UNIVERSIDADE ESTADUAL PAULISTA “Júlio de Mesquita Filho” FACULDADE DE CIÊNCIAS E TECNOLOGIA Campus de Presidente Prudente PPGCC – Programa de Pós-Graduação em Ciências Cartográficas Retrieval of euphotic zone and Secchi disk depth in Bariri reservoir using OLI/Landsat-8 data PRESIDENTE PRUDENTE 2018 Ana Carolina Campos Gomes Retrieval of euphotic zone and Secchi disk depth in Bariri reservoir using OLI/Landsat-8 data Dissertation presented to the Department of Cartography of the Faculty of Science and Technology of São Paulo State University “Júlio de Mesquita Filho” as part of the requirements for obtaining the Master degree in Cartographic Sciences. Advisor: Prof. Dr. Enner Alcântara Co-advisor: Dra. Thanan Walesza Pequeno Rodrigues Presidente Prudente 2018 FICHA CATALOGRÁFICA AGRADECIMENTOS Meu primeiro agradecimento é a Deus, por todas as oportunidades concedidas e por ser luz e esperança em momentos árduos. Agradeço imensamente aos meus pais, Marlene e Luís Antonio, que sempre estiveram presentes em minhas decisões, sendo meus apoios constantes, garantindo um lar cheio de amor e carinho. À minha irmã, Ana Laura, pela amizade, companheirismo e amor. Ao André, pela paciência e companheirismo. Agradeço à orientação do Prof. Dr. Enner, por todo o comprometimento e competência dedicados ao desenvolvimento deste trabalho. Pelo incentivo na escrita de artigos e por proporcionar meu crescimento na pesquisa. À Dra. Thanan Rodrigues, pela coorientação, contribuindo com a evolução do trabalho. Agradeço aos meus tios, primos e amigas de infância pelas conversas, apoios, orações e pela torcida em todos os momentos da minha vida. Agradeço aos meus amigos do grupo de pesquisa SERTIE, em especial, ao grupo de pesquisa de Sensoriamento Remoto em sistemas aquáticos por me acolherem tão bem no grupo e no ambiente da Pós-Graduação, em especial agradeço: À Nariane, Alisson, Fernanda e Luiz por toda atenção, contribuição; pela disponibilidade, disposição, confiança nas atividades do grupo e por toda a organização dos trabalhos de campo que possibilitaram as coletas dos dados usados neste trabalho. Agradeço ao Dr. Luiz e a Prof. Dra. Fernanda pela composição da banca no Exame de Qualificação do Mestrado. À Sarah, Carol Ambrósio, Carol Piffer e Bruno Frias (GeoGrupo), pelo companheirismo e parceria, tornando as longas horas de aulas e estudos mais divertidas e menos árduas e pela amizade, que ainda permanece com a distância. Agradeço também a Nariane e ao Alisson, pela grande ajuda com o processamento dos dados, pela companhia nas horas de trabalho e nos intervalos e pelos laços criados, que culminaram em um convite muito especial. Agradeço aos amigos da Pós-Graduação: Francielle, Guilherme, Tobias, Luiz Eduardo e Gabriel pelas conversas e cafés compartilhados, pelo acolhimento, ajuda e amizade. Agradeço a toda infraestrutura fornecida pelo Programa de Pós-Graduação em Ciências Cartográficas (PPGCC), às funcionárias, Cida e Zilda, e a todos os professores que enriqueceram minha caminhada. À estrutura da FCT/UNESP, a todos os funcionários da unidade e aos professores da Graduação em Engenharia Ambiental, em especial a Profa. Dra. Maria Cristina Rizk. Agradeço ao Prof. Dr. Nilton Imai pelo laboratório TIE e sua infraestrutura para a realização das análises dos dados. Ao Prof. Dr. Paulo César Rocha pela autorização do uso do Laboratório de Geologia, Geomorfologia e Recursos Hídricos. Ao Prof. Dr. Edivaldo Velini e aos técnicos da UNESP de Botucatu, pela permissão do uso dos equipamentos para análises laboratoriais. Agradeço ao Conselho Nacional de Desenvolvimento Tecnológico e Científico - CNPq por conceder a bolsa de mestrado (Processo N.° 131737/2016-3) e pelo financiamento de projetos (Processos 400881/2013-6 e 472131/2012-5) e à Fundação de Pesquisa do Estado de São Paulo - FAPESP (Projetos 2012/19821-1 e 2015/21586-9) pelo financiamento dos trabalhos de campo. A todas as pessoas envolvidas nesses dois anos de trabalho, meus sinceros agradecimentos! RESUMO O presente trabalho teve como objetivo estimar as profundidades da zona eufótica (Zeu) e do disco de Secchi (ZSD) a partir do coeficiente de atenuação da luz (kd) utilizando dados do sensor Operational Land Imager (OLI)/Landsat-8 no reservatório de Bariri. Como importantes parâmetros de medida da claridade da água, kd, Zeu e ZSD são afetados pelas substâncias opticamente significativas (SOS). A caracterização óptica do reservatório foi realizada a partir de duas campanhas de campo realizadas no período seco, aqui nomeadas como BAR1 (agosto/2016) e BAR2 (junho/2017), que contaram com análises das propriedades ópticas inerentes (POIs), das SOS e da coleta de dados radiométricos para o cálculo da reflectância de sensoriamento remoto (Rsr). A localização do reservatório de Bariri como o segundo do Sistema de Reservatórios em Cascata (SRC) do Rio Tietê promove a heterogeneidade dos seus níveis de eutrofização na direção montante-jusante além de caracterizá-lo como altamente produtivo. As campanhas de campo foram marcadas por uma significativa diferença nos valores de concentração de clorofila-a ([Chl-a]) que apresentou variação média entre 7,99 e 119,76 μg L -1 com os maiores valores em BAR1, com decréscimo das SOS em BAR2 em relação a BAR1 e predomínio de material particulado orgânico (MPO) nas duas campanhas de campo; a turbidez variou entre 5,72 e 16,60 NTU. A absorção por matéria orgânica colorida dissolvida (aCDOM) foi predominante nas duas campanhas de campo, sendo mais expressiva em BAR2. Para as estimativas de kd, nove modelos empíricos e três modelos semi-analíticos baseados em dados radiométricos como razões entre as bandas azul/verde e azul/vermelho do sensor OLI/Landsat-8 e baseados em [Chl-a] foram avaliados. Considerando a propriedade óptica aparente (POA) do kd, um modelo semi-analítico baseado em POIs e na distribuição angular da luz apresentou os menores erros (erro médio percentual absoluto – MAPE) de 40% em relação aos modelos empíricos de [Chl-a] com 60% e de 80% para os modelos empíricos baseados