Spaceborne Estimation of Leaf Area Index in Cotton, Tomato, and Wheat Using Sentinel-2

: Satellite remote sensing is a useful tool for estimating crop variables, particularly Leaf Area Index (LAI), which plays a pivotal role in monitoring crop development. The goal of this study was to identify the optimal Sentinel-2 bands for LAI estimation and to derive Vegetation Indices (VI) that are well correlated with LAI. Linear regression models between time series of Sentinel-2 imagery and ﬁeld-measured LAI showed that Sentinel-2 Band-8A—Narrow Near InfraRed (NIR) is more accurate for LAI estimation than the traditionally used Band-8 (NIR). Band-5 (Red edge-1) showed the lowest performance out of all red edge bands in tomato and cotton. A novel ﬁnding was that Band 9 (Water vapor) showed a very high correlation with LAI. Bands 1, 2, 3, 4, 5, 11, and 12 were saturated at LAI ≈ 3 in cotton and tomato. Bands 6, 7, 8, 8A, and 9 were not saturated at high LAI values in cotton and tomato. The tomato, cotton, and wheat LAI estimation performance of ReNDVI (R 2 = 0.79, 0.98, 0.83, respectively) and two new VIs (WEVI (Water vapor red Edge Vegetation Index) (R 2 = 0.81, 0.96, 0.71, respectively) and WNEVI (Water vapor narrow NIR red Edge Vegetation index) (R 2 = 0.79, 0.98, 0.79, respectively)) were higher than the LAI estimation performance of the commonly used NDVI (R 2 = 0.66, 0.83, 0.05, respectively) and other common VIs tested in this study. Consequently, reNDVI, WEVI, and WNEVI can facilitate more accurate agricultural monitoring than traditional VIs.


Introduction
Monitoring crop growth and performance during developmental stages is an essential aspect of agricultural management. Leaf Area Index (LAI) is a good proxy of the vegetation state [1][2][3] and a good yield predictor [4][5][6]. LAI is a dimensionless quantity that characterizes plant canopies. It is defined as the one-sided green leaf area per unit ground surface area. The LAI is an important parameter in plant ecology and a measure of the photosynthetic active area, and at the same time of the area subjected to transpiration. It is also the area that comes in contact with air pollutants. LAI is often a key biophysical variable used in biogeochemical, hydrological, and ecological models. LAI is also commonly used as a measure of crop growth and productivity at spatial scales ranging from the plot to the globe. Moreover, activities such as herbicide and fertiliser management, leaf chlorophyll content estimation, detection of crop disease, and yield prediction can be based on LAI monitoring [7].
LAI can be estimated from VIs [8][9][10][11] produced from imagery acquired by optical satellites, but this approach suffers from a low correlation between LAI and some bands that the VIs are based on. Many studies showed that LAI estimation from optical imagery suffers from saturation when LAI is greater than 3 (i.e., the LAI changes at a faster rate than the reflectance) [11][12][13][14]. Since the LAI of many crops typically exceeds this level by a large margin, optical sensors have limited use for LAI estimation. Most previous studies that defined this saturation effect were based on older sensors (e.g., Landsat, Modis, SPOT) [15][16][17], and accordingly, Vegetation Indices (VIs) intended for those sensors. In 2015 the first Sentinel-2 became operational, which marked the arrival of the new generation of satellites. The Multi-Spectral Instrument (MSI) onboard Sentinel-2 observes the earth at 13 spectral bands with a spatial resolution from 10 to 60 m (depending on the band) and a five-day revisit time. MSI is a spaceborne multispectral instrument that thoroughly covers the red edge spectral range, which is highly sensitive to the chlorophyll reflectance in plants [18]. The red-edge spectral range covers the wavelengths of 680-750 nm, where the change of leaf reflectance is sharp [19,20]. In order to estimate LAI from Sentinel-2, there is a need to evaluate which bands suffer from the saturation that was observed in previous generations of spaceborne sensors and explore ways to overcome this limitation.
In addition to LAI modelling based on VIs, several machine learning algorithms for LAI estimation based on Sentinel-2 bands were studied and showed mixed results [11,[21][22][23]. Previous studies on different wavebands [24], including simulated Sentinel-2 bands, concluded that the red edge is the best spectral region for LAI estimation in several crops [2,3,[25][26][27]. Therefore, careful selection of the bands used to derive VIs and machine learning algorithms can improve the performance and generality of the LAI estimation models based on Sentinel-2 imagery. Nevertheless, while several studies investigated the performance of MSI-based VIs and machine learning algorithms for LAI estimation of tomato, wheat, and cotton [11,[28][29][30], very few studies investigated the performance of the real MSI bands (as opposed to synthetic data) in the LAI estimation of these crops [31].
Therefore, this study's first goal was to model LAI using real Sentinel-2 imagery and field-measured LAI to quantify the performance of individual bands and their saturation levels in cotton, tomato and wheat. The second goal of the study was to suggest wellperforming VIs that employ bands not commonly used for VI derivation and facilitate better agricultural monitoring.

