Exploring the utility of Sentinel-2 for estimating maize chlorophyll content and leaf area index across different growth stages

ABSTRACT This study investigated the utility of Sentinel-2 spectral data for estimating leaf area index (LAI), leaf and canopy chlorophyll content of maize at different growth stages. Vegetation indices based on the visible-near infrared and red-edge regions of the spectrum were computed from Sentinel-2 imagery acquired within one or two days of field data collection. Results showed that green chlorophyll index (CIgreen) and red-edge chlorophyll index (CIred-edge), using the red-edge variant centred at 705 nm, consistently showed higher relationship to maize LAI with r 2 of 0.65 and 0.63 during the early stages of growth, respectively, and an r 2 of 0.79 and 0.81 during tassel stage, respectively. Regarding canopy chlorophyll content the results indicated the spectral advantage of the Sentinel-2 sensor with the presence of two red-edge bands for continuous monitoring of maize chlorophyll content. Overall, the results indicated that maize biophysical variables can be monitored at satellite level using Sentinel-2 data.


Introduction
Ease of access to high-resolution remote sensing data has propelled interest in precision agriculture to new levels. The advent of European Space Agency (ESA) Sentinel-2 sensor series with relatively high temporal resolution has been lauded as a potential source of near real-time data for monitoring crop growth conditions (Clevers et al. 2017, Delloye et al. 2018. The five-day revisit period of the Sentinel-2 sensor series makes it possible to frequently collect data corresponding to each growth stage of crops. Maize crops (Zea mays L) undergo several growth stages before maturity where chlorophyll content and leaf area index vary rapidly and there are time-specific management decisions (e.g. weed control, top dressing etc.) that should be implemented within a short time span (approximately 20 days) (Du Plessis 2003). For instance, when maize reaches growth stage two with eight leaves, its leaf area increases five-to ten-fold (Du Plessis 2003). Therefore, the availability of high temporal resolution Sentinel-2 data not only enables timely monitoring of crop growth condition, but also facilitates evaluation of the efficacy of management decisions executed at each growth stage.
It has been argued in the literature that the rich spectral configuration of Sentinel-2 sensors, featuring two spectral bands in the red-edge region and a high spatial resolution of 10 m, facilitates monitoring of key variables indicative of crop health such as chlorophyll content and leaf area index (LAI). Frampton et al. (2013) demonstrated the potential utility of Sentinel-2 spectral bands using simulated data by developing new vegetation indices for estimating vegetation biophysical variables. The study concluded that the presence of red-edge bands in Sentinel-2 would particularly enhance the accuracy of retrieving canopy chlorophyll content. Pasqualotto et al. (2019) emphasized the importance of the spectral content in the red-edge region of Sentinel-2 sensors when developing a new Sentinel-2 LAI green Index (SeLI) for accurate quantification of green LAI in heterogeneous agricultural areas. Li et al. (2018), experimenting with Sentinel-2A data, observed that the vegetation index formulated with red-edge bands presents useful spectral information for estimating canopy chlorophyll content in apple trees.
The aforementioned literature indicates that the enthusiasm around Sentinel-2 data emanates largely from the spectral resolution of the sensors and accentuates two points relevant to precision agriculture, (i) the need to accurately monitor key crop variables indicative of its growth condition and (ii) that Sentinel-2 data would enhance the accuracy with which these variables are monitored. Remote sensing of crop chlorophyll content and LAI could serve as an essential means to deliver information on crop growth, nutrient status and photosynthetic capacity to farmers (Gitelson et al. 2015, Gabriel et al. 2017, Xie et al. 2018, Delloye et al. 2018. Chlorophyll content and LAI have long been shown to be closely linked to crop nutrient status (i.e. nitrogen content), physiological activity and potential gross primary productivity (Yoder and Pettigrew-Crosby 1995, Medlyn 1998, Gitelson et al. 2006. LAI, which is defined as the amount of leafy area (m 2 ) in a canopy per unit surface area (m 2 ) (Asner et al. 2003), affects light-use efficiency as it defines the proportion of photosynthetically active radiation absorbed by canopies (Medlyn 1998). Meanwhile, leaf nitrogen (N) content is the principal component influencing both optimum canopy light use efficiency and canopy photosynthesis rate. N deficiency acts to reduce light use efficiency and may reduce the proportion of photosynthetically active radiation absorbed by the canopy through reduction of LAI (Haxeltine andPrentice 1996, Kergoat et al. 2008).