em razões de bandas. A partir das estimativas de kd, modelos de estimativa de Zeu e ZSD foram avaliados. Para as estimativas de Zeu, cinco modelos empíricos, baseados na relação entre o coeficiente de atenuação da luz da radiação fotossinteticamente ativa [kd(PAR)] e de kd em 490 nm [kd(490)], e um modelo semi-analítico, baseado na equação de transferência radiativa, foram considerados; para as estimativas de ZSD, um modelo semi-analítico foi testado. Os resultados obtidos foram melhores para um modelo empírico (erro percentual absoluto – ε) de Zeu com 16% em relação ao modelo semi-analítico (ε 30%) e os erros nas estimativas de ZSD foram de 57%. Os erros nas estimativas de kd revelaram que a acurácia dos modelos empíricos foi comprometida devido à influência por CDOM e que o modelo semi-analítico, por considerar a natureza óptica de kd como uma POA, apresentou os melhores resultados. As estimativas de ZSD também foram afetadas pelas características ópticas de Bariri, não apresentando correlação com a matéria orgânica em BAR2, marcado pelo decréscimo de [Chl-a] e aumento dos valores de aCDOM. Zeu mostrou melhores resultados a partir de um modelo empírico calibrado com dados ópticos semelhantes aos do reservatório de Bariri em comparação ao modelo semi-analítico, desenvolvido para abranger as variações bio-ópticas sazonais e regionais. kd, Zeu e ZSD foram espacializados a partir de imagens do sensor OLI/Landsat-8 permitindo a avaliação espaço- temporal desses parâmetros que apresentaram um padrão sazonal quando analisados em relação aos dados de precipitação. kd apresentou variação entre 0,89 e 5,60 m -1 para o período analisado (2016) e Zeu e ZSD apresentaram variação entre 0,30 e 7,60 m e entre 0,32 e 2,95 m, respectivamente, para o período de 2014-2016. Pode-se concluir então, que apesar das estimativas de kd, Zeu e ZSD terem sido afetadas pela influência de CDOM no reservatório de Bariri, o esquema semi-analítico foi capaz de estimar kd com menor erro e permitiu as estimativas de Zeu e ZSD. Palavras-chave: atenuação da luz; qualidade da água; dados de satélite; modelos bio-ópticos. ABSTRACT The objective of this present work was estimate the euphotic zone (Zeu) and Secchi disk (ZSD) depths from the light attenuation coefficient (kd) using the Operational Land Imager (OLI)/Landsat-8 data in Bariri reservoir. The kd, Zeu and ZSD are important water clarity parameters and are influenced by the optically significant substances (OSS). The optical characterization was carried out with data collected in two field campaigns in the dry period, here called BAR1 (august/2016) and BAR2 (june/2017), that included analysis of the inherent optical properties (IOPs), of the OSS and radiometric data to calculate the remote sensing reflectance (Rrs). The location of Bariri reservoir as the second of the Cascading Reservoir System (CRS) of Tietê River promotes the heterogeneity of the eutrophication levels from upstream to downstream besides characterizes the reservoir as highly productive. The field campaigns presented a significant difference in chlorophyll-a concentrations ([Chl-a]) with mean variation between 7.99 and 119.76 μg L -1 with the highest values in BAR1, with reduce of the OSS in BAR2 in relation to BAR1 and predominance of organic particulate matter (OPM) in both field campaigns and variation in turbidity from 5.72 to 16.60 NTU. The absorption of chromophoric dissolved organic matter (CDOM) was dominant in both field campaigns and more expressive in BAR2. For the kd estimates, nine empirical models and three semi-analytical models based on radiometric data such as ratios of blue-green and blue-red bands of (OLI)/Landsat 8 sensor and based on [Chl-a] were evaluated. Considering the apparent optical property (AOP) of kd, a semi-analytical model based on IOPs and the light angular distribution presented the lowest errors (mean absolute percentage error – MAPE) of 40% in relation to the empirical models of [Chl-a] with 60% and of 80% for the empirical models based on the band ratios. Through the kd estimates, models to derive Zeu and ZSD were evaluated. For the Zeu estimates, five empirical models were considered based on the relation between the attenuation coefficient of the photosynthetically active radiation [kd(PAR)] and the kd at 490 nm [kd(490)], and one semi-analytical model, based on the radiative transfer equation; for the ZSD estimates, one semi-analytical model was tested. The empirical model of Zeu showed the better results with the (unbiased absolute percentage error – ε) 16% in relation to the semi-analytical model (ε 30%) and the estimates errors of ZSD were 57%. The errors in kd estimates revealed that the accuracy of the empirical models was affected by the CDOM influence in Bariri reservoir and the semi-analytical model presented a better performance when considering the optical nature of kd as an AOP. The ZSD estimates were also affected by the optical characteristics of Bariri with no correlation to the SPM in BAR2, where the [Chl-a] decreased and the aCDOM increased. Zeu showed better results from an empirical model calibrated with similar optical data to Bariri reservoir in relation to the semi-analytical model developed to be applied in a wide range of bio-optical seasonal and regional variations. The kd, Zeu and ZSD were spatially distributed through OLI/Landsat-8 images allowing the temporal-spatial assessment of theses parameters, which presented a seasonal pattern when analyzed in relation to rainfall data. kd presented variation from 0.89 to 5.60 m -1 to the analyzed period (2016) and Zeu and ZSD presented variations between 0.30 and 7.60 m and between 0.32 and 2.95 m, respectively, for 2014-2016 period. It can be concluded, therefore, that despite of the CDOM have affected the kd, Zeu and ZSD retrievals in Bariri reservoir, the semi-analytical scheme was able to estimate kd with lowest error and enable the Zeu and ZSD estimates. Key-words: light attenuation; water quality; satellite data; bio-optical models. LIST OF FIGURES Page Figure 2.1 Graphics showing the location of (a) São Paulo State in Brazilian context; (b) Bariri Reservoir in the Cascading Reservoir System of Tietê River; (c) sampling points of the two field campaigns carried out, BAR1 and BAR2 and (d) precipitation rate data (mm) from NASA‟s GIOVANNI database for the period of 2014-2017. In 2017, the precipitation monthly averages were available just until July month. 