Test Sites and Field Measurements
The field data used in this study were collected during one cotton, two wheat, and three processing tomatoes experiments conducted in five locations in Israel ( Figure 1). The Sentinel-2 image inventory used for this study is presented in Table 1. Overall, 56 Sentinel-2 images were used in the study (14-wheat, 33-processing tomatoes, 9cotton). During these experiments, LAI was measured by a SunScan Canopy Analysis System-SS1 developed by Delta-T Company (Cambridge, United Kingdom). The Sun-Scan is a widely used, accurate, nondestructive LAI measurement system successfully em- The Sentinel-2 image inventory used for this study is presented in Table 1. Overall, 56 Sentinel-2 images were used in the study (14-wheat, 33-processing tomatoes, 9-cotton). During these experiments, LAI was measured by a SunScan Canopy Analysis System-SS1 developed by Delta-T Company (Cambridge, United Kingdom). The SunScan is a widely used, accurate, nondestructive LAI measurement system successfully employed in many previous studies [5,30,32]. The SunScan system measures LAI by calculating the difference in solar radiance received by the dome sensor installed under unobscured Sun view and the hand-held probe placed below vegetation canopy on the ground level ( Figure 2).  The Sentinel-2 image inventory used for this study is presented in Table 1. Overall, 56 Sentinel-2 images were used in the study (14-wheat, 33-processing tomatoes, 9cotton). During these experiments, LAI was measured by a SunScan Canopy Analysis System-SS1 developed by Delta-T Company (Cambridge, United Kingdom). The Sun-Scan is a widely used, accurate, nondestructive LAI measurement system successfully employed in many previous studies [5,30,32]. The SunScan system measures LAI by calculating the difference in solar radiance received by the dome sensor installed under unobscured Sun view and the hand-held probe placed below vegetation canopy on the ground level ( Figure 2).  Each LAI value used for model calibration was an average value of at least 30 field measurements. LAI was measured in the center of the fields and was correlated to average values of Sentinel-2 bands and VIs of homogenous areas in the fields' centers. In the Megido 2020 experiment, LAI was measured in two areas of the field (in the center of the field (six LAI measurements) and on the northwest corner (four LAI measurements)) where the crop developed at different rates and, thus, LAI was different. Accordingly, both time series of the field measurements were correlated with the average values of bands and VIs within defined polygons. In-field paths and their surrounding area were masked out from analysis polygons of the tomato experiments to remove bare soil areas and avoid border effects. These excluded areas consisted of approximately 20% of the overall polygon areas in the tomato fields. Therefore, each tomato polygon consisted of either two or four vegetated regions separated by the paths.
Overall, 11 averaged LAI values taken during two growing seasons were used for deriving the wheat models, nine for cotton (one season), and 23 for tomato (three seasons).
The linear regression models in this study utilised the average values derived from satellite imagery within the analysis polygons and same-date field measurements or linearly interpolated LAI values of field measurements from adjacent dates.

Satellite Imagery
Sentinel-2 is an Earth observation mission and part of the European Space Agency (ESA) Copernicus program. It includes two satellites with a payload of MSI, namely Sentinel-2A (launched 23 June 2015) and Sentinel-2B (launched 7 March 2017). Table 1 lists the spectral bands of Sentinel-2 that were used in this study. The inventory of the atmospherically and topographically corrected Level-2A Sentinel-2 images used in this study alongside the information on the LAI measurements can be found in Tables 2 and A1. Level-2A and Level-1C imagery were downloaded from the ESA Copernicus site (https: //scihub.copernicus.eu/dhus/#/home, accessed on 6 April 2021) (# means "Number"). Level-1C images were processed to Level-2A using Sen2Cor algorithm [33]. Table 2. Sentinel-2 imagery and LAI measurements used in the study.