Therefore, the emphasis on accurate monitoring of chlorophyll content and LAI should not be surprising particularly within the context of precision agriculture. Timely and accurate information on chlorophyll content and LAI is crucial for triggering efficient use of farm inputs, thus meeting crop requirements and improving crop productivity (Mulla 2013). The retrieval of crop biophysical variables from remotely sensed data is often implemented with empirical methods, which relate these variables to spectral vegetation indices. Although the Sentinel-2 sensor series could provide near real-time data given the 5-day revisit period, conflicting observations, from various studies, have been made regarding the accurate retrieval of key biophysical variables from the Sentinel-2 data. Contrary to Frampton et al. (2013), Pasqualotto et al. (2019). Li et al. (2018), Dalloye et al. (2018) noted poor estimation accuracy of green area index when using the red-edge bands of Sentinel-2 in an organic farm and attributed it to systematic underestimation in mixed pixels with high heterogeneity. Clevers et al. (2017) observed that vegetation indices derived from the visible and near infrared (VNIR) bands of Sentinel-2 were more suitable for estimating chlorophyll content than the red-edge-based vegetation indices. Poor performance of the red-edge chlorophyll index was attributed to an equilibrium of effects by chlorophyll absorption and canopy scattering, which occurred at around 705 nm for potatoes. Kooistra and Clevers (2016) expressed a suggestion that red-edgebased indices could be replaced by indices using a green band, despite confirming the importance of red-edge bands for agricultural applications.
These diverging observations underline the need for greater scrutiny of the capability of Sentinel-2 sensors in different contexts. In fact, vegetation indices from the VNIR bands are highly affected by differences in leaf structures and canopy architectures compared to red-edge-based vegetation indices. As such, the performance of vegetation indices from VNIR are specific to a particular crop while red-edge-based VIs are species-independent (Nguy-Robertson et al. 2014). Moreover, the two red-edge bands featured in Sentinel-2 sensors should enhance its sensitivity to chlorophyll content in various ranges. Studies have observed that an increase in chlorophyll content activates the broadening of its absorption feature towards longer wavelengths (Curran et al. 1990). This implies that the 740 nm band of Sentinel-2 may be useful for higher ranges of chlorophyll content while the 705 nm band helps with lower ranges. This assumption is supported by Clevers and Gitelson (2013), who noted that the maximum sensitivity of the red-edge chlorophyll index to maize chlorophyll content at high ranges was obtained at 724 nm. While the second red-edge band of Sentinel-2 is centred at 740 nm, it is reasonable to assume that it would be useful for estimating high ranges of chlorophyll content, given that it is still within the position of inflexion point in the red-edge region (Baranoski and Rokne 2005).
Meanwhile, in another study Nguy-Robertson et al. (2014) observed that the accuracy of the red-edge chlorophyll index for estimating maize LAI decreases dramatically when using red-edge variants above 730 nm. Nguy-Robertson et al. (2014) attributed the poor accuracy of LAI estimation to the fact that reflectance beyond 730 nm is more highly influenced by leaf scattering than chlorophyll absorption. However, it must be noted that Nguy-Robertson et al. (2014) only used LAI data from the vegetative stage of maize when chlorophyll content might have been low and that introduced errors in the estimation of LAI. Xie et al. (2018) observed that the red-edge region is mostly affected by chlorophyll content and that the use of the red-edge band for LAI estimation may not be ideal during particular growth stages in which chlorophyll content and LAI vary together. Generally maize chlorophyll content varies throughout the growth stages and, at higher ranges, it is the longer wavelength of the red-edge band that achieved maximum sensitivity (Clevers and Gitelson 2013). Therefore, we can assume that at peak growth stages of maize when chlorophyll content is high, the red-edge band above 730 nm could be useful for estimating LAI because at that stage the reflectance beyond 730 nm would be highly affected by chlorophyll content. This study aimed to test assumptions put forward by investigating the utility of Sentinel-2 spectral bands for estimating maize chlorophyll content and LAI at different growth stages of maize, i.e. cob initiation, brace root development, tassel and blister stage. Chlorophyll content was estimated at both leaf and canopy level to understand individual plants and the spatial variability of crop conditions in fields.

