically require some ancillary information to enable the parameterization of the physical model, which may not always be available [
13,
29]. An additional problem hereby is that if uncertainties are introduced the likelihood increases that the model inversion will not lead to a unique solution and extra steps are required to overcome the ill-posed problem [
30]. Second, regardless of the availability of auxiliary data, there is the intrinsic risk of oversimplifying the architecture of canopy for those RTMs fast enough for operational applications. The difficulty in describing canopy structure increases in heterogeneous scenes, such as mosaics of crops at different phenological stages or complex mixtures of woodlands and/or grasslands [
2,
31,
32]. Non-linear regression techniques are standardly used for operational LAI products. For Sentinel-2 an operational LAI product, associated with a quality indicator, is provided through the SNAP (Sentinel Application Platform) toolbox and produced through a neural network which has been trained by simulated spectra generated from well-known RTMs [
33]. The algorithm is trained with simulated LAI
green values generated from the SAIL radiative transfer model [
34], which describes the canopy as a homogenous and horizontal turbid-medium, and the PROSPECT radiative transfer model [
35], which considers the leaf as a succession of absorption layers. However, the accuracy of this product is shown improvable [
36]. Other machine learning algorithms than neural networks have been proposed to study the retrieval opportunities of LAI from Sentinel-2 and -3 [
37], solving the black box problem. However, although machine learning approaches can be fast and can capture the non-linear relationship between different parameters, they are time variant and location dependent [
38].
Alternatively, linear empirical models, i.e., vegetation indices (VIs), are one of the most straightforward implementable method in an operational data processing chain. These indices relate a few spectral bands with the biophysical parameter of interest [
39] in a way that enhances the spectral characteristics of a given vegetation property while minimizing the soil, atmospheric, and sun-target-sensor geometry effects [
22]. Despite the positive aspects of VIs developed for LAI retrieval, their major weakness is the lack of a generally applicable index for multiple vegetation types. The best way to find efficient and robust indices is to use large and diverse field datasets, with a large variety of canopy structures [
22,
40]. Early studies identified the red and near-infrared (NIR) regions as sensitive to LAI
green, resulting in the common use of the reflectance broad-bands in these regions through simple ratios [
41] or normalized difference ratios [
42,
43]. It should be mentioned that while these indices were found to be sensitive to low LAI
green values
, they usually lose sensitivity as LAI
green increases (typically above 2–3 according to Haboudane et al. [
19]). This saturation of the reflectance at moderate to high LAI
green values in the red range (600–700 nm) is due to the high chlorophyll absorption in this spectral range [
9]. The wavelength region located in the visible–near infrared (VIS-NIR) transition, i.e., between 690 and 750 nm, generally referred as the “red-edge”, is the region between maximum chlorophyll absorption in the red, and maximum reflection (high scattering) in the NIR caused by leaf cellular structure abundance, i.e., LAI [
44,
45,
46]. Reflectance in the red-edge transition region is much higher than in the visible range especially for these moderate to high LAI
green values, where the upwelling radiance in the red-edge range provides a higher and less noisy signal compared to the low values in the red region. It has been specifically demonstrated, through real [
47,
48] and simulated spectral data [
42,
46], that the shape of the red-edge region and mainly the slope is strongly influenced by chlorophyll density and, hence, by LAI
green. Despite this well-known sensitivity, practically no established indices use the red-edge region for the LAI
green retrieval as until now no free operational satellites had narrow-bands in this region. With the Sentinel-2 satellites (13 spectral bands) not only optimal and temporal resolution for crop monitoring is guaranteed, but, moreover, also spectral configuration in the red-edge is improved, with narrow-bands centred at 705 nm (B5) and 740 nm (B6). Recent studies have explored the potential of Sentinel-2 for the LAI
green retrieval based on simulated datasets [
49,
50]. But at this moment, few studies have used real Sentinel-2 images in combination with in situ datasets for agricultural applications. Moreover, these studies using the red-edge Sentinel-2 bands for LAI
green retrieval, calibrated and validated their products for only a few crop types [
51,
52], leaving the robustness of a generic retrieval application still an open issue.