22 Figure 2.2. The Rrs spectra for BAR1 (a) and BAR2 (b) surveys. 28 Figure 2.3. Absorption spectra of NAP, phytoplankton, CDOM and pure water (aw) in BAR1 (a) and BAR2 (b). 30 Figure 2.4. Ternary plots depicting BAR1 and BAR2 for center OLI bands (a) 443 nm, (b) 482 nm, (c) 561 nm and (d) 655 nm. 32 Figure 2.5. Spatial distribution of kd(490) in Bariri reservoir using the semi-analytical model from Lee et al. (2013) considering the months from February to November of 2016. 36 Figure 2.6. Boxplot of kd(490) spatial distribution regarding the months of February to November of 2016 in Bariri reservoir. 37 Figure 3.1. Maps showing the location of (a) São Paulo State in Brazilian territory and the path/row of OLI/L8 images which the coordinates of Bariri reservoir are contained, highlighted by the red square; (b) CRS of Tietê River in São Paulo State with the six reservoirs and their respective dams (1 – Barra Bonita; 2 – Bariri; 3 – Ibitinga; 4 – Promissão; 5 – Nova Avanhadava; 6 – Três Irmãos) and (c) Bariri reservoir boundaries with the sampling points of the two field campaigns. 43 Figure 3.2. Flowchart showing the sequential steps followed to mapping ZSD and Zeu. 49 Figure 3.3. The Rrs spectra for BAR1 (a) and BAR2 (b) surveys. 51 Figure 3.4. Comparison between derived-Zeu from in situ data and those estimated from Zeu models analyzed. 56 Figure 3.5. The spectral distribution of the sample points in Rrs in situ and Rrs_LaSRC derived (a) and the Rrs_LaSRC derived performance in relation to Rrs in situ. 57 Figure 3.6. Graphics of monthly average of (a) rainfall (mm) data for Bariri reservoir, obtained from NASA‟s GIOVANNI database between 2014 and 2017. In 2017, the precipitation data were available just until October and (b) the flow rate (m³ s -1 ) of Bariri reservoir in the period 2014-2017 available in the SAR/National Water Agency (http://sar.ana.gov.br/MedicaoSin). 58 Figure 3.7. Spatial distribution of Zeu using the model based on Zhao et al. (2013) for February, May, August and October months of (a) 2014, (b) 2015 and (c) 2016. 60 Figure 3.8. Boxplots of Zeu spatial distributions in Bariri reservoir for 2014 (a), 2015 (b) and 2016 (c). 61 Figure 3.9. Spatial distribution of ZSD using the model from Lee et al. (2015) for February, May, August and October months of (a) 2014, (b) 2015 and (c) 2016. 63 Figure 3.10. Boxplots of ZSD spatial distributions in Bariri reservoir for 2014 (a), 2015 (b) and 2016 (c). 64 Figure 3.11. Relationships between in situ Zeu and (a) [Chl-a], (b) [SPM], (c) turbidity and between the aphy(443) with (d) [SPM] and (e) [Chl-a]. 66 Figure 3.12 Relationships between in situ ZSD and (a) [Chl-a], (b) [SPM], (c) turbidity. 67 LIST OF TABLES Page Table 2.1. kd(490) models developed for clear ocean, coastal, turbid ocean, slightly turbid and global waters. 25 Table 2.2. Descriptive statistics for optical and water quality parameters for BAR1, BAR2 and the mixed data. (S.D. = Standard Deviation. C.V. = coefficient of variation.) 29 Table 2.3. Error assessment of kd(490) models developed for clear, slightly turbid, turbid ocean waters, coastal waters and global waters applied to all dataset of Bariri. 33 Table 3.1. Zeu models developed from coastal and ocean waters, slightly turbid ocean waters and turbid lake waters data. 47 Table 3.2. Summary of statistical parameters of the optical properties acquired in BAR1 and BAR2 showing the minimum, maximum, mean, standard deviation (SD) and coefficient of variation (CV). CV = SD/mean. 52 Table 3.3. Errors assessment of Zeu models developed for ocean, costal, slightly turbid ocean and turbid lake waters that were tested for Bariri dataset. 53 Table 3.4. Errors assessment of ZSD (Lee et al., 2015) considering the two field campaigns and the aggregated data. 56 LIST OF ABREVIATIONS AND ACRONYMS a Absorption coefficient ap Absorption Coefficient of Total Particulate aNAP Absorption Coefficient of Non-Algal Particle aCDOM Absorption Coefficient of Chromophoric Dissolved Organic Matter aphy Absorption Coefficient of Phytoplankton bb Backscattering coefficient ρ Effective Surface Reflectance Esky Sky Irradiance Ed Downwelling Irradiance Downwelling Irradiance at Water Surface Downwelling Irradiance at Water Subsurface Normalized Downwelling Irradiance ε Unbiased Absolute Percentage Error F Fresnel Reflectance Fr Spectral Response Function F0 Extraterrestrial Solar Irradiance kd(490) Attenuation Coefficient at 490 nm kd(PAR) Attenuation Coefficient of the Photosynthetically Active Radiation Lsky Sky Radiance Lt Total Radiance LSR Surface-Reflected Radiance nLw Water-Leaving Radiance Rrs Remote Sensing Reflectance Remote-Sensing Reflectance Simulated R Irradiance Reflectance Beneath the Water Surface R² Coefficient of Determination Srs Sky Remote Sensing Reflectance Trs Total Remote Sensing Reflectance Zeu Euphotic Zone Depth ZSD Secchi Disk Depth APHA American Public Health Association Protocol AOP Apparent Optical Property CDOM Chromophoric Dissolved Organic Matter Chl-a Chlorophyll-a CRCC Cascading Reservoir Continuum Concept CRS Cascading Reservoir System CV Coefficient of Variation IOP Inherent Optical Property IPM Inorganic Particulate Matter LaSRC Landsat Surface Reflectance Code MAPE Mean Absolute Percentage Error MODIS Moderate Resolution Imaging Spectroradiometer NAP Non-Algal Particle OPM Organic Particulate Matter OSS Optical Significant Substance OLI/Landsat-8 Operand Land Imager onboard the Landsat-8 Satellite PAR Photosynthetically Active Radiation PCA Principal Component Analysis QAA Quasi-Analytical Algorithm RMSE Root Mean Square Error RMSD Root Mean Square Difference SeaWiFS Sea-viewing Wide Field-of-View Sensor 6SV Second Simulation of the Satellite Signal in the Solar Spectrum Vectorial SD Standard Deviation SPM Suspended Particulate Matter USGS United States Geological Survey CONTENTS CHAPTER 1: Introduction………………………………………………………….. 