Study area
The study was conducted on the maize farms located in Vanderbijlpark (27.70695 E, −26.51581 S), Bronkhorstspruit (28.83328 E, −25.55871 S) and Magaliesburg (27.45368 E, −26.07293 S) in the Gauteng province of the Republic of South Africa (RSA) (Figure 1). The study consisted of six farms with over 100 hectares each. These farms are located on a generally flat surface dominated by sandy, acidic soil commonly used to cultivate maize crops. Generally, subtropical climatic conditions prevail in the area with an annual average rainfall of 659 mm, which varies from year to year. Rainy season usually occurs between November and February, while the May to August period is characterized as the dry winter season (Hlahane 2018). The farms primarily depend on natural rainfall as a water source, and regularly apply fertilizers and herbicides for soil nutrient supplement and weed control respectively. As a result, the maize cultivation period corresponds to rainfall months in the study area.

Remote sensing data and pre-processing
The study used multispectral data from the Sentinel-2 sensor series. Sentinel-2 data cover the visible, near infrared (NIR) and shortwave infrared (SWIR) regions of the electromagnetic spectrum, including two bands in the red-edge region. The sensor carries the Multi-Spectral Instrument (MSI), which delivers four spectral bands at 10 m spatial resolution (490, 560, 665 and 842 nm), six bands at 20 m spatial resolution (705, 740, 783, 865, 1610 and 2190 nm) and three bands at 60 m spatial resolution (443, 945 and 1375 nm). Cloud-free Sentinel-2 images acquired within one or two days of field data collection were downloaded from the Copernicus Open Access Hub https://scihub.copernicus.eu/dhus/. These images were atmospherically corrected using the image correction for atmospheric effects (iCOR) algorithm to deliver bottom of the atmosphere reflectance. iCOR is a plugin in the Sentinel Application Platform (SNAP) toolbox. It employs a MODTRAN5 Look Up Table (LUT) to complete atmospheric correction and relies on the scene metadata to obtain information regarding solar and viewing angles (Sun zenith angle (SZA), view zenith angle (VZA) and relative azimuth angle (RAA)) and a Digital Elevation Model (De Keukelaere et al. 2018, Pereira-Sandoval et al. 2019. Following atmospheric correction, the four visible-near infrared bands and those spectral bands with a 20 m resolution were resampled to 10 m spatial resolution using the nearestneighbour tool in ENVI 4.8 software. The resampling was completed for ease of stacking bands from different spectral regions and 10 m spatial resolution is consistent with field plots. To address our research objective, the analysis focused on the resampled Sentinel-2 bands. Vegetation indices (VIs) that are known to be correlated to LAI, leaf chlorophyll content (LCC) and canopy chlorophyll content (CCC) (Haboudane et al. 2002, Vincini et al. 2008, Clevers and Gitelson 2013, Gitelson et al. 2015 were computed from Sentinel-2 band combinations. Six VIs were computed using spectral bands in the visible and NIR regions and other vegetation indices were computed using combinations of bands in the visible, red-edge and NIR regions (Table 1).

