Estimation Of Chlorophyll-A Concentrations Using Field Spectral Measurement And Multi-source Satellite Data In Lake Qiaodao,

China (Project ID :10668)

Prof. Gong Jianhua, P.I. (China) Dr. Apostolos Sarris, P.I. (Greece) Prof. Qigen Liu, Co-P.I. (China) Dr Tao CHENG, Co-P.I. (UK)

Dr. Ibrahim Abdoul Nasser, Co. I. (China) Dr. Andrew Gill, Co. I. (UK) Dr. Li Yi, Co. I. (China)

Presentation Contents 1. Introduction of study area

2. Measured reflectance spectral of Lake Qiaodao

3. Four-band model to estimate Chl-a concentration

4. SVM and random forests to estimate Chl-a concentration

5. Other work in progress

Introduction of Lake Qiaodao

Lake Qiaodao, located in Zhejiang Province, China, which is the most important drinking water source and ecological buffer for Hangzhou Bay and Yangtze River Delta regions.

Lake Qiandao has an area of 573 km2 with more than 1078 islands, storage capacity of 17.8 billion m3, an average depth of 34 m.

In recent years the eutrophication of Lake Qiandao has been exacerbated

A joint field survey on Oct. 13-16, 2014, Lake Qiaodao. Field spectral + Chlorophyll concentration measurement

Measured reflectance spectra of Lake Qiandao

In this study, 50 sampling points were set covering the five Lake sections (northeast, northwest, southeast, southwest, central lakes).

The Northeast Lake is sampled with denser points because of an obvious concentration gradient of chlorophyll-a.

Measured reflectance spectra of Lake Qiandao

SVC HR-1024 (measure scope of 350-2500nm)

Algae Field Analyzer (Fluoroprobe) is adopted to measure the concentration of chlorophyll-a

ρ=0.028 (sunny, breeze wind) Rp=0.3

Absorption coefficients of each component measured in lab, including Chl-a, CDOM and suspended matter.

Abs

orpt

ion

coef

ficie

nt(m

-1)

Yellow material and non-algal particles have a strong cover effect on absorption characteristics of chlorophyll, which will affect the accuracy of chlorophyll concentration inversion model

Correlation between spectral reflectance and chlorophyll a concentration (Pearson analysis). Highest positive correlation coefficient(686.4 nm） Highest negative correlation coefficient(493.1 nm）

Band ratio model

Band ratio model built(using 35 out of 50 points)

Validation (using rest 15 points)

Band (493 nm) whose first derivative result has the highest correlation coefficients with chlorophyll-a concentration.

First order differential model

Model built(using 35 out of 50 points)

Validation (using the rest 15 points)

Three-band model 1 1

1 2 3[ ( ) ( )] ( )Chla R R Rλ λ λ− −∝ − ×

λ1- Absorption valley in red band （Initial value: 663 nm） λ2- Fluorescence peak λ3- Near infrared （Initial value: 740 nm） 优化次数 λ1 λ2 λ3 最优波段 Rmax

1 663 —— 740 695.9 663 695.9 —— 762.3 —— 695.9 762.3 660.2 2 660.2 —— 762.3 701.3 660.2 701.3 —— 758.4 —— 701.3 758.4 663 3 663 —— 758.4 706.7 663 706.7 —— 758.4 —— 706.7 758.4 663

优化结果 663 706.7 758.4 0.9441

Four-band model to estimate Chl-a concentration of Lake Qiandao

Iterative circle λ1 λ2 λ3 λ4 Rmax Optimum band

1 —— 690.4 705.4 740 0.8811 665.7 665.7 —— 705.4 740 0.9173 691.8 665.7 691.8 —— 740 0.9252 687.7 665.7 691.8 687.7 —— 0.9298 714.8 2 —— 691.8 687.7 714.8 0.9426 661.6 661.6 —— 687.7 714.8 0.9466 706.7 661.6 706.7 —— 714.8 0.9481 682.2 661.6 706.7 682.2 —— 0.9513 714.8 3 —— 706.7 682.2 714.8 0.9513 661.6

The four-band model is to apply the principle of biological optical model. The optimized bands combination λ1 - λ4 is achieved by loop iterations calculating method to obtain the largest correlation with the concentration of chlorophyll-a .

