Sentinel-2 Red-Edge Bands Capabilities for Retrieving Chlorophyll-a: Lake Burullus, Egypt

M. S. Salama1, Hanan Farag1,2 and Maha Tawfik2 The TIGER initiative

1- University of Twente, The Netherlands 2- National Water Research Center, Egypt

UNIVERSITY TWENTE. Outlines

• Why Snetinel-2 for water quality? • Challenges. • Requirements and objectives. • Study area and data. • Why Red-bands models? • Calibration and validation. • Rapid Eye Chl-a products. • Sentinel-2 expected accuracy. • Demonstration with SPOT time series. • Preliminary conclusions.

UNIVERSITY TWENTE. Why Sentinel-2 for water quality?

The aquatic biosphere is uniquely monitored by EO sensors as they provide synoptic information on key water quality variables at high temporal frequency.

The Multi-Spectral Instrument (MSI) payload of SENTINEL-2 mission will sample 13 spectral bands: four bands at 10 m, six bands at 20 m and three bands at 60 m spatial resolution.

The twin satellites of SENTINEL-2 mission offer data at frequency of 5 days at the Equator.

UNIVERSITY TWENTE. Challenges

1 2 3 4 5 6 7 8

9 10 11 12 13 14 15 16

17 18 19 20 21 22 23 24

25 26 27 28 29 30 31 32

Part of the 33 34 35 reality36 37 38 39 40

41 42 43 What you see is not what you get!

UNIVERSITY TWENTE. Requirements and objective

Consistent EO-estimates of water quality parameters in inland and coastal waters requires three components: – (i) a reliable atmospheric correction method; – (ii) an accurate retrieval algorithm and – (iii) an objective method for calibration and validation.

The objective : – Investigate the capabilities of red-bands of the Sentinel-2 MSI sensor in detecting Chl-a in inland waters; – Calibrate and validate such a model.

UNIVERSITY TWENTE. Study area The Lake Burullus is located on Studythe coastal Area shore of the north Nile Delta (30º 35 to 31º 8 E, 31º 22 to 31º 37 N);

The Lake surface area is about 65X14 km2 with an average depth of 1 m;

The gradient between the brackish (southern) part of the lake and marine waters provides a unique ecotonal zone where many marine and fresh water organisms flourish.

UNIVERSITY TWENTE. Field and EO data

• MERIS images and Rapid Eye acquired in 2011 and 2013 concurrent with the field measurements. • In situ Data (Chl-a), 22 in situ measuring locations (2 campaign In Jan and July 2011) matching MERIS images. • 11 in situ measuring locations, for Rapid Eye (12 May 2013)

Resolution Rapid MSI MERIS nm nm Rapid nm nm Eye nm nm 𝝀𝝀 𝚫𝚫𝝀𝝀 MSI MERIS 𝝀𝝀 𝚫𝚫𝝀𝝀 𝝀𝝀 𝚫𝚫𝝀𝝀 Eye 2 490 65 1 475 70 3 490 10

3 560 35 2 555 70 5 560 10 10 m 5 m 300 m 4 665 30 3 655.5 55 7 665 10

5 705 15 4 710 40 9 709 10 20 m

UNIVERSITY TWENTE. Field and EO data

WQV min max mean Stdev nr Chl-a matchup MERIS [mg.m-3] 12.28 121.00 57.33 28.97 22 Chl-a matchup RapidEye 28.73 88.27 51.92 19.72 11

UNIVERSITY TWENTE. Why Red-bands models?

The model establishes a linear relationship between Chl-a concentration and the : 1- Normalized Difference Chlorophyll Index (NDCI):NDCI=(R(708)-R(665))/ R(708)+R(665)) Mishra and Mishra (2012) Basically based on the Gitelson et al., (2006) model 2- RedEdg2/(RedEdge1-Red)

These models did work for MERIS, Rapid Eye, SPOT, but how about the Sentinel-2?

UNIVERSITY TWENTE. Calibration and validation

NDCI MERIS NDCI Rapid Eye 0.4 0.2 y = 0.0064x - 0.4415 y = 0.0064x - 0.4577

0.2 R² = 0.834 0.1 R² = 0.8364

0 0 0.00 50.00 100.00 0.00 50.00 100.00 150.00 -0.1 -0.2 -0.2 NDCI MERIS NDCI NDCI Eye Rapid -0.4 -0.3 -0.4 -0.6 -3 Chl-a mg.m-3 Chl-a mg.m

There is a similarity between the results of the two data sets. NDCI is positive at about >75 mg.m-3 Ideally, the validity of any model is tested against an independent data set. But how to subdivide? Many combinations of matchups can be used!