Field data collection
As part of field data collection activities, measurements of LAI and LCC were conducted on 16-17 January 2018, 29 January 2018 and 19 February and 3-4 March 2020. Additional data were also collected in the 2020/2021 growing season in three separate farms that were at different growth stages to increase the sample size. Plots of 10 m × 10 m were designed to correspond to Sentinel-2 pixel locations, thus avoiding mismatch between field data and Sentinel-2 pixel content. The locations of plots were guided by Global Position System (GPS) coordinates preloaded on the Garmin GPSMAP 64 device. LAI and LCC measurements were made within these randomly distributed plots on the farms. To complete LAI measurements, the LAI-2200 C Plant Canopy Analyzer instrument was used with a 45° view cap under clear sky conditions. For each plot, a minimum of five LAI measurements were recorded. Each measurement consisted of one reading above the canopy and three readings below the canopy. Two below-canopy readings were collected within the rows of maize and one between the rows. The five LAI measurements per plot were used to calculate the average LAI of each plot. For LCC measurements, the SPAD-502® chlorophyll meter was used to estimate leaf chlorophyll content based on the light transmitted by the plant leaf in the red (650 nm) and NIR (940 nm) wavelengths. The SPAD-502® chlorophyll meter is used by clipping it to a portion of the leaf within a small dark sample slot (± 1 cm) and illuminating an LED light corresponding to red and NIR wavelength regions. The instrument receptor receives the transmitted light and computes the ratio to generate a digital reading that is highly correlated with leaf chlorophyll content (Dwyer et al. 1991). The SPAD-502® chlorophyll meter reading was used as a proxy for LCC. In this study, a minimum of five measurements were recorded within a plot. These  (Gitelson et al. 2005) measurements were recorded on five different rows within a plot and targeted the uppermost leaves. The SPAD-502® values for chlorophyll content were converted to LCC using the wheat-specific relationship given by Uddling et al. (2007) and presented as g/ m −2 . CCC was estimated by combining leaf area index with leaf chlorophyll content (LAI × LCC) (Gitelson et al. 2005). Table 2 presents the basic statistics of the field data collected.

Predicting LAI, LCC and CCC
The study used simple linear regression to develop prediction models. Linear regression is a statistical method of explaining the variation in the response variable based on the linear relationship with one or more predictor variables (ter Braak and Looman, 1995). For this study, VIs from different settings of Sentinel-2 bands were used as input variables in a simple linear regression to predict LAI, LCC and CCC. The study predicted LAI, LCC and CCC for each growth stage and then a global model was adopted to predict these biophysical variables using all the data from different growth stages. A bootstrapping approach was applied in modelling these biophysical variables. The study established 1000 random replicates of the original dataset and then separated it into two-thirds for model calibration and one-third for evaluating the model performance. Validation statistics such as r 2 were used to quantify the strength of the relationship statistically while the accuracy of the model prediction was assessed with the root mean square error (RMSE) and normalized RMSE (NRMSE).

Relationship between LAI and VIs
The results show that all the vegetation indices except CI red-edge2 had high relationships with LAI in the cob initiation and brace root development stages sampled in the study (Table S3). At the early stages of growth, the NDVI, GNDVI, CI green and CI red-edge1 were closely related to maize LAI with coefficients of determination (r 2 ) above 0.60 and NRMSE below 30%. However, the CI red-edge2 using the second red-edge band centred at 740 nm as a variant had a low relationship with LAI in the early stages of growth with an r 2 of 0.29 and an NRMSE of above 40%. This observation on the second red-edge band resembles that of Nguy-Robertson et al. (2014), who also noted that using a red-edge band beyond 730 nm decreased the accuracy of maize LAI estimation during the vegetative stage. However, our study observed that during tassel stage, the CI red-edge2 had an improved relationship with maize LAI, boasting an r 2 of 0.76 and NRMSE of 20.15%. These observations regarding the second red-edge band of Sentinel-2 are consistent with the argument of Xie et al. (2018) that the red-edge region is mostly affected by chlorophyll content and this affects the ability of red-edge bands to estimate LAI during particular growth stages. In the early stages of maize growth when canopy chlorophyll content was low (Table 2), the reflectance at 740 nm was affected more by leaf scattering than chlorophyll absorption and this led to a poor relationship between LAI and CI red-edge2 . During the tassel stage when maize chlorophyll content was high, the relationship between maize LAI and CI red-edge2 improved. Nonetheless, it was the CI red-edge1 and CI green that had the best relationships with maize LAI during the tassel stage with an r 2 of 0.81 and 0.79, respectively, while the GNDVI had an r 2 of 0.77 and NDVI had the lowest relationship with LAI with an r 2 of 0.69. This suggests that although the result pertaining to CI red-edge2 does support our assumption that at the peak growth stage the red-edge band positioned beyond 730 nm may be used for LAI estimation, it is not a better option in the presence of CI red-edge1 and CI green . The blister stage saw a systemic decline in the relationship between maize LAI and vegetation indices in general. In the blister stage, the relationship between VIs and LAI deteriorated because of saturation in the case of NDVI and GNDVI and signal scattering in the case of red-edge and green chlorophyll indices (Figure 2). NDVI and GNDVI displayed saturation when LAI reached 2.5 and this sensitivity to maize LAI is lower than the range of LAI values usually observed from maize, which may be up to 6 as observed in Nguy-Robertson et al. (2012). CI green and CI red-edge1 showed no sign of saturation along the entire range of LAI values although there is too much signal scattering in the blister stage ( Figure S2).