1 1 1 1 11 2 3 4[ ( ) ( )] [ ( ) ( )]Chla R R R Rλ λ λ λ− − − − −∝ − × −

λ4 is introduced to eliminate scattering and absorption effect in NIR band caused by suspended matter

Brief summary

1 1 1 1 131.97 [ (661.6) (706.7)][ (714.8) (683.2)] 28.66rs rs rs rsChla R R R R− − − − −= × − − +

Model R2 RMSE MAE

Band ratio 0.7145 2.122 1.720

First order differential

0.7667 1.851 1.662

Three bands 0.9014 1.308 0.972

Four bands 0.9115 1.097 0.858

Compared with the band ratio model and the first derivative model, the precision of three band and four band model are higher. The four-band model has a higher precision than three band model.

Remote sensing Images gathered

Month Satellites Sensors Day Track 2014.10 HJ1 HJ1A-CCD1 2014-10-14 L20002208328

GF1 GF1-WFV2 2014-10-15 L1A0000391134

HJ-CCD Pseudo-color image (NIR-R-G) GF-WFV Pseudo-color image(NIR-R-G)

SVM to estimate Chl-a concentration in Lake Qiandao

Work flow chart

SVM to estimate Chl-a concentration in Lake Qiandao

2 2

1 1

( )( )

( ) ( )XY n n

i ii i

X X X Yr

X X Y Y= =

− −=

− • −

∑∑ ∑

Pearson correlation calculation method

Band component Correlation coefficient

Band component Correlation coefficient

B4/B1 0.8007 B4/(B1+B2) 0.7878 B4/B2 0.7426 B4/(B2+B3) 0.7293 B4/B3 0.6641 B4/(B1+B2+B3) 0.7725

(B4-B1)/(B4+B1) 0.7967 B4/(B1*B2) 0.7486 (B4-B2)/(B4+B2) 0.7416 B4/(B2*B3) 0.6701 (B4-B3)/(B4+B3) 0.6560 (B4*B3)/(B1*B2) 0.7937

Correlation coefficient of chl-a concentration and HJ1A-CCD1 band combination (2014-10-14).

Correlation coefficient of chl-a concentration and GF1-WFV2 band combination.

Band component Correlation coefficient

Band component Correlation coefficient

B4/B1 0.8063 B4/(B1+B2) 0.7836 B4/B2 0.7424 B4/(B2+B3) 0.7113 B4/B3 0.5239 B4/(B1+B2+B3) 0.7649

(B4-B1)/(B4+B1) 0.8089 B4/(B1*B2) 0.6779 (B4-B2)/(B4+B2) 0.7446 B4/(B2*B3) 0.4022 (B4-B3)/(B4+B3) 0.5253 (B4*B3)/(B1*B2) 0.7928

Totally 12 bands compositions are selected (6 from HJ1A-CCD1 band compositions and 6 from GF1-WFV2 band compositions), and they are taken as input of SVM model.

SVM model inversion results based on HJ-CCD single data source (6 bands )

(a) Model built (using 35 out of 50 points ) (b) Validation (using the rest 15 points )

(a) Model built (using 35 out of 50 points ) (b) Validation (using the rest 15 points )

SVM model inversion results based on GF-WFV single data source (6 bands )

(a) Model built (using 35 out of 50 points ) (b) Validation (using the rest 15 points )

12 bands are selected from HJ-CCD and GF-WFV as independent variables, Chl-a concentration measured in situ is dependent variable.

SVM to estimate Chl-a concentration in Lake Qiandao(2014.10)

SVM-based collaborative multi-source image retrieval model obtains a high accuracy (83.3% of the variance of chlorophyll a concentration), and the RMSE is 1.598 μg / L, much lower than the average of the chlorophyll a concentration 11.774 μg / L.

Data source model N R2 RMSE μg/L MRE % MAE μg/L Nr

HJ-GF SVM 35 0.8328 1.598 11.181 1.208 0.7457

HJ SVM 35 0.7510 1.923 13.010 1.428 0.6124

GF SVM 35 0.7972 1.789 11.986 1.183 0.7233

HJ-GF Multiple regression 35 0.7561 1.905 14.037 1.462 0.6445

Data source model N R2 RMSE μg/L MRE % MAE μg/L Nr

HJ-GF SVM 15 0.8283 1.635 10.977 1.253 0.7539

HJ SVM 15 0.7676 1.949 14.509 1.576 0.6013

GF SVM 15 0.8042 1.855 13.309 1.534 0.7780

HJ-GF Multiple regression 15 0.7708 2.3965 18.318 1.963 0.6806

Evaluation of modeling precision

Evaluation of validation precision

Random forests to estimate Chl-a concentration

Random Forest is a non-integrated learning algorithm parameters, specifically a collection of tree classifier:

{ ( , ), 1, 2... ...}kh x k iθ =Random Forest algorithm requires less sample with high precision advantages

RF model inversion results based on HJ-CCD single data source

(a) Model built (using 35 out of 50 points ) (b) Validation (using the rest 15 points )

Random Forest model inversion results and accuracy assessment

(a) Model built (using 35 out of 50 points ) (b) Validation (using the rest 15 points )

RF model inversion results based on GF-WFV single data source

(a) Model built (using 35 out of 50 points ) (b) Validation (using the rest 15 points )

12 bands are selected from HJ-CCD and GF-WFV as independent variables, Chl-a concentration measured in situ is dependent variable.

RF model shows a good prediction of the chlorophyll distribution. The inversion results are spatially accord with the results of SVM. But the phenomenon that high values underestimation exists, especially in the northwest edge of the Lake District.

Data source model N R2 RMSE μg/L MRE % MAE μg/L Nr HJ-GF RF 35 0.8442 1.559 11.651 1.244 0.7904

HJ RF 35 0.7476 2.073 15.471 1.661 0.6860 GF RF 35 0.8235 1.671 11.588 1.324 0.6948

HJ-GF Multiple regression 35 0.7561 1.905 14.037 1.462 0.6445

Data source model N R2 RMSE μg/L MRE % MAE μg/L Nr HJ-GF RF 15 0.8277 1.668 10.643 1.320 0.7194

HJ RF 15 0.6559 2.281 16.597 1.867 0.5514 GF RF 15 0.7666 1.872 12.325 1.468 0.6704

HJ-GF Multiple regression 15 0.7708 2.3965 18.318 1.963 0.6806

Evaluation of modeling precision

Evaluation of validation precision

The contribution of GF-WFV accuracy of the model is generally higher than HJ-CCD combinations.

Based on multi-source remote sensing data (HJ-CCD, GF-WFV), the support vector machine (SVM) and Random Forest (RF) models were constructed to estimate chlorophyll-a concentrations, and the model performance is verified and assessed. The main conclusions of this chapter are as follows: 1) compared to a single data source, the introduction of multi-source data can reduce errors inversion models of chlorophyll-a. 2) SVM retrieval higher accuracy than random forest, but only to enhance the precision of 0.06% -0.32%, while SVM parameter estimation process is more complex, RF model is more simple and practical; 3) Support vector machine and the random forest accuracy are higher than the regression model. Mainly due to the former two have strong nonlinear fitting capabilities, so they can better describe the complex nonlinear relationship between parameters and water spectral characteristics.

Summary

Work in progress

1、Relationship analysis between LULCC of Xinanjiang watershed and the water quality of Lake Qiaodao. 2、Chl-a concentration remote sensing inversion method coupled with hydrodynamic and water quality model.

Work in plan

1、Chl-a concentration remote sensing inversion using UAV equipped with multi-spectral camera.

The cropland and residential area mainly distributes in the northwest of the Lake Qiaodao, which is upstream of the Lake. So the agricultural non-point source pollution and domestic wastewater contributed greatly to make the Northwest lake chlorophyll-a concentrations significantly higher than the other lake sections. Forest occupies most of Xin'an River basin, which is the main reason of high water quality.

[1] Feng Quanlong, Liu Jiantao, Gong Jianhua. UAV Remote Sensing for Urban Vegetation Mapping Using Random Forest and Texture Analysis [J]. Remote Sensing, 2015, 7(1): 1074-1094. [2] Feng Quanlong, Liu Jiantao, Gong Jianhua. Urban Flood Mapping based on Unmanned Aerial Vehicle Remote Sensing and Random Forest Classifier – A Case of Yuyao, China [J]. Water, 2015, 7(4): 1437-1455. [3] Feng Quanlong, Gong Jianhua, Wang Ying, et al. Estimating Chlorophyll-a Concentration based on a Four-band Model Using Field Spectral Measurement and HJ-1A Hyperspectral Data in Qiandao Lake, China [J]. Remote Sensing Letters, 2015. (accepted)

Paper published supported by dragon 3 project:

Young scientists cultivated: 4 persons of Ph.D. level, and 2 persons of M.Sc. level

Thank you!