Area
Crop Period * # of Images Polygon Size (Sentinel-2 Pixels)

Saad
Wheat

Model Calibration and Validation
Linear regression models were derived to estimate LAI for specific crops based on field measurements and Sentinel-2 bands. Similarly, regression models between LAI and VIs were derived, including NDVI [34] and NDVI based on the Narrow NIR Band-8A instead of NIR Band-8. Additionally to NDVI, models were also derived for reNDVI [35], MTCI [36], WDVI [37], EVI [38], SAVI [39], MSAVI [40], DVI [34], and two new indices: WEVI (Water vapor red Edge Vegetation Index) and WNEVI (Water vapor narrow NIR red Edge Vegetation index). For every model, the R 2 and root mean square error (RMSE) values were calculated using the Microsoft Excel software. WEVI and WNEVI were developed based on combinations of the best performing bands for LAI estimation. The following equations and Sentinel-2 bands were used for deriving the aforementioned VIs: where: L = 1 − 2 × s × NDVI × WDVI and s is the soil line slope = 0.5 3. Results Table 3 shows the performance of the separate Sentinel-2 bands and VIs for LAI estimation of cotton, tomato, and wheat. Overall, the bands that modelled LAI best were Band-7 (Red edge-3), Band-9 (Water vapor), and two NIR bands (8 and 8A). Notably, Band-8A (Narrow NIR) showed a higher correlation with LAI and lower RMSE in LAI estimation than Band-8 (NIR) in all three crops. Consequently, NDVI8A performed better than NDVI. Importantly, Band-4 (Red) showed average performance, and Band-5 (Red edge-1) showed weak performance relative to other bands in tomato and cotton. Therefore VIs based on the better performing bands might be beneficial for LAI estimation. One such VI, namely reNDVI, showed a very high estimation performance. Finally, the high performance in LAI prediction by the Water vapor Band-9 suggests that this band might be useful for creating VIs with good correlation to LAI. This result was confirmed by low RMSE and high R 2 values of the new WEVI and WNEVI that are based on Band-9. The two new VIs proposed in the study (WEVI and WNEVI) alongside reNDVI showed superior performance in LAI predictions compared to NDVI and NDVI8A in all three crops, with the largest difference in wheat.  Figure 3 shows the reflectance in each band and the corresponding LAI measurements in this study's experiments. The reflectance in bands 1, 2, 3, 4, 5, 11, 12 in cotton and processing tomatoes start saturating from LAI ≈ 3 and almost no longer changing at LAI ≈ 6. This result is especially important because bands 4 and 5 are used in many VIs. On the other hand, bands 6, 7, 8, 8A, 9 were not saturated. Insufficient satellite imagery and field measurements of LAI were acquired during the wheat experiments and hindered estimating the saturation levels of this crop.      Figure 5 shows reNDVI, WEVI, WNEVI, NDVI, and MTCI linear regression models for tomato, cotton, and wheat. Figure 6 shows the LAI measurements and LAI estimation based on the VIs used in this study using the models described in Table 3. While several VIs showed similar behavior in LAI estimation, MTCI, MSAVI, reNDVI, WEVI, and WNEVI were notably different. MTCI, affected by the low performance of the Band-5, did not perform well in tomato LAI estimation in Gadot 2019 and 2020. MSAVI notably underestimated wheat LAI values. Conversely, reNDVI, WEVI, and WNEVI show closer resemblance to measured LAI than all other VIs. In the present study, no difference in the spectral response of Sentinel-2A and -B satellites was observed owing to an excellent radiometric cross-calibration of the MSI on both satellites.  Figure 5 shows reNDVI, WEVI, WNEVI, NDVI, and MTCI linear regression models for tomato, cotton, and wheat. Figure 6 shows the LAI measurements and LAI estimation based on the VIs used in this study using the models described in Table 3. While several VIs showed similar behavior in LAI estimation, MTCI, MSAVI, reNDVI, WEVI, and WNEVI were notably different. MTCI, affected by the low performance of the Band-5, did not perform well in tomato LAI estimation in Gadot 2019 and 2020. MSAVI notably underestimated wheat LAI values. Conversely, reNDVI, WEVI, and WNEVI show closer resemblance to measured LAI than all other VIs. In the present study, no difference in the spectral response of Sentinel-2A and -B satellites was observed owing to an excellent radiometric cross-calibration of the MSI on both satellites.  Table 2, the RMSE values of the models are given in Table 3.  Table 2, the RMSE values of the models are given in Table 3.