15 1.1. Background……………………………………………………………………… 15 1.2. Hypothesis……………………………………………………………………….. 17 1.3. Objectives…………………………………………………………………………18 1.4. Outline of the Dissertation……………………………………………………… 18 CHAPTER 2: Retrieval of diffuse attenuation coefficient in inland waters dominated by colored dissolved organic matter…………………………………… 19 2.1. Introduction……………………………………………………………………...19 2.2. Materials and Methods………………………………………………………….. 20 2.2.1. Study Area and Sampling Planning……………………………………………………20 2.2.2. Water quality and Optical Data………………………………………………………..22 2.2.3. Kd(490) models………………………………………………………………………….. .24 2.2.4. OLI/Landsat-8 Data Processing and Acquisition…………………………………... 25 2.2.5. Statistical Analysis and Accuracy Assessment………………………………………. 26 2.3. Results and Discussion…………………………………………………………...26 2.3.1. Remote Sensing Reflectance Spectral………………………………………………… 26 2.3.2. Optical Properties and Water Constituents…………………………………………. 27 2.3.3. Assessment of the vertical attenuation coefficient models………………………… 32 2.3.4. Assessment of the OLI atmospheric correction product…………………………… 34 2.3.5. Application of kd(490) on OLI/Landsat-8 images…………………………………... 35 2.4. Conclusions………………………………………………………………………. 37 CHAPTER 3: Remotely sensed estimation of euphotic zone and Secchi disk depths in a CDOM dominated inland waters………………………………………………. 39 3.1. Introduction………………………………………………………………………39 3.2. Material and Methods…………………………………………………………... 41 3.2.1. Study Site…………………………………………………………………………………. 41 3.2.2. Planning of Sampling and Fieldworks……………………………………………….. 42 3.2.3. Field Data Collection…………………………………………………………………... 43 3.2.4. Euphotic Zone Depth Models………………………………………………………….. 45 3.2.5. Secchi Disk Depth Model………………………………………………………………. 47 3.2.6. OLI/Landsat-8 Data Acquisition and Processing…………………………………... 48 3.2.7. Statistical Analysis and Accuracy Assessment………………………………………. 49 3.3. Results……………………………………………………………………………. 49 3.3.1. In situ Measurements…………………………………………………………………… 49 3.3.2. Zeu evaluation using in situ data………………………………………………………. 51 3.3.3. ZSD evaluation using in situ data……………………………………………………… 55 3.3.4. Assessment of LaSRC…………………………………………………………………… 56 3.3.5. Rainfall Data…………………………………………………………………………….. 57 3.3.6. Zeu using satellite-derived data………………………………………………………... 58 3.3.7. ZSD using satellite-derived data……………………………………………………….. 61 3.4. Discussions……………………………………………………………………….. 64 3.4.1. Relationships between Zeu; ZSD and water quality constituents…………………… 64 3.5. Conclusions………………………………………………………………………. 67 CHAPTER 4: Final Considerations………………………………………………… 69 4.1. Highlights...………………………………………………………………………. 69 REFERENCES……………………………………………………………………….. 71 15 CHAPTER 1: Introduction 1.1 Background Inland waters embrace the reservoirs, lakes and rivers and provide important support to diverse ecosystems and habitats. The reservoirs are commonly found in the Brazilian hydropower context and are subject of anthropogenic interferences as agricultural productions, grazing lands, urban centers and waste water which alter the natural biogeochemical characteristics of the water body. The reservoirs become environments where the algal proliferation and eutrophication processes are favorable by the nutrient increments through runoff and the increase of retention time, reducing the water quality (Lira et al., 2009; Calijuri et al., 2002). The water quality is affected by the alteration of the quantities of the optical significant substances (OSS) such as the non-algal particles (NAP), phytoplankton and chromophoric dissolved organic matter (CDOM). The increase of OSS alters the turbidity and reduces the water clarity; the changes in turbidity affect the zooplankton community, increase the water temperature due to the high absorption of the sunlight by the suspended particles and reduce the dissolved oxygen rates in water column (Alcântara et al., 2010). The complex environment of the reservoir requires a qualitative and quantitative monitoring. The remote sensing based on satellite data allows the spatial-temporal study of a wide region in comparison with the traditional methods as point stations of sampling collection. Recent researches have shown suitable results in estimate water properties and constituents through remote sensing data in turbid inland waters (Yang et al., 2013; Mishra et al., 2014; Watanabe et al., 2015; Alcântara et al., 2016; Bernardo et al., 2016; Rotta et al., 2016; Zheng et al., 2016; Rodrigues et al., 2017). An important study of water clarity monitoring was developed in Minnesota lakes using a 20 years of satellite data from Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) sensors (TM/Landsat 5 and ETM+/Landsat 7 series, respectively) generating satisfactory results in relation to in situ data and revealing the potentiality of the Landsat for optical complex waters (Olmanson et al., 2008). The potentially of the Landsat systems for water clarity and CDOM measurements were evaluated in inland waters optically complex and CDOM-dominated. The OLI/Landsat-8 and ETM+/ 16 Landsat-7 presented satisfactory results (coefficient of determination – R²) for CDOM measures (R² = 0.81 and R² = 0.79, respectively) and for water clarity as the Secchi disk depth - ZSD (R² = 0.818), relating the slight better performance of OLI/Landsat-8 to the higher radiometric sensitivity (Olmanson et al., 2016). The OLI/Landsat-8 data presented a satisfactory performance (R² = 0.82) in kd retrieval in extremely turbid inland lakes (Zheng et al., 2016) and in kd (mean absolute percentage error – MAPE = 10.35%) and ZSD (linear correlation – R = 0.73) retrievals in optically variant inland waters producing good outcomes to spatial modeling of water transparency (Rodrigues