In this respect, we aim to develop a simple, accurate empirical algorithm for deriving LAIgreen from Sentinel-2 real data of multi-crop agricultural fields, using two large in situ field datasets. The first objective is to determine if the commonly used VIs for estimating LAIgreen may be applicable for a variety of crop types. Secondly, we want to identify the Sentinel-2 spectral bands that present the highest correlation for the estimation of a wide variation in crop LAIgreen. Based on this analysis and on a parallel study of the importance of the new Sentinel-2 red-edge bands, a new robust LAIgreen index is defined. The performance of the new index and established VIs indices are validated and applied over two distinct agricultural test sites.
4. Discussion
With the availability of a narrow band in the red-edge region by Sentinel-2, an improved and simple estimation of LAI
green based on a simple index becomes possible at high spatial resolution. The proposed SeLI index shows a significant improvement towards indices using the saturating bands in the red (B4 in Sentinel-2). Moreover, no saturation appeared in the obtained LAI
green product based on the red-edge bands (B5: 705 nm and B6: 740 nm). The B4 (665 nm) saturation at high LAI
green values is clearly shown with real Sentinel-2 TOC reflectance spectra for different LAI
green values, higher than 2 (
Figure 5 and
Figure 6). The red-edge bands (B5: 705 nm and B6: 740 nm) in contrast are both affected by higher scattering, whereby the B5 band is still driven by chlorophyll absorption. This agrees with numerous authors who emphasize the importance of the red-edge bands for the estimation of biophysical parameters, mainly the LAI
green and chlorophyll estimation [
42,
46,
71]. The proposed Sentinel-2 LAI Index (SeLI) exploits the B5 red-edge band, which has been widely demonstrated that is highly influenced by the LAI
green parameter [
46,
47,
48], and the B8a NIR band, which is driven by the scattering changes in moderate-to-high LAI values in crops [
72]. Very few previous indices have used bands in the red-edge region because no free operational previous sensors had narrow bands in this spectral area [
42]. Both linear and non-linear empirical regression techniques have been tested for the LAI retrieval on simulated spectrally resampled airborne data [
50,
73] and recently on real Sentinel-2 data [
10,
52]. The band selection obtained from these methodologies appeared to favor (1) green and SWIR bands in the case of linear regression by VIs, and (2) red, NIR and SWIR bands in the case of non-linear regression by machine learning approaches [
74]. These bands were also chosen by several of our tested VIs (
Table 5), with the difference that the NDGI formulation indicated the use of red-edge (705 nm) and a NIR band (865 nm) as best band selection.
The robustness and generality of the SeLI index is demonstrated by applying it to an independent in situ field dataset from a distinct geographical location with crop types different from those included in the testing dataset, obtaining equally good statistics (R
2 of 0.732, RMSE of 0.69). Specifically, SeLI does not present problems of saturation when it is applied to a multi-crop Valencia in situ dataset composed of 13 different crop types and LAI
green values that go up to 4.5, obtaining R
2 of 0.708 and RMSE of 0.67. Furthermore, when SeLI is applied to the Foggia region, characterized by high LAI
green values, the limits of the different crop fields and LAI
green variability within the crop field appears even for the high value ranges, indicating variable growing conditions. Such a clear distinction in LAI
green variability allows evaluation of management practices at the field level. Hence, it is shown that the SeLI index generally can be applied for LAI
green retrieval of different crop types and distinct areas. A limitation of the index is that it has been calibrated and validated with LAI data up to 5, so it is only applicable to agricultural areas with this range of values, although it is the common range of in situ LAI measured values in a lot of studies with a great variety of crop types; such as wheat [
75,
76], corn [
19], potato [
10] and sugar beet [
2,
47]. Currently, ongoing scientific debate is taking place on the discussion if there is a linear relationship or not between in situ LAI values and estimated values [
77]. Our study shows that in situ data are linearly related to SeLI, in the value range of 0 to 5. This result is in accordance with other results in which also LAI of agricultural areas is estimated through indices and linear models in similar ranges [
10,
75,
78,
79,
80]. Generally, no values higher than 5 are used in these studies constraining the model applicability to this range. As SeLI has a physiological foundation, the index will be applicable to a higher value range, but the SeLI-LAI fitting relationship might change depending on the dataset range. To verify this, a LAI in situ data range >5 would be required. However, one must also consider the instrumental limitations for in situ measurements. The LAI-2200 Plant Canopy Analyzer [
55] calculates LAI by comparing differential light measurements above and below canopy. The maximum measurable LAI is generally lower for these devices measuring gap fraction with LAI reaching an asymptotic saturation level at a value of about 5, compared to that assessed via destructive methods. The cause for this is gap fraction saturation as LAI approaches five or six [
81,
82,
83].
