SPECTRAL FEATURE ANALYSIS FOR QUANTITATIVE ESTIMATION OF CYANOBACTERIA CHLOROPHYLL-A

In recent years, lake eutrophication caused a large of Cyanobacteria bloom which not only brought serious ecological disaster but also restricted the sustainable development of regional economy in our country. Chlorophyll-a is a very important environmental factor to monitor water quality, especially for lake eutrophication. Remote sensed technique has been widely utilized in estimating the concentration of chlorophyll-a by different kind of vegetation indices and monitoring its distribution in lakes, rivers or along coastline. For each vegetation index, its quantitative estimation accuracy for different satellite data might change since there might be a discrepancy of spectral resolution and channel center between different satellites. The purpose of this paper is to analyze the spectral feature of chlorophyll-a with hyperspectral data (totally 651 bands) and use the result to choose the optimal band combination for different satellites. The analysis method developed here in this study could be useful to recognize and monitor cyanobacteria bloom automatically and accurately. In our experiment, the reflectance (from 350nm to1000nm) of wild cyanobacteria in different consistency (from 0 to 1362.11ug/L) and the corresponding chlorophyll-a concentration were measured simultaneously. Two kinds of hyperspectral vegetation indices were applied in this study: simple ratio (SR) and narrow band normalized difference vegetation index (NDVI), both of which consists of any two bands in the entire 651 narrow bands. Then multivariate statistical analysis was used to construct the linear, power and exponential models. After analyzing the correlation between chlorophyll-a and single band reflectance, SR, NDVI respectively, the optimal spectral index for quantitative estimation of cyanobacteria chlorophyll-a, as well as corresponding central wavelength and band width were extracted. Results show that: Under the condition of water disturbance, SR and NDVI are both suitable for quantitative estimation of chlorophyll-a, and more effective than the traditional single band model; the best regression models for SR, NDVI with chlorophyll-a are linear and power, respectively. Under the condition without water disturbance, the single band model works the best. For the SR index, there are two optimal band combinations, which is comprised of infrared (700nm-900nm) and blue-green range (450nm-550nm), infrared and red range (600nm-650nm) respectively, with band width between 45nm to 125nm. For NDVI, the optimal band combination includes the range from 750nm to 900nm and from 700nm to 750nm, with band width less than 30nm.For single band model, band center located between 733nm-935nm, and its width mustn’t exceed the interval where band center located in. This study proved, as for SR or NDVI, the centers and widths are crucial factors for quantitative estimating chlorophyll-a. As for remote sensor, proper spectrum channel could not only improve the accuracy of recognizing cyanobacteria bloom, but reduce the redundancy of hyperspectral data. Those results will provide better reference for designing the suitable spectrum channel of customThe International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLI-B7, 2016 XXIII ISPRS Congress, 12–19 July 2016, Prague, Czech Republic This contribution has been peer-reviewed. doi:10.5194/isprsarchives-XLI-B7-91-2016 91 ized sensors for cyanobacteria bloom monitoring at a low altitude. In other words, this study is also the basic research for developing the real-time remote sensing monitoring system with high time and high spatial resolution

ized sensors for cyanobacteria bloom monitoring at a low altitude.In other words, this study is also the basic research for developing the real-time remote sensing monitoring system with high time and high spatial resolution

INTRODUCTION
In recent years, lake eutrophication caused a large of Cyanobacteria bloom which not only brought serious ecological disaster but also restricted the sustainable development of regional economy in our country, it has become a pretty serious environmental problem.In order to reduce the water pollution caused by cyanobacterial bloom and ensure the safety of drinking water, many satellite data have been used to study the spatial distribution (Matthew, 2012), and the biomass quantitative retrieval (LIU Tangyou, 2002;Heng Lyu, 2013) for water bloom.
Biological risk caused by cyanobacteria bloom was evaluated, while cyanobacteria bloom was monitored and predicted, and water quality early warning system was established (Gower, 1994;Adam, 2000;FENG Jiangfan et al, 2009).These satellite data can be classified into wide band multispectral data and narrow band hyperspectral data according to the spectral resolution, hyperspectral data generally comes from field spectrometer and aerial image.According to the previous researches, narrow band can provide more significant information about the physical characteristics for quantitative research.
Chlorophyll-a is a kind of important environment factor which can be utilized to evaluate water quality, nutrient load and pollution level.In the past decades, many vegetation indices were used to quantify the biological variation of aquatic plants, for example Simple Ratio (SR) is often used to estimate the biological variation, such as cyanobacteria, leucocyan etc. Normalized Difference Vegetation Index (NDVI) is an effective index that can monitor vegetation and ecotope (Ekstran, 1992;MA Ronghua et al, 2005;Sachidananda Mishra, 2013;Wesley J.Moses, 2012;Changchun Huang, 2014).In spite of the same vegetation indices, there also would be discrepancy in quantitative estimation precision, since the band location and width of satellite spectral channel are different.As for spectral index, the center location and band width have a lot of impact on the precision of quantitative estimation.
The main purpose of this paper is to analyze the quantified abil-

