Spectral Unmixing And Inverse Estimation Of Chlorophyll-a Concentration In Turbid Water Of Taihu Lake

The optical properties of turbid water are complex and varied,which influenced by phytoplankton,suspended matter,colored soluble organic matter(CDOM)and other optically active substances,so it is a kind of mixed spectrum.The mixed spectra not only affect the estimated accuracy of optical active components,but also influence the further development of quantitative development of water color remote sensing.Concentration of Chlorophyll a(Chla)is an important indicator of eutrophication.As the main pigment,it exists in all kinds of plankton algae and its concentration is closely related to the amount of algae in the water body,influencing the water color,quality and primary productivity of water.Therefore,the estimation of the chlorophyll a concentration is an important content to monitor the water quality.The study is of the project 41471283 supported by NSFC.In this paper,the new model was built based on 120 lake water samples of Taihu in 15 periods from 2006 to 2016(2/3 of water samples used for modeling,1/3 of water samples used for validation),water quality data and MERIS multispectral simulation data.It improved the inversion accuracy of chlorophyll a concentration by using linear unmixing algorithm with spectral signatures analysis and representative endmember extraction.The main contents and conclusions are as follows:(1)Feasibility analysis of spectral unmixing method based on artificial mixing experimentThe remote sensing reflectance and water color parameters of the suspended sediment and the algae were measured with different gradients mixture of suspended sediment and algae in natural water body.The endmember of optical active components and their abundance were determined by using three typical endmember extraction algorithms PPI,N-Findr,and VCA combined with Unconstrained Least Squares(UCLS)algorithm.The results showed that three different endmembers can be extracted by PPI,N-Findr and VCA algorithms effectively,and the endmember abundance has significant correlation with Chla concentration,so the endmember extraction-abundance calculation is feasible for remote sensing estimation of Chla in turbid water.The N-Findr algorithm can extract the endmembers with significant spectral signatures accurately.The PPI algorithm can determine the range of target endmembers roughly.Therefore,the three endmembers of optical active component based on field measured spectrum of Taihu were extracted by combining N-Findr and PPI algorithm.(2)Validation of spectral unmixing algorithm in turbid waterThe endmembers of optical active component of Taihu meansured spectral data from 2006 to 2016 were determined by combining PPI with N-Findr.Six kinds of abundance were calculated by using UCLS and Fully Constrained Least Squares(FCLS)algorithms based on different feature bands like 425-800nm,690-710nm,665-750nm.Three of the models using FCLS got zero values in the abundances,which account for nearly one-third of results,because of too many constraints,while the other three results showed better correlation with Chla concentration.(3)Development and validation of chlorophyll a concentration estimation modelThree Chla concentration remote sensing estimation models were constructed by using relationship between Chla concentration and unmixing coefficients calculated by UCLS algorithm,based on different feature bands of hyperspectral data in Taihu.Compared with the models:ratio model,three-bands model and four-bands model,the three new models using spectral unmixing method got a better accuracy.Among these three models,the one with feature bands 665-750 nm has the highest accuracy,of which R2 is 0.73,RMSE is 11.19μg/L,MAE was 8.44μg/L,and MRE was 27.12%.The validation results show that the new model validation accuracy is also stable,of which R2 is 0.70,RMSE is 8.35μg/L,MAE is 7.46μg/L and MRE is 26.92%.The calculation of simulated multi-spectral MERIS remote sensing image showed advantages of the spectral unmixing model obviously.The spectral unmixing model established based on the 3rd,5th,7th,8th,9th,and 10th bands of MERIS data,which preserves the spectral characteristics of the reflectance curve.The inversion accuracy is significantly better than the ratio model,three-bands model and four-bands models,of which R2 increased to 0.68,RMSE is 11.65 μg/L,MAE is 8.35 μg/L MRE decreased to 26.89%.In summary,remote sensing estimation model of chlorophyll-a concentration was constructed based on the measured hyperspectral data of Taihu and MERIS multispectral simulation data by using spectral unmixing theory.The new model indicated better estimation accuracy and less estimation error compared with the existing typical models.On multispectral data,the advantages of new model over the existing model are more obvious due to the band setting of the multispectral sensor.Compared with the ratio model,three-bands model and four-bands estimation models,the development of spectral unmixing model has a better physical foundation and theoretical support.The results of this paper provide a new reference of model construction for remote sensing estimation of chlorophyll a concentration in turbid water.