Overall, the results show that the highest relationship between vegetation indices and LAI occurred during the tassel stage, which represents the peak of vegetative phenology where maize leaves have fully developed and background contribution to the reflectance signal captured by the Sentinel-2 sensor would have been reduced. NDVI, GNDVI, CI green and CI red-edge1 are sensitive to canopy foliage beside the greenness of vegetation, which is highly influenced by chlorophyll content (Haboudane et al. 2004). This explains the observed higher relationship with LAI during the tassel stage. With a shift in maize growth from cob and brace root development stages to tassel stage, the amount of canopy foliage and chlorophyll increased (Table 2) and this improved the relationship between vegetation indices and LAI. This is more evident in the case of CI red-edge2 .
Notably, the regression models show that NDVI and GNDVI had a non-linear relationship with LAI while CI green and CI red-edge showed a linear relationship ( Figure S2). The results show that using green and red-edge bands in the chlorophyll index overcame the saturation problem and are consistent with the observations in other studies using Sentinel-2 data and field spectrometers (Sun et al. 2019;Nguy-Robertson et al. 2014). The sensitivity of Sentinel-2 derived CI green and CI red-edge to LAI variation across different growth stages of maize indicates their suitability for empirical modelling of maize LAI in the early to mid-growth stages of maize.

Relationship between LCC and VIs
Results showed that CVI and TCARI/OSAVI had changing patterns of relationships with LCC across different stages of maize growth (Table S4). In the early growth stage, CVI had no relationship with maize LCC while the TCARI/OSAVI relationship to maize LCC was very low with an r 2 of 0.16 ( Figure S3). Meanwhile, during the tassel stage both CVI and TCARI/ OSAVI improved their relationship to maize LCC with an r 2 of 0.59 and 0.60 respectively ( Figure S3). The improvement of the CVI and TCARI/OSAVI relationship to LCC coincided with an increase in maize canopy cover as the crop was growing and this pattern was also observed with LCC of potato crop in Clevers et al. (2017). The blister stage of growth saw a slight decline in the relationship between VIs and maize LCC. Combining the data from all growth stages did not improve the model performance, with NRMSE deteriorating for both VIs. However, the good relationship between VIs and maize LCC during the tassel stage should be considered as a further confirmation at satellite level that remote sensing platforms such as Sentinel-2 can track leaf chlorophyll content. The results of Sentinel-2 data in this study are comparable to those of Kooistra and Clevers (2016), who used RapidEye data for LCC estimation in potato farms.

Relationship between CCC and VIs
Vegetation indices showed varying relationships with maize CCC across different growth stages (Table S5). In the early stages of growth, CI green had the highest relationship with maize CCC with an r 2 of 0.62 and NRMSE of 29.9%. GNDVI and CI red-edge1 had the second highest relationship with maize CCC with r 2 of 0.85 and 0.56 and an NRMSE of 30.98% and 31.45%, respectively. NDVI and CI red-edge2 had the lowest relationship with maize CCC with r 2 of 0.48 and 0.33 and an NRMSE of 34.27% and 39.61%, respectively. Overall, the relationship between VIs and maize CCC was low in the early stages of growth, with only CI green achieving moderate results. However, in the peak of the vegetative stage the relationship between VIs and maize CCC improved except for NDVI (Table S5). These improvements coincide with an increase in maize canopy foliage and chlorophyll content as it transitions from cob initiation and brace root development stages to tassel stage ( Table 2). The CI red-edge2 , which uses the second red-edge band sitting at 740 nm, had an r 2 of 0.80 and an NRMSE of 21.31%, while in the early stages of growth CI red-edge2 had a poor relationship with maize CCC. The results are consistent with those of Clevers and Gitelson (2013), who observed that red-edge bands in the longer wavelength were more useful for maize CCC estimation when the chlorophyll content is high. In this study the CI red-edge1 , which uses the first red-edge band sitting at 705 nm, had a lower relationship with maize CCC compared to CI red-edge2 during the tassel stage.