et al., 2017). The Operational Land Imager (OLI) data collected by the Landsat-8 satellite present a high spatial resolution of 30 meters and a swath width of 185 kilometers with increase of the signal-to-noise ratio (SNR) indicating that the Landsat- 8/OLI is suitable to monitoring water-quality parameters in a regional scale (Zheng et al., 2016). The water clarity can be quantified from the vertical diffuse attenuation coefficient (kd) which is estimated by measuring the decrease of downwelling irradiance with depth. The kd is an intermediary product to estimate other two water clarity parameters: the euphotic zone (Zeu) and the Secchi disk (ZSD) depths. The Secchi disk is a black- and-white disk and the oldest instrument used to measure the water clarity (Lee et al., 2015). The kd is controlled by the inherent optical properties (IOPs) of water such as absorption (a) and scattering (bb) processes and the angular distribution of light; the kd at the wavelength of 490 nm is commonly derived from spectral remote sensing reflectance (Austin and Petzold, 1981; Mueller, 2000; Lee et al., 2013) and is considered as the inverse of ZSD which can be derived through satellite data in a semi- analytical model (Lee et al., 2015; Lee et al., 2016). The Zeu corresponds to the depth where the downwelling irradiance achieves 1% of that measure at the water subsurface (Kirk, 1994). Considering the homogeneity of water column, a relation between kd(PAR), the kd at the photosynthetically active radiation (PAR), is used to estimate Zeu in empirical and semi-analytical ways (Lee et al., 2007; Zhao et al., 2013; Liu et al., 2016). Lee et al. (2013) proposed a semi-analytical equation to derive kd as the inversion of the IOPs estimated via a quasi-analytical algorithm (QAA) using the fifth version (QAA_v5). Yang et al. (2014) and Rodrigues et al. (2017) applied the semi-analytical equation and found that the lowest accuracy of kd estimation and, consequently, the ZSD 17 and Zeu estimations, in turbid inland waters can be related to the propagated errors of the IOPs estimations. The QAA is a multi-band effective algorithm with easily applicable sequential steps to estimate IOPs via remote sensing reflectance (Rrs) data for optically deep waters (Lee et al., 2002). The sequence consists, firstly, in the conversion of Rrs into the subsurface reflectance (rrs) and is followed by empirical, analytical and semi-analytical steps in four levels to acquisition of a(λ) and particle backscattering coefficient (bbp(λ)) and able to separate the a(λ) in absorption coefficients of phytoplankton and CDOM (aϕ and aCDOM, respectively), allowing the estimations of Chl-a and CDOM concentrations. The empirical steps such as the estimation of a and bbp were calibrated to seawater properties limiting the application of the QAA in optically complex waters (Yang et al., 2013). Versions of QAA were created since the original version with modifications in the empirical steps and reference wavelengths (λ0) in order to improve the results. The fifth version (QAA_v5), set for λ0 = 560 nm, was applied in turbid inland waters, making possible the estimations of IOPs through adjustments in the algorithm (Le et al., 2009; Yang et al., 2013; Li et al., 2016). This work was carried out in Bariri reservoir, part of the Cascading Reservoir System (CRS) of Tietê River, São Paulo. Bariri is dominated by CDOM which compromise the phytoplankton activity, ecosystem productivity and present other effects on aquatic ecology and water chemistry that affect the water quality for human use (Zhang et al., 2009; Brezonik et al., 2015). The CDOM-rich freshwaters are greatly influenced by allochthonous source, susceptible to environmental factors such as hydrodynamic and anthropogenic activities (Zhu et al., 2014) that can compromise the kd, Zeu and ZSD estimations. 1.2. Hypothesis Considering the nature of kd, the semi-analytical algorithm developed by Lee et al. (2013) respects the dependence on the IOPs and the angular distribution light and, therefore, will be able to produce good results in inland waters through the original form of QAA_v5, set for the Landsat-8/OLI center bands, in relation to others semi-analytical and empirical models analyzed in this study. Thus, our hypothesis bases on the fact that the semi-analytical algorithm used to derive kd, even with errors related to the 18 estimative of the IOPs via QAA_v5 when applied to turbid inland waters, will enable the estimation of Zeu and ZSD through a semi-analytical scheme. 1.3. Objectives This study aimed to investigate the performance of a semi-analytical scheme to estimate Zeu and ZSD using kd estimates from remotely sensed data in Bariri reservoir. For this, the specific objectives were to: - Characterize the optical properties of Bariri reservoir; - Evaluate the performance of nine empirical and three semi-analytical models of kd to choose the more suitable one for Bariri reservoir; - Evaluate the performance of six Zeu algorithms and one semi-analytical model of ZSD using as input the kd model best fitted for the Bariri reservoir and choose the best models to map the Zeu and ZSD in Bariri reservoir; - Assess the temporal-spatial distribution of derived kd, Zeu and ZSD; - Assess the effects of CDOM and the seasonality in rainfall terms on the performance of the water clarity parameters. 1.4. Outline of the Dissertation This dissertation is organized in 4 chapters. The chapter 1 introduces the theme, delineating the problems of this research and the steps followed to answer the questions proposed. The chapter 2 corresponds to the investigation of the performance of kd models in the aquatic environment of Bariri reservoir, exploring the limiting factors of the CDOM dominance on the absorption coefficients in the evaluated models structure. The chapter 3 corresponds to the evaluation of models to retrieve Zeu and ZSD through the derived kd data and their performances in the field campaigns. The relation of Zeu and ZSD with the water constituents also was evaluated. The chapters 2 and 3 present the characterization of the study site, the description of the field campaigns, sampling planning and the optical properties characterization of the study site. Lastly, the chapter 4 highlights the main findings and challenges of this research. 