Control experiment
The FieldSpecFR spectrometer manufactured by American ASD Company has been applied in this research.The spectrum consist of visible and near-infrared (VNIR), short wave infrared 1 (SWIR1) and short wave infrared 2 (SWIR2).After interpolating the spectral channels, spectral data was obtained, of which the spectral resolution is 1nm.The VNIR range (350nm-1000nm) was used in our research.The spectral reflectance is calculated by the following equation: Where T  = spectral reflectance

LT=target radiance
Lr=reference plate radiance κ=reference plate reflectance In our research, a standard gray plate was used as the reference plate, whose reflectance is 30%.The quantitative estimation ability of narrow band vegetation models, broad band vegetation indices models and chlorophyll-a were compared respectively, based on which the optimal spectral width can be extracted.Meanwhile Landsat ETM+ bands were selected as the representative of broad bands.Spectral response function was used, and spectral response values of Landsat ETM+'s first 4 bands were simulated by ASD spectral reflectance (Steven et al, 2003).Before applying convolution to ASD spectral reflectance, smoothing processing was not used while obtaining the broad band spectral data.

Multivariate statistical analysis
Multivariate statistical analysis is generally applied in the study of ocean color remote sensing (Andrew Clive Banks, 2012).
Original spectrum, simple ratio (SR) and normalized difference vegetation index (NDVI) (Table 1) were chosen as independent variables, and the concentration of Microcystic aeruginosa chlorophyll-a was regarded as the dependent variable.Pearson product-moment correlation coefficient (r) was calculated to describe the correlation between dependent variable (y) and independent variable (x): Where x = average of independent variables y=average of dependent variables n= the number of dependent variables/independent variables Then intervals in which x, y located with high degree of correlation were chosen to construct a linear prediction model.
Actually, there might be non-linear relationship between original spectrum, SR, NDVI and chlorophyll-a respectively, so another two non-linear prediction models were used: power and exponential in addition: According to the result of determination coefficient (R 2 ): The optimal prediction model would be chosen and the precision of predicted value would be evaluated with root mean square error (RMSE) and mean relative error (MRE).Data ob- tained in the first experiment were applied as forecast samples, then the data from second experiment were regarded as test samples.

RESULTS AND ANALYSIS
Spectral reflectance of water and Microcystic aeruginosa with

Correlation of narrow band vegetation indices and Microcystic aeruginosa chlorophyll-a
Narrow band SR and NDVI indices consist of any two bands in the entire 651 narrow bands (651*651=423801) were calculated.
All narrow band SR and NDVI were chose as independent variable and chlorophyll-a is as dependent variable to construct linear regression model.R 2 was calculated and shown in Figure  Where Δ1 = (0-Δλ1/2)nm Δ2 = (0-Δλ2/2)nm There were totally 24 different results for six indices in this experiment (two types, two conditions).The most optimal narrow band width was found in the NDVI index under the condition of water stationary (Δλ1=8nm, Δλ2=4nm); while the most optimal broad band width was found in the SR5 index under the The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLI-B7, 2016 XXIII ISPRS Congress, 12-19 July 2016, Prague, Czech Republic condition of water disturbance (Δλ1=362nm, Δλ2=227nm).

Non-linear correlation of vegetation indices and chlorophyll-a
Figure 2 illustrates the linear correlation of SR, NDVI indices and Microcystic aeruginosa chlorophyll-a.However, sometimes there may be non-linear correlation between vegetation indices and chlorophyll-a.In our research, power model is the optimal non-linear model.6 groups of optimal linear models' band center (λ1 and λ2) and band width (Δλ1 and Δλ2) were applied to calculate non-linear models.R 2 and RMSE (Table 3) were calculated by optimal linear models and power models.
As shown in Table 3, for SR indices, SR1/2/3 linear models are better than those power models.While SR4/5/6 power models are better than those linear models under the condition of disturbance.Only SR4 power model performed better than those linear models under the condition with disturbance.For NDVI, under the condition of disturbance, estimation precision of power models is superior to their linear models.Under the condition of stationary, NDVI 1/2/3/5 linear models performed better than their power models.NDVI 4/6 and ETM_NDVI power models performed better than their linear models.0.9694 57.45 Table 3: 6 optimal linear models and non-linear models

Distribution of optimal bands
For the above 6 groups of indices, the optimal 4 groups were chosen to calculate the distribution of corresponding band center and band width(shown in Figure 4 & 5).band centers of ρi are between 450nm-550nm and 600nm-650nm.And the most optimal band center of ρj (λ1, NIR) locates in the range of 700nm-900nm.The most optimal band widths locate between 46nm-125nm and more than 125nm.For NDVI, most ρi locate in the range from 700nm to 750nm, and most ρj are between 750nm-900nm.Obviously, for SR indices, broad bands take an absolute advantage for quantitative estimating Microcystic aeruginosa chlorophyll-a.For NDVI, the narrower bands have, the better quantitative ability is.

