Available online freely at www.isisn.org Bioscience Research Print ISSN: 1811-9506 Online ISSN: 2218-3973 Journal by Innovative Scientific Information & Services Network RESEARCH ARTICLE BIOSCIENCE RESEARCH, 2018 15(3): 2534-2541 OPEN ACCESS Evaluation of multi-temporal sentinel-2 capabilities for estimation of leaf chlorophyll concentration R. S. Morgan 1, M. Faisal 2, Y. Atyah 2 and I.S. Rahim 1 1 Soils and Water Use Department, Agricultural and Biological Research Division, National Research Centre, Dokki, Cairo, Egypt. 2 Drainage Research Institute, National Water Research Center, Egypt. *Correspondence: randa_sgmm@yahoo.com Accepted: 04 Aug 2018 Published online: 30Sep. 2018 Remote sensing can provide a fast and non-destructive estimation of leaf chlorophyll content and consequently can be used in plant nitrogen status evaluation. Such evaluation can be of valuable assistance in modifying the fertilization strategy in an area of interest. Sentinel-2 has the advantage of the high spatial, spectral and temporal resolution and therefore can be considered a powerful tool in such studies. Furthermore, it incorporates three new spectral bands in the red-edge region which enhances its capabilities for vegetation studies. The present work addresses the comparison between the sensitivity of three approaches in processing Sentinel 2 data for the estimation of leaf chlorophyll concentrations of wheat and olive cultivated in a selected area in the North West of Sinai Peninsula, Egypt. These approaches included the use of reflectance data of selected bands and vegetation indices with or without the red edge bands. The proposed approach has an overall accuracy expressed as coefficient of determination of about 0.90 between the actual and predicted chlorophyll. This approach utilized neural networks and used two vegetation indices the Green Simple Ratio (GSR) and the Inverted Red-Edge Chlorophyll Index (IRECI) in addition to the plant type. Keywords: Sentinel-2 leaf chlorophyll neural networks INTRODUCTION Leaf Chlorophyll is one of the most important plant components which can provide valuable information about its nutritional status (Wu et al., 2008). Furthermore, there is a strong linear correlation between nitrogen and chlorophyll in various plants (Evans, 1989; Bojović and Marković, 2009; Skudra and Ruza, 2017). Therefore, monitoring leaf chlorophyll can be considered as a promising tool for improving nitrogen management. Furthermore, the distribution of leaf chlorophyll content obtained by destructive sampling methods are somewhat considered inaccurate (Curran et al. 1990). Remote sensing techniques have the potentially to provide information on leaf chlorophyll because chlorophyll is a key factor in determining the reflectance in the visible region of the spectrum (Clevers and Gitelson, 2013). Such measurements of the leaves chlorophyll contents are faster and considered non-destructive (Gitelson et al., 2001). In the visible range of the spectrum, chlorophyll has two strong peaks in the red and blue regions. However, chlorophyll content estimations cannot be estimated in the blue peak because it overlaps with the absorbance of the carotenoids (Wu et al., 2008). Furthermore, the importance of the reflectance in red-edge region for chlorophyll content estimation has shown to be highly significant in chlorophyll estimations (Clevers et al., 2001; Clevers and Gitelson, 2013; Schlemmer et al., 2013). The red edge occurs between wavelengths of 680 and 750nm, where
the reflectance changes from very low due to the chlorophyll absorption in the red region to very high in the near infrared because of leaf scattering (Curran et al., 1990). Furthermore, several vegetation indices (VIs) have been proposed for estimating leaf chlorophyll (Gamon and Surfus, 1999; Richardson et al., 2002; Vincini et al., 2007; Lu et al., 2015). Additionally the VIs occasionally included a red edge spectral band (Ramoelo et al., 2012; Clevers and Gitelson, 2013; Schlemmer et al., 2013; Li et al., 2014) and all these studies strengthened the usefulness of these VIs in predicting chlorophyll content. The aim of our study is to evaluate the potentialities of using multi temporal S2 data for the estimation of chlorophyll content of two different plants (wheat and olive) utilizing different approaches with special emphasis on the bands in the