19 CHAPTER 2: Retrieval of diffuse attenuation coefficient in inland waters dominated by colored dissolved organic matter 2.1. Introduction The diffuse attenuation coefficient (kd) is an apparent optical property (AOP) and is defined as the exponential decrease of the ambient downwelling irradiance (Ed) with depth, therefore is related to light penetration and availability and can be used to predict the euphotic depth. AOP are those properties that depend both on the medium and on the geometric structure of the radiance distribution (Mobley, 1994). kd is largely determined by the inherent optical properties (IOPs), absorption (a) and backscattering (bb) coefficients in first order and, in lesser magnitude dependent on the incident radiation field as the Sun angle. IOPs are properties of the medium and do not depend on the ambient light field (Mobley, 1994). Non-algal particles (NAP), phytoplankton, chromophoric dissolved organic matter (CDOM) and water are considered the four optically significant substances (OSS) that control the kd. The kd at 490 nm [kd(490)] is generally considered and classified as a parameter of water quality, therefore is essential for monitoring the eutrophication process due to light attenuation by phytoplankton growth or suspended matter (Zheng et al., 2016). Since light availability is a critical regulator of physical, chemical and biological processes the accurate estimation of kd is critical to better understanding and modeling primary productivity, heat and gas transfer in aquatic systems. In order to estimate kd(490), some empirical algorithms were developed using a direct form from normalized water-leaving radiance (nLw) or remote sensing reflectance (Rrs) ratios or indirect form through products based on chlorophyll-a concentrations ([Chl-a]) (Morel et al., 2007; Mueller, 2000; Mueller and Tress, 1997; Chauhan et al., 2003; Werdell, 2005; Kratzer et al., 2008; Zhang and Fell, 2007; Wang et al., 2009). However, these models are considered site specific, which is a limiting factor for a broader application. Wang et al. (2009) and Lee et al. (2013) came up with a semi- analytical approach covering a wide range of waters. Wang et al. (2009) used a combination of two models suitable for ocean and turbid coastal waters, aiming to improve the application range of kd(490) estimations. The model from Lee et al. (2013) considered the IOPs derived from Rrs through a quasi-analytical algorithm QAA (Lee et al., 2002), and the solar zenith angle using the radiative transfer theory. The algorithm 20 make it possible the estimation of spectral kd, as well as kd(490), for water bodies ranging from the clearest ocean to turbid coastal waters. Even so, studies showing the use of remote sensing for kd(490) estimation in tropical productive inland waters are hampered by the lack of in situ data and theoretical framework to predict and interpret ocean color data in such waters (Gallegos et al., 1990). The optically complexity of turbid waters (mineral suspended solids, algae and associated organic particles) produce multiple scattering making it difficult the mathematical analysis of radiative transfer (Gallegos et al,. 1990). From our understanding there are no investigation of kd(490) estimation in inland waters dominated by CDOM, which is an optically active component of dissolved organic matter and plays an important role in the cycling carbon. CDOM is considered an important water quality indicator due to its impact on the drink water, carbon balance and aquatic ecosystems; mainly because CDOM affects penetration of photosynthetically active radiation into the water column, which affects the primary productivity. Most models use Rrs or nLw from the blue-green spectral region, in which the CDOM has high absorption. Because of that the following questions came up: (1) The kd(490) models that use the blue/green ratio will be impacted in such way that the errors will make it impossible to use? (2) How will these models be impacted in the presence of high [Chl-a]? (3) Semi-analytical models, which use IOPs as input data will perform better than the empirical ones? (4) What are the perspectives for kd monitoring from space in inland waters dominated by CDOM? Therefore the main goal of this paper was to assess the performance of algorithms to estimate kd(490) in inland waters dominated by CDOM, which are widely available to end users. We tested nine empirical and three semi-analytical algorithms. The selected models will help us to answer the above questions. 2.2. Materials and Methods 2.2.1. Study Area and Sampling Planning Bariri hydroelectric reservoir (22° 9‟ 49.260” S 48° 44‟ 21.420” W) is in the middle of the São Paulo State and is part of the Cascading Reservoir System (CRS) of the Tietê River (Figure 2.1). Bariri is the second of a total of six reservoirs and presented the smallest flooded area of 63 km² in an average altitude of 450 meters. The Bariri 21 reservoir is situated in a tropical climate with a dry period (April-September) and a wet period (October-March) according to Köppen classification. The water retention time varies among 7 and 24 days (Tundisi et al., 2008). Figure 2.1. Graphics showing the location of (a) São Paulo State in Brazilian context; (b) Bariri Reservoir in the Cascading Reservoir System of Tietê River; (c) sampling points of the two field campaigns carried out, BAR1 and BAR2 and (d) precipitation rate data (mm) from NASA‟s GIOVANNI database for the period of 2014-2017 (TRMM Data Product; Spatial Resolution of .0.25º; Monthly Temporal Resolution). In 2017, the precipitation monthly averages were available just until July month. The amount of wastewater coming from the metropolitan region of Tietê River characterizes