UNIVERSITY TWENTE. Calibration and validation

Data set • We use the n GeoCalVal model of split Salama et al., 2012

• GeoCalVal derives Cal set = k Val set = n-k the probability K=k+j distributions (PDs) of regression errors calibration model store coefficients and coefficients store validation errors apply the form new cal set model [k,j] N finish randomly picks j

Y j=j+1

UNIVERSITY TWENTE. Calibration and validation, MERIS

80 80

60 60

40 40

Frequency 20 Frequency 20

0 0 4 5 6 7 8 9 10 -0.6 -0.5 -0.4 -0.3 -3 -3 -3 Slope for NDCI, [mg.m ] x 10 Intercept for NDCI ratio, [mg.m ]

100 80

80 60 60 40 40 Frequency Frequency 20 20

0 0 5 10 15 20 25 0.65 0.7 0.75 0.8 0.85 0.9 0.95 MAE of derived Chla, [mg.m-3] R2 of derived Chla

UNIVERSITY TWENTE. Calibration and validation, Rapid Eye

200 150

150 100 100 50 Frequency 50 Frequency

0 0 0 0.002 0.004 0.006 0.008 0.01 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 -3 Slope for NDCI, [mg.m-3] Intercept for NDCI ratio, [mg.m ] 1500 100

80 1000 60

40 500 Frequency Frequency 20

0 0 0 10 20 30 40 50 0.2 0.4 0.6 0.8 MAE of derived Chla, [mg.m-3] R2 of derived Chla

UNIVERSITY TWENTE. Rapid Eye products of Chl-a

UNIVERSITY TWENTE. Rapid Eye products of Chl-a

UNIVERSITY TWENTE. Sentinel-2 expected accuracy

250 Noise model: original_SNR updated_SNR S 200 Noise model = + 𝑆𝑆 150

𝑁𝑁 𝛼𝛼𝛼𝛼 𝛽𝛽 Which can be converted into: SNR 100

= + 50 R2=0.75 𝑁𝑁 𝛼𝛼𝛼𝛼 𝛽𝛽 RMSE=26.5 First and were estimated for 0 all bands , with a correction for 0 50 100 150 200 the spatial𝛼𝛼 (20x20𝛽𝛽 m) and spectral Lref[Wm-2 sr-1 ] (15 nm) resolutions. Sentinel-2 bands (2-8b) 490-865 nm

UNIVERSITY TWENTE. Sentinel-2 expected accuracy

Improvement in SNR of Sentinel-2 MSI over the Rapid Eye

0.8

0.7 Sentinel-2 has about 28- 0.6 70% more SNR w.r.t Rapid Eye 0.5

0.4

The expected accuracy Improvement 0.3

from Chl-a models would SNR be the same or greater 0.2

0.1

0 475 555 655.5 710 Rapid Eye vis bands

UNIVERSITY TWENTE. Expected noise over water

The Noise model is used to simulate 40 the noise level above water for four 35

Sentinel-2/MSI bands: 30 - Estimate the reflectance: 25 = cos 20 𝜋𝜋𝜋𝜋 15 and the noise𝜌𝜌 of it𝑠𝑠 𝑠𝑠 𝐸𝐸 𝜃𝜃 10 = + 5

data field mean to noise relative 𝑁𝑁 𝛼𝛼𝜌𝜌 𝛽𝛽 0 490 560 665 705 In comparison to field measured values, Sentinel-2/MSI bands where Chl-a varies between 12-120 The atmospheric mg.m-3 , the noise level may cancel out correction residuals may to be 13% of error! add additional errors to this!

UNIVERSITY TWENTE. Demonstration with SPOT time series

UNIVERSITY TWENTE. Preliminary conclusions

• NDCI is consistent between MERIS and Rapid Eye matchup data. • The increased SNR of Sentinel-2 over Rapid Eye is about 28-70% and is therefore expected to improve on the results of Chl-a (<12 mg.m-3). • The noise level of Sentinel-2 can adequately be modeled and adjusted to nominal radiometric and spatial resolutions: – The SNR of the Red-band (665) is better than that of the 705 nm by ~40% (due to the course radiometric resolution 30 nm). – It is because of this wide band the reflectance at the Red band will be larger than that at the Red Edge band 705 nm for low Chl-a. i.e. the NDCI will positive for Chl-a>75 mg.m-3 • We’ve demonstrated the potential of Sentinel-2 in deriving Chl-a in productive inland waters.

UNIVERSITY TWENTE. Acknowledgment

This project is financed by the (ESA) , The Tiger initiative, Alcantara

Contact: [email protected]

UNIVERSITY TWENTE.