red-edge region. MATERIALS AND METHODS Study area and field work The study area is part of Tena Plain in the North West of Sinai peninsula, Egypt and covers an area of about 6.3 hectares (Map 1). The study area was cultivated Olive trees (perennial plant) and wheat (annual plant). The planned field work included the collection of leaf chlorophyll readings at approximately 50-70 m grid based on accessibility (Map 1) with a total of 29 samples collected within 2-3 days of satellite overpass. Plant leaf chlorophyll content was measured at these locations using the Minolta Chlorophyll Meter (SPAD-502). The field work started from mid December/2016 to May/2017 throughout the growth stages of both plants but special attention was given to wheat due to its shorter life spam which was planted from 15-25 November to the end of its growing season in May 2017. The initial stages including germination and seedling growth were not considered for field work while readings were collected starting from tillering stage on mid December 2016 to the end of growing season (May 2017). Moreover, within the study period, the olive trees were in semi-dormancy stage from December till March when it started the Leaf development. Within the study period six dates were available that were cloud free in the study area at the rate of one image in each of the six months from December 2016 to May 2017 except in March where no cloud free images were available and in April where two cloud free images were available. But unfortunately the first three dates from December to February the field work was not available within two to three days of the satellite overpass of the cloud free dates. Therefore only three dates were cloud free and in consistence with field work. These available dates were on 4/4/2017, 14/4/2017 and 4/5/2017. Sentinel-2 Sentinel-2 (S2) was developed by the European Space Agency (ESA) and provides high spatial, spectral and temporal resolution images of the earth. The full S2 mission comprises twin polar-orbiting satellites in the same orbit, phased at 180 to each other at a mean altitude of 786 km. Sentinel-2 revisit time is 10 days at the equator with one satellite and 5 days with two satellites. The satellites provide a set of 13 spectral bands. Four bands at 10m, six bands at 20m and three bands at 60m spatial resolution and it also incorporates three 20 m bands in the red-edge region, which are centered at 705, 740 and 783 nm (Clevers and Gitelson 2013; Vajsová and Aastrand, 2015). Data Pre-Processing In this study Sentinel-2 data were available at ten days interval. These Level-1C radiometrically and geometriacally corrected reflectance data were in Geographic Markup Language JPEG2000 (GMLJP2) format and projected into Universal Transverse Mercator (UTM) and World Geodetic System 1984 (WGS 84) with datum-zone 36 N. The data were imported into QGIS 2.18.0 software which supports the GMLJP2 format and exported as Geotiff into ILWIS software for further processing. Additionally, the observation points were developed into a geographic database with the same projection of the S2 data and frequent chlorophyll data was added to it. Data Processing In our study we used two forms of S2 data the first represented the reflectance data of 6 bands from B3 to B8 (from the green to the infrared bands including the red edge bands), while the second form represented vegetation indices (VIs). Furthermore, the vegetation indices were divided into two groups, the first with the red edge bands while the second without it. The latter vegetation indices included the Green Simple Ratio (GSR) which was developed by Gamon and Surfus (1999) (Eq1) and the Chlorophyll Vegetation Index (CVI) which was developed by Vincini et al., (2007) (Eq2). On the other hand, the red-edge vegetation indices included the Sentinel-2 Red- Bioscience Research, 2018 volume 15(3): 2534-2541 2535