the Bariri Reservoir as highly productive water with high average concentrations of total nitrogen (2750 μg L -1 ), phosphorus (87 μg L -1 ) and [Chl-a] (55.8 μg L -1 ) with the phytoplankton community dominated by cyanobacteria. As the second reservoir, the Cascading Reservoir Continuum Concept (CRCC) effect promotes the heterogeneity of the eutrophication levels in upstream to downstream in Bariri, revealed by the water transparency of the water (Barbosa et al., 1999). Two fieldworks were carried out where in the first one (BAR1), 30 samples were collected from 15 to 18 August 2016 and in the second one (BAR2), 18 samples were 22 taken from 23 to 24 June 2017 (see Figure 2.1 for samples locations). The days of the field campaigns were determined according to the temporal resolution of OLI/Landsat-8 aiming to match the data collection with the satellite overpass the study site. 2.2.2. Water quality and Optical Data The wind speed (m s -1 ), pH, depth (m), Secchi disk depth (ZSD; m) and turbidity (NTU) were measured in the field with an anemometer, pHmeter, Secchi disk and turbidity meter, respectively. The water samples were collected in the field to derive the OSS, such as the Chl-a and the suspended particulate matter (SPM) as well as the inorganic (IPM) and organic particulate matter (OPM) concentrations according to Golterman et al. (1978) and the American Public Health Association Protocol (APHA, 1998), respectively. The IOP data analysis followed the methodology proposed in Bricaud et al. (1981) for aCDOM while for the absorption coefficients of non-algal particle (aNAP) and the total particulate (ap), that consists in the sum of aNAP and aϕ, the acquisition and analyses were made in accordance with the Transmittance-Reflectance method according to Tassan and Ferrari (1995, 1998, 1999). The radiometric data were obtained from the two hyperspectral RAMSES sensors (TriOS, Rastede, Germany). The sky and total radiance data (Lsky and Lt, respectively, both in W m -2 sr -1 ) and the downwelling and sky irradiance data (Ed and Esky, respectively, both in W m -2 ) were taken with the sensors fixed by a steel frame in the boat, in a configuration of 40° from nadir (zenith) and to azimuthal angle of 90° in order to minimize the specular reflection (Mobley (1999). The radiometric quantities were used to calculate the remote sensing reflectance according to Equation (2.1). = (2.1) where + is the downwelling irradiance measured at the water surface. The LSR is the surface-reflected radiance and consists in a multiplicative product of Lsky and ρ - an effective surface reflectance. The ρ value ( 0.02-0.05) depends on the viewing geometry and spectral variation. In agreement with that, Lee et al. (2010) proposed a calculation approach of Rrs as a function of the total remote-sensing reflectance (Trs, 23 ratio of Lt to Ed) and sky remote-sensing reflectance (Srs, ratio of Lsky to Ed) for each Lt and Lsky scan from the Equation (2.2): (2.2) where F refers to Fresnel reflectance and it was set as 0.021 according to the viewing geometry; is a spectrally constant settled before can be derived (Lee et al., 2010). In oceanic waters, the is negligible in the red and near-infrared wavelengths and can be assumed as zero beyond 700 nm, however, for turbid inland waters, the IOPs have a significant influence in , thus, one alternative is to model the spectral based on the IOPs, and then resolve Δ after comparing modeled and derived from Equation (2.2). In sequence, the Rrs used in estimates models of kd must be simulated to the satellite signal at each spectral channel centered at a wavelength (λ = 443; 482; 561; 655 nm). The band simulation consists in the convolution of the radiation signal of hyperspectral sensor and the spectral response function [Fr(λ)] of the OLI sensor in wavelength interval of the spectral resolution (Barsi et al., 2014): ∫ ∫ (2.3) where is the remote-sensing reflectance simulated at center wavelength; xmax and xmin are, respectively, the maximum and minimum values of the sensor spectral channel. According to Ed data, the kd(490) in both field campaigns was determined as the slope of Ed at subsurface depth (z) (Mobley, 1994): (2.4) In order to eliminate the noise in the due to changes in the sun illumination condition caused by cloud cover during the measurements, a normalization factor was required in all scans. The Ed normalization consists in a division of Esky at first scan t(z1) to Esky at subsequent scans t(zi) as factor normalization of the t(zi) according Mishra et al. (2005) and Mueller (2000): 24 (2.5) where is the normalized 2.2.3. Kd(490) models The Table 2.1 summarizes the kd(490) algorithms tested in this study. Table 2.1. kd(490) models developed for clear ocean, coastal, turbid ocean, slightly turbid and global waters. * kw(490) = 0.016m -1 is a constant of the diffuse attenuation coefficient for pure sea water (Zhang and Fell, 2007). Type Model Formula Calibration Dataset Mueller (2000) Mueller and Tress (1997) Chauhan et al. (2003) Werdell (2005) Global Waters Kratzer et al. (2008) Coastal Waters Zhang and Fell (2007)* Turbid Ocean Waters Wang et al. (2009) Slightly Turbid Waters Lee et al. (2013) Global Ocean Waters Semi-analytical model Wang et al. (2009) Empirical model with normalized water-leaving radiance (nL w ) or remote sensing reflectance (R rs ) Clear Ocean Waters Slightly Turbid Coastal Waters Morel et al. (2007) Empirical model with Chl-a concentration Clear Ocean Waters + x x (λ) 25 We have tested three types of model: (1) empirical relationships between the kd and [Chl-a]; (2) empirical relationship between the water-leaving radiance (nLw) and kd and (3) semi-analytical models which are based on radiative transfer models. The nLw is calculated trough the conversion of as where is the extraterrestrial solar irradiance (Wang et al. 2009; Zhao et al., 2013). The R(λ) is the irradiance reflectance beneath the water-surface and is calculated as a function of extraterrestrial solar irradiance [F0(λ)] and (Wang et al., 2009). The equations from Morel et al. (2007) are based on [Chl-a] whereas the others empirical algorithms use a simple ratio of (λ). The Zhang and Fell (2007) and Wang et al. (2009) are the unique empirical models based on a ratio. The semi- analytical model of Wang et al. (2009) was developed using the R(λ) ratios and the Lee et al. (2013) model previously requires the calculation of a(λ) and bb(λ) which are obtained from the QAA_v5 (the version 5 was used in this study) with as input at the reference wavelength ( of 561 nm, the nearest OLI center band of = 55x of QAA_v5. In the equation, the zenith solar angle (θs) was set as 40º following the geometry used in the radiometric data collection. 