Discussion
This study investigated the performance of the individual Sentinel-2 bands and VIs in estimating LAI of tomato, cotton, and wheat. This study's most important finding is that bands 6, 7, 8, 8A, 9 performed well in LAI estimation and did not saturate at high LAI in cotton and processing tomatoes. At the same time, the wheat data was insufficient to make this determination. Therefore, these bands can be used to create VIs for LAI monitoring. VIs such as reNDVI and two new VIs introduced in this study for the first time, WEVI and WNEVI, which are based on these bands, performed well in LAI estimation, better than the commonly used NDVI as well as all the other VIs used in the study.
Band-8A (Narrow NIR) showed better performance in LAI estimation compared to Band-8 (NIR). Therefore, NDVI derived based on Band-8A performed better than NDVI based on Band-8. Band-4 (Red) was found to have an average performance. Therefore, substituting Band-8 with Band-8A and possibly substituting Band-4 with a better-performing band (such as Band-6 used in reNDVI) is likely to improve the correlation of VIs with LAI, and facilitate more accurate agricultural monitoring. The high performance of the reNDVI achieved in the study supported this hypothesis. Unlike red edge and NIR bands, Band-9 (Water vapor) is not commonly used as a VI formulae but can be used in VIs such as WEVI and WNEVI developed in this study. The analysis of Band-9 performance, which is not commonly used for agricultural monitoring, and developing VIs based on this band that perform well in LAI estimation of the three crops, is the unique feature of the present study.
Unlike red edge bands 6 and 7 that showed high performance, Band-5 (Red edge-1), at the tail of the chlorophyll absorption peak [41], showed the lowest overall performance out of all the red edge bands. This might be explained by the negative effect of the chlorophyll content present in the leaves [10,14,[42][43][44], which reaches maximum absorbance at about 690 nm [45]. Moreover, chlorophyll content may vary independently from LAI [46]. In this study, MTCI, based on Band-5, showed low performance in tomato LAI estimation. MTCI was previously found to have low correlation with tomato crop coefficient (K c ) and height [11]. Nevertheless, MTCI was highly correlated with LAI of cotton and wheat in the present study. MTCI was also previously found to have very high correlation with cotton K c [47,48] as well as a very good correlation with leaf chlorophyll concentration [25,49] and LAI of many crops [3,23,50]. Consequently, despite its effective use for crop variable estimation in many cases, Band-5 and VIs based on this band (e.g., MTCI) should be used with caution to model tomato variables. Similarly, careful selection of Sentinel-2 bands might improve the performance of various machine learning algorithms, for example, the SNAP Biophysical processor [51].
The results and approach demonstrated in this study can be useful in many agricultural applications based on remote sensing data, for example Zaeen et al., [52] who developed in-season potato yield prediction models based on several VIs, and Kganyago el al., [22] that studied the performance of SNAP Biophysical processor machine learning algorithm in LAI estimation of several crops. These applications might benefit from further investigation of the correlations between Sentinel-2 bands and various vegetation variables.
In the present study, all the Sentinel-2 bands and the majority of VIs (except reNDVI, WEVI, and WNEVI) showed low performance in LAI estimation of wheat. Therefore, despite the achievements in estimating LAI using Sentinel-2 bands in tomato, cotton, and wheat, additional measurements of wheat are needed to estimate Sentinel-2 bands saturation levels in that crop. Moreover, owing to the spectral resemblance of the Sentinel-2 MSI and the VENµS sensors [2,11,53], a combination with VENµS might facilitate better agricultural monitoring, considering its high two-day temporal resolution.
Overall, the study quantified the performance of the individual Sentinel-2 bands and several VIs (including two newly developed VIs) in the LAI estimation of tomato, cotton, and wheat. Such a result facilitates deriving efficient algorithms and methods for agricultural monitoring via optical satellite imagery.

Conclusions
This study is a step towards improving agricultural practices such as variable rate irrigation, fertilizer and herbicide application, yield prediction, disease monitoring, and many others. This achievement is made possible because of the newly-derived VIs and models that can estimate LAI throughout the season without saturation. As a result, agricultural practices informed through remote sensing can potentially improve agricultural production.
This study found that Sentinel-2 Band-8A (Narrow NIR) is more accurate for LAI estimation than Band-8 (NIR). A very important achievement of the study is that the Band-5 (Red edge-1) showed a low correlation with LAI. Band 9 (Water vapour) showed a very high correlation with LAI alongside the red-edge bands 6 and 7 and NIR bands. Band-9 was demonstrated to be effective for LAI estimation when incorporated into new VIs suggested here for the first time, WEVI and WNEVI. Importantly, Bands 1, 2, 3, 4, 5, 11, 12 were saturated at LAI ≈ 3 and were practically not responsive to a further increase in LAI around LAI ≈ 6. Bands 6, 7, 8, 8A, 9 did not saturate at high LAI. ReNDVI, WEVI, and WNEVI were found to be the best performing VIs for LAI estimation of all three crops tested in this study. 1 A defective red edge band in a Sentinel-2 image acquired on 10 June 2019 prevented the derivation of red edge-based models for that date. 2 Polygon in the centre of the field. 3 NW polygon.