Edge Position (S2REP) and Inverted Red-Edge Chlorophyll Index (IRECI) and both were developed by Frampton et al., (2013). GSR=B8/B3 Eq1 CVI=(B8/B3)*(B4/B3) Eq2 S2REP= 705 + 35 * (0.5 * (B07 + B04) - B05) / (B06 - B05) Eq3 IRECI = (B07 - B04) * B06 / B05 Eq4 Where B refers to repentance, and the subscripts refer to a specific spectral band of S2. The ENVI software was used to calculate and produce the VIs images. The software was also used to extract the image values of the S2 bands and VIs images corresponding to the field observation points. In this study we tried to build an innovative procedure for predicting the leaf chlorophyll. Initially we used the linear regression methods (including the Simple Linear Regression (SLR) and the Multiple Linear Regression (MLR) analyses) to select the best subset of data (bands or Vis) for best predicting the dependent variable (leaf chlorophyll). Based on these selected variables, Artificial Neural Networks were employed to improve the prediction of the linear models, taking advantage of their nonlinear modeling capability. The data extracted from the images were exported into MATLAB for further analysis. A feed-forward neural network was used and when designing the neural network 70% of the samples were used for training and 15% for testing and 15% for validation. The appropriate neural network was assessed based on the coefficient of determination (R 2 ) between the actual and predicted leaf chlorophyll. The structure of the neural network was manually alternated using a trial-and-error process until the highest R 2 was achieved. RESULTSAND DISCUSSION Preliminary statistical analysis Within the period (April and May) the data revealed that the average content of chlorophyll was higher in olive than wheat and it slightly decreased approaching the harvest stage of wheat in May while in the same period it slightly increased for olive. Wheat average chlorophyll content decreased from 40.2 to 36.6, while olive average chlorophyll content increased slightly from 66.2 to 68.2. The correlation analysis between the leaf chlorophyll content of both plants and S2 bands from the green to the near infrared revealed that B5 (the first red edge band) had the highest correlation coefficient (r = 0.75) followed closely by red and green bands (0.73 and 0.72 respectively) then the second red edge band (r = 0.62). The lowest correlation coefficient was reported for the infrared and the third red edge bands reaching 0.53 each. We did not initially process each plant separately when processing the multi-temporal S2 data. Furthermore, the MLR analysis was also used to enhance the selection of the various sets of S2 bands used as input data of the neural network to predict the leaf chlorophyll content. The first set included only bands B3, B4 and B8 (green, red and near infrared bands respectively) which are similar to the bands included in other satellites such as Landsat and SPOT. Secondly, we used these bands in addition to the three red edge bands. The results revealed that using these sets and utilizing MLR, the coefficient of determination (R 2 ) for both data sets did not exceed 0.60 (0.56 for the first set and 0.58 for the second). On the other hand, the use of multi-temporal S2 data with plant type (expressed in numerical coding) could significantly increase R 2 to reach 0.82. Furthermore, two types of vegetation indices were used as alternative option to using reflectance data for chlorophyll predictions. The results revealed that the correlation between the two types of vegetation indices were very weak whereas IRECl and GSR gave the highest correlation and reached only 0.33 and 0.37, respectively. On the other hand, CVI and S2REP did not exceed 0.15. Therefore, in an attempt to increase the accuracy of the prediction the two indices with the highest R 2 were used utilizing MLR. Nevertheless, the results were not satisfying as R 2 only increased slightly to reach about 0.42. We also tried to make use of muti-temporal values of the normalized differential vegetation index (NDVI), which is one of most used vegetation indices in an attempt to differentiate between the different crops based on the different growth stages of both plants and consequently enhance the prediction accuracy. However, the use of such approach increased the accuracy to reach only 0.55. This could be caused by the interactions between numerical NDVI values for both plants within the time of the study. The NDVI values ranged from 0.24 to 0.52 and 0.16 to 0.44 for wheat and olive, respectively. However, when replacing NDVI with plant type the correlation enhanced noticeably reaching 0.81. Bioscience Research, 2018 volume 15(3): 2534-2541 2536