CONCLUSIONS
By performing two control experiments, the correlation between chlorophyll-a and single band reflectance, SR, NDVI were analyzed respectively.Linear and non-linear regression models were constructed.Then various spectral indices, which were suitable for quantitative estimation of Microcystic aeruginosa chlorophyll-a, their central band wavelength and band width were obtained, which are summarized herewith: ( (2) Selecting band centers and width of spectral indices: For the SR index, there are two optimal band combinations, which is comprised of infrared (700nm-900nm) and blue-green range(450nm-550nm), infrared and red range (600nm-650nm) respectively, with band width between 45nm to 125nm.For NDVI, the optimal band combination includes the range from 750nm to 900nm and from 700nm to 750nm, with band width less than 30nm.For single band model, band center located between 733nm-935nm, and its width mustn't exceed the interval where band center located in.
Our experiments analyzed the quantified ability between Microcystic aeruginosa chlorophyll-a and different vegetation indices under the disturbance and stationary conditions.The following two aspects could be improved: 1) Besides Microcystic aeruginosa, there is much suspended matter, chromophoric dissolved organic matter and so on which can influence the spectrum in lakes; 2) Atmospheric influence was ignored during simulating Landsat ETM data.So our results should be verified in the lake water.
sa was added drop by drop, meanwhile stirred uniformly.Each time after adding Microcystic aeruginosa, three samples were collected to obtain the concentration of chlorophyll-a, and measured spectral data three times (under the condition of water disturbance).Five minutes later, when the water came back calm and Microcystic aeruginosa distributed uniformly on water surface, spectral data was measured three times again (under the condition of water stationary).These two containers were utilized alternately during the experiment.Totally, 16 groups of different concentration spectral data were collected.The concentration of chlorophyll-a increased at the speed of 1.3 times, ranging from 0-1190.46μg/L.To obtain more reference data, in the second experiment, the same procedure was performed.A group of sub-experiments were under the condition of disturbance, and the range of chlorophyll-a concentration was 15.31μg/L-450.9μg/L(12 gradients).Then a group of sub-experiment was under the condition of water stationary, the range of chlorophyll-a concentration was 143.64μg/L -1362.11μg/L(9 gradients).

Figure 1
Figure1shows the correlation coefficient (r) between single band reflectance and chlorophyll-a under the condition of water disturbance and stationary.

Figure 1
Figure 1 Correlation coefficient between Microcystic aeruginosa spectral reflectance and chlorophyll-a concentration ETM_SR and ETM_NDVI were used to represent broad band indices and C_SR and C_NDVI to narrow band indices.Water disturbance and stationary was abbreviated as D and S respectively.The relation between chlorophyll-a real value Y and preregarded as a scale used to compare each model's predictive ability.By comparing RMSE, under the condition of water disturbance, the predictive precision of SR model is higher, even though the R 2 of single band model and SR model are both very high.Hence there is almost little discrepancy between narrow SR model and broad band SR model.Under the condition of water stationary, SR model's predictive precision is lower than that single band model's, where RMSE of SR model is 1.7 times higher than that of single band model.The predictive ability of ETM+'s band 4 model is close to which of 794nm model.

Figure 2
Figure 2 R 2 's isogram of narrow band SR, NDVI and chlorophyll-a After comparing SR and NDVI, the indices' quantitative estimation precision of SR 1/2/3/4/6 indices and NDVI 5/6 can perform pretty well under the condition of water disturbance.SR indices and NDVI indices' quantitative estimation precisionto chlorophyll-a are similar, which both performed worse than their single band models.

Figure 3 :
Figure 3: Distribution of SR and NDVI indices' band center 1) Selecting spectral indices: Under the condition of water disturbance, SR and NDVI indices are both suitable for quantitative estimation of chlorophyll-a, which both performed better than those single band models.Linear model was suitable for SR indices, and power model was suitable for NDVI, under the condition of water stationary, single band models can estimate the concentration of chlorophyll-a pretty well, while SR and NDVI indices performed worse.