Contrary to Clevers and Kooistra (2011), these results in this study indicate that the position of the red-edge band affects the performance of the CI red-edge in estimating maize CCC. At the early stages of growth when canopy chlorophyll content is still low, the CI red-edge using the red-edge band sitting at 705 nm achieved higher sensitivity to maize CCC than 740 nm. Meanwhile, at the peak of vegetative stage the red-edge band centred at 740 nm achieved higher sensitivity to maize CCC than 705 nm. However, it must be mentioned that the Clevers and Kooistra (2011) study focused on potato crops and grassland. Our observation in this study highlights the spectral advantage that the Sentinel-2 sensor series has with the presence of two red-edge bands which are useful for chlorophyll content estimation at different growth stages of maize.
Moreover, the CI green continued its high relationship with maize CCC into the tassel stage with an r 2 of 0.81 and an NRMSE of 21.91%. The maize CCC relationship with GNDVI also improved during the tassel stage. However, the blister stage saw a systemic decline in the relationship between maize CCC and VIs arising due to saturation of NDVI and GNDVI and signal scattering in the case of red-edge and green chlorophyll indices where higher levels of maize CCC were observed. Overall, NDVI and GNDVI reached saturation at a CCC of 7 g/m 2 and its sensitivity to increasing maize CCC decreased. The two VIs had nonlinear relationships with maize CCC while CI green and CI red-edge showed no sign of saturation with increasing maize CCC ( Figure S4). The global models combining all data performed equally well, particularly the CI green for estimating maize CCC.

Conclusion
The study investigated the utility of Sentinel-2 spectral data for estimating maize LAI, LCC and CCC at different growth stages with the assumption that the presence of two rededge bands could provide a spectral advantage for estimating these biophysical variables. With regard to LAI, the results showed that CI green and CI red-edge1 , using the red-edge band sitting at 705 nm, were the most useful for estimating maize LAI during the vegetative growth stages and therefore they could support key management decisions that should be implemented during vegetative stages, e.g. top dressing. However, maize LCC was poorly estimated in the early vegetative stages and only improved during the tassel stage. The improvement in the tassel stage and the high estimation of LAI in the vegetative stages indicate that maize biophysical variables indicative of growth condition can be monitored at satellite level using Sentinel-2 data. The spectral advantage of the Sentinel-2 sensor was demonstrated when estimating maize CCC at different growth stages. The results showed that in the early growth stages when maize CCC is low, the red-edge band at 705 nm could be useful for accurate estimation of CCC. Meanwhile, at the peak of the vegetative stage when maize CCC is high, the presence of the red-edge band at 740 nm makes it possible to enhance the accuracy of maize CCC estimation. At the peak stage, the use of the red-edge band at 705 nm as a variant in the CI red-edge was unable to deliver the best relationship with maize CCC because it had saturated.
However, it must be noted that the spectral behaviour of the red-edge band centred at 740 nm when used as a variant in the CI red-edge2 had the same effect of introducing error in the estimation of canopy chlorophyll content during the vegetative stage as was observed with LAI by Nguy-Robertson et al. (2014). While Nguy-Robertson et al. (2014) attribute the poor performance of the red-edge band centred beyond 730 nm to simply the effect of leaf scattering, the results in this study showed the utility of the red-edge band centred at 740 nm for estimating LAI and canopy chlorophyll content, which is affected strongly by the growth stage of maize. Vegetation indices are affected by background, leaf pigment, canopy structure and density and the effects of these factors vary between growth stages (Daughtry et al. 2000, Hunt et al. 2011. In view of improving the estimation of crop biophysical variables for agricultural applications, further research is recommended in the area of radiative transfer modelling to counter the limitations of VIs and facilitate the retrieval of LAI and LCC independently.