In this work, the Sentinel-2 LAI
green product obtained from the SNAP toolbox was also tested with the multi-crop dataset from Valencia. The results show underestimated LAI estimations (R
2 = 0.475, RMSE = 0.91). There are some studies, which have also compared this Sentinel-2 LAI
green product with in situ LAI
green crop data, that obtain better R
2 [
84,
85], but they only tested the product with few crop types. When the product is analysed in different areas and plant species, the results can be improved [
36]. This finding could be explained by the fact that the Sentinel-2 algorithm used for land surface parameters, including LAI
green product, ingests almost all spectral bands and applies a nonlinear regression to estimate each parameter [
33], in addition to the fact that it has been proven that there is a substantial sensitivity of Sentinel-2 biophysical products to the implemented rugged terrain corrections [
36].
The other main challenge in the retrieval of biophysical parameters with vegetation indices is the difficulty of finding a simple index with such a general character that it can estimate the parameter of a wide variety of crop types. In this work it has also been shown that the established indices do not present this general character. This may be because they were developed and calibrated based on limited experimental data in terms of species, presenting improvable statistics (R
2 between 0.234–0.663) when applied to multi-crop datasets. In an attempt to improve estimations over this multi-crop dataset, all band combinations were systematically calculated for each index in order to achieve the highest possible correlation for the estimation of LAI
green. More promising results were obtained, with a R
2 between 0.701 and 0.737. However, when inspecting these sensitive bands whether they are physically meaningful, i.e., if the selected bands are actually influenced only or mostly by LAI
green, then these indices turned out to be questionable. In the majority of cases, the selected bands were influenced by leaf constituents such as lignin, cellulose and water (e.g., 1610 nm, 2190 nm) affecting the scattering properties in the NIR and SWIR [
69,
70], and being less related to photosynthetically based LAI
green. At the field or landscape scale, canopy reflectance patterns represent the integrated effects of all biophysical parameters. Co-variation mechanisms of leaf constituents is typically causing the selection of bands related to other covarying biochemicals such as pigments or lignin due to their high effect on spectral variability [
86]. Similarly, it was earlier observed that due to the covariation between water content and chlorophyll content (related with LAI parameter), typically bands in the water content absorption region are selected as most sensitive [
87]. To improve the estimation of LAI
green (aside from LAI
brown), bands only affected by structural leaf components should be omitted. Structurally-related NIR and SWIR bands may improve the LAI
green retrieval when the model is trained on healthy vegetation [
74] but may be less generally applicable for scenarios with different structural types or stress conditions. With the band selection B5 and B8a, SeLI is functionally related to green LAI, avoiding absorption saturation in the red region.
It should be mentioned that this is the first time that this kind of LAI
green retrieval can be carried out for agricultural areas with plots sizes of only 40–100 m, such as Huerta of Valencia, due to the lack of an operational satellite with the required spatial and temporal resolution. ESA’s satellite Sentinel-2 aims to replace and improve the older generation of satellite sensors such as Landsat and SPOT, with improved spectral and spatial capabilities. Therefore, the Sentinel-2 satellites provide a great opportunity for global vegetation monitoring, and specifically crop field monitoring, due to its enhanced spatial, spectral and temporal characteristics [
42].
Finally, further validation is required with other field campaigns and synthetic Sentinel-2 data to reinforce findings. Considering appropriate instrumental tools, the index behavior for LAIgreen values higher than 5 should be tested, as well as the fitting behavior of these further ranges.