Neural Network (NN) Similarly to the previous approach and using the NN the multi-temporal S2 data were processed. The first (B3, B4 and B8) and second (B3 B8) data sets were used as NN inputs. The designed neural network included one input layer representing the first or the second data set with three and five nodes, respectively and a hidden layer with 10 nodes and one output layer with one node representing the chlorophyll content. The network performance reached R 2 of 0.66 and 0.73 for the first and second data set, respectively [Figure 2a and b]. As the second set enhanced the performance of the NN additional modification adding the plant type to the data sets was done and the network performance could reach R 2 of 0.90 (Figure 3). The designing of the network included one input layer with seven nodes representing the six used S2 bands and the plant type (represented in numerical coding) and a hidden layer with 15 nodes and one output layer with one node representing the chlorophyll content. On the other hand, using the two types of indices whether including the red edge bands or without it individually resulted a relatively weak correlation and R 2 did not exceed 0.54. The indices without red edge bands ranged from R 2 = 0.20 for CVI to 0.53 for GSR (Figure 4b and a respectively). While the indices with red edge bands was 0.30 and 0.46 for S2REP and IRECl respectively (Figure 5 a and b). The designing of each network for these indices included one input layer with one node representing an index and a hidden layer with 10 nodes and one output layer with one node representing the chlorophyll content. Based on these results the two indices which had the best performance (GSR and IRECI) were chosen to design a new network and test their ability to increase the accuracy of the prediction. Using both of these indices resulted in an increase in R 2 to reach 0.60 (Figure 6 a). The designing of the network was similar to the single index network except for the input layer which included two nodes. Further increment could be achieved when adding NDVI as input node where R 2 reached 0.73 (figure 6b). The designing of the network included one input layer with three nodes representing the two chosen indices (GSR and IRECI) and NDVI and a hidden layer with 15 nodes and one output layer with one node representing the chlorophyll content. Furthermore, using GSR and IRECI and replacing NDVI with plant type had the highest accuracy of R 2 of 0.86 (figure 7). The designed network included an input layer with three nodes and one hidden layer with 10 nodes and an output layer with one node repressing the chlorophyll content (Figure 8) Figure 1.Study area location and distribution of soil samples Bioscience Research, 2018 volume 15(3): 2534-2541 2537
(a) (b) Figure 2. Performance of the designed ANN for reflectance data a) the first data set b) the second data set Figure3. Performance of the designed ANN for the second data set and plant type (a) (b) Figure 4. Performance of the designed ANN for vegetation indices a) GSR b) CVI Bioscience Research, 2018 volume 15(3): 2534-2541 2538
(a) (b) Figure 5. Performance of the designed ANN for vegetation indices a) S2REP b) IRECI (a) (b) Figure:6.Performance of the designed ANN for vegetation indices a) GSR and IRECI b) GSR, IRECI and NDVI Figure 7. Performance of the designed ANN for the GSR, IRECI and plant type Bioscience Research, 2018 volume 15(3): 2534-2541 2539
Figure 8. The architect of the proposed neural network CONCLUSION Utilizing NN profoundly enhanced the prediction accuracy of leaf chlorophyll using multitemporal S2 data compared to the linear regression methods. Furthermore, the use of plant type is crucial in predicting the leaf chlorophyll content and the replacement of plant type by the multi-temporal NDVI values could not effectively discriminate between the two plants but still enhanced the prediction accuracy when used with GSR and IRECI. Using the red-edge bands when utilizing with multi-temporal reflectance data has no effect or slightly enhanced the prediction accuracy. The use of GSR (no red-edge band index) gave the highest prediction accuracy of all the used vegetation indices even so R 2 was above 0.5 only when using NN. This could be rendered to the high correlation coefficient between the near infrared and green bands which are the component of the GSR and the chlorophyll content. On the other hand, IRECI gave higher R 2 values than CVI and S2REP but were all still below 0.5. The use of the multi temporal six studied S2 bands (B3-B8) with plant type gave the highest accuracy reaching R 2 of 0.90 when processed with NN compared to all the studied indices with or without the red-edge bands even when they were integrated with plant type. CONFLICT OF INTEREST No conflict of interest AUTHOR CONTRIBUTIONS RS and IS formulated the research idea and designed field work and also wrote the manuscript. Y performed all field work and data collection. RS and M processed and analyzed the remote sensing and GIS data and reviewed the manuscript. All authors read and approved the final version. Copyrights: 2017 @ author (s). This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author(s) and source are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. REFERENCES Bojović B and Marković A, 2009. Correlation Between Nitrogen and Chlorophyll Content in Wheat (TriticumaestivumL.). Kragujevac J. Sci., 31: 69-74. Clevers JGPW and Gitelson AA, 2013.Remote estimation of crop and grass chlorophyll and nitrogen content using red-edge bands on Sentinel-2 and -3. International Journal of Applied Earth Observation and Geoinformation, 23: 344 351. Clevers JGPW, de Jong SM, Epema GF, van der Meer F, Bakker WH, Skidmore AK, Addink EA, 2001. MERIS and the red-edge position. International Journal of Applied Earth Observation and Geoinformation, 3: 313 320. Curran PJ, Dungan JL and Gholz HL, 1990.Exploring the relatioship between reflectance red edge and chlorophll content in slash pine. Tree physiology, 7: 33-48. Frampton WJ, Dash J, Watmough G, Milton EJ, 2013. Evaluating the capabilities of Sentinel- 2 for quantitative estimation of biophysical variables in vegetation. ISPRS Journal of Photogrammetry and Remote Sensing, 82:83 92. Gamon JA and Surfus JS, 1999.Assessing leaf pigment content and activity with a reflectometer. New Phytologist, 143: 105 117. doi:10.1046/j.1469-8137.1999.00424.x Gitelson AA, Gritz Y and Merzlyak MN, 2003.Relationships between leaf chlorophyll content and spectral reflectance and Bioscience Research, 2018 volume 15(3): 2534-2541 2540
algorithms for non-destructive chlorophyll assessment in higher plant leaves.journal of Plant Physiology, 160 (3): 271-282. http://digitalcommons.unl.edu/agronomyfacp ub/771 Evans JR, 1989. Photosynthesis and nitrogen relationships in leaves of C3 plants.oecologia, 78 (1): 9 19. Li F, Miao Y, Feng G, Yuan F, Yue S, Gao X, Liu Y, Liu B, Ustin SL, Chen X, 2014. Improving estimation of summer maize nitrogen status with rededge-based spectral vegetation indices. Field Crops Research, 157: 111 123 Lu S,Lu X,Zhao W,Liu Y,Wang Z,and Omasa K, 2015. Comparing vegetation indices for remote chlorophyll measurement of white poplar and Chinese elm leaves with different adaxial and abaxial surfaces. Journal of Experimental Botany, 66(18): 5625 5637. doi: 10.1093/jxb/erv270 PMCID: PMC4585420 Richardson AD, Duigan SP and Berlyn GP, 2002.An evaluation of noninvasive methods to estimate foliar chlorophyll content. New Phytologist, 153: 185 194. doi:10.1046/j.0028-646x.2001.00289.x Schlemmer M, Gitelson AA, Schepers, James S, Ferguson RB, Peng Y, Shanahan J and Rundquist D, 2013.Remote estimation of nitrogen and chlorophyll contents in maize at leaf and canopy levels.international Journal of Applied Earth Observation and Geoinformation, 25: 47 54. doi: 10.1016/j.jag.2013.04.003. Skudra I and Ruza A, 2017.Effect of Nitrogen and Sulphur Fertilization on Chlorophyll Content in Winter Wheat. Rural Sustainability Research, 37(332):29-37. Scientific Journal of Latvia University of Agriculture. DOI:10.1515/plua-2017-0004. Vajsová B, Aastrand PJ, 2015. New sensors benchmark report on Sentinel -2A. EUR 27674 EN.JRC Technical Reports, Publications Office of the European Union.doi: 10.2788/544302. Vincini M, Frazzi E, and D Alessio P, 2007. Comparison of narrow-band and broad-band vegetation indexes for canopy chlorophyll density estimation in sugar beet. In J. V. Stafford (Ed.), Precision agriculture 07: Proceedings of the 6th European Conference on Precision Agriculture (pp. 189 196). Wageningen Academic Publishers, Wageningen, the Netherlands. Wu C, Niu Z, Tang Q, Huang W, 2008. Estimating chlorophyll content from hyperspectral vegetation indices: Modeling and validation.agricultural and Forest Meteorology, 148:1230 1241. Bioscience Research, 2018 volume 15(3): 2534-2541 2541