2.2.4. OLI/Landsat-8 Data Processing and Acquisition The Landsat 8 Surface Reflectance on-demand data generated by the Landsat Surface Reflectance Code (LaSRC) were obtained in the U.S. Geological Survey platform (http://earthexplorer.usgs.gov/). The LaSRC algorithm for atmospheric correction was developed using the Second Simulation of the Satellite Signal in the Solar Spectrum Vectorial (6SV) model. The algorithm uses the OLI Coastal Aerosol Band (0.433–0.450 μm) which works as cover for shorter wavelength as the blue band in previous Landsat and is helpful to retrieve aerosol properties (Vermote et al., 2016). Pahlevan et al. (2017) and Bernardo et al. (2016) showed that LaSRC is a consistent product to derive aquatic estimates. A total of 10 atmospherically corrected images (Path/Row 221/75) were acquired during the year of 2016, covering the months from February to November. The criteria used for the images choice was attributed to cloud free data over the reservoir. To obtain the nLw(λ) and R(λ) and to test the kd(490) models, the surface reflectance images were divided by π to convert them into Rrs (Moses et al., 2012). The kd(490) model with the best performance was applied to the time-series of LaSRC 26 images selected to obtain the spatial distribution of the vertical attenuation coefficient. The kd(490) model best fitted for Bariri reservoir was validated through the use of the satellite image from August 15 th 2016, correspondent to the first day of BAR1. In the same way, the LaSRC accuracy in kd(490) estimations was evaluated from the analysis between the Rrs converted images with the Rrs calculated from TriOS data in BAR1. 2.2.5. Statistical Analysis and Accuracy Assessment Statistical analyses, including calculations of the maximum, minimum and average values and linear and non-linear regressions were performed. The kd(490) analyzed models were applied for BAR1 and BAR2 datasets. The Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE) and the bias were used to assess the accuracy of the kd(490) models. √ ∑ ( ) (2.6) ∑ | | (2.7) ∑ ( ) (2.8) where is the kd(490) estimated value and is the in situ measure of kd(490). 2.3. Results and Discussion 2.3.1. Remote Sensing Reflectance Spectra In BAR1 spectral curves (Figure 2.2) is possible to see a high absorption feature at about 680 nm and a reflectance peak at approximately 710 nm that can be associated with the Chl-a pigment. The sampling points with the highest [Chl-a] presented a significant peak at 550 nm, expected for water with a great amount of algae, just as seen in field campaign. The increase of reflectance in longer wavelength was an indicative of total suspended matter concentration increases. The [Chl-a] in BAR1 was approximately 37 times higher than in BAR2 dataset (see Table 2.2), which explains why the absorption by pigments was more evident during the first survey. 27 In all data set, the organic particles were predominant (Table 2.2), therefore the spectra of BAR1 and BAR2 were quite similar with lesser reflectance magnitude of BAR2 mainly because lowest OSS concentrations. In BAR2 spectra, the feature characteristic of phytoplankton absorption was not evident such as was for BAR1 due to the decrease of [Chl-a] in relation to BAR1, with the exception of the feature of the sampling point that presented the maximum value of [Chl-a] in BAR2 (19.11 μg L-1). Figure 2.2. The Rrs spectra for BAR1 (a) and BAR2 (b) surveys. The light absorption by CDOM was the highest in both field campaigns (1.57 0.37 m -1 in BAR1 and 1.98 0.76 m -1 in BAR2) in relation to light absorption by others OSS. In CDOM-rich lakes, the spectra shapes vary according to CDOM levels and the influence of other substances in absorption processes (Brezonik et al., 2015). In BAR1, the light absorption was dominated by CDOM, however the reflectance spectra assumed a shape of Chl-a due to the high concentrations and presented low reflectance values at 400-500 nm. In BAR2, with the predominance of CDOM absorption and reduction of OSS concentrations, the reflectance values assumed a nearly flat feature at 600 nm. 2.3.2. Optical Properties and Water Constituents Table 2.2 presents the descriptive statistics for optical and water parameters for the two field campaigns, BAR1 and BAR2. In BAR1, the sky was sunny in the most of the days with some moments of overcast. In BAR2, the sky was more favorable with sunny days during all field campaign. In all dataset, the average wind speed was 3.32 1.77 m s -1 producing some small waves on the water surface. The minimum values for both 0 0.005 0.01 0.015 0.02 0.025 0.03 400 500 600 700 R rs ( sr -1 ) Wavelength (nm) 0 0.005 0.01 0.015 0.02 0.025 0.03 400 500 600 700 R rs ( sr -1 ) Wavelength (nm) (a) (b) 28 field campaigns were 0 m s -1 and the maximum value was 8 m s -1 for BAR1 and 6.50 m s -1 for BAR2. Table 2.2. Descriptive statistics for optical and water quality parameters for BAR1, BAR2 and the mixed data. S.D. = Standard Deviation. C.V. = coefficient of variation. Field Campaigns Parameters Minimum Maximum Mean S.D. C.V. BAR1 (N = 30) ZSD (m) 0.50 1.60 1.16 0.23 20.03% Turbidity (NTU) 7.80 80.90 16.60 7.61 45.82% pH 6.10 9.90 7.94 0.83 10.46% Wind Speed (m s -1 ) 0.00 8.00 3.47 1.80 51.73% kd(482) (m -1 ) 1.87 3.98 2.80 0.30 10.69% [Chl-a] (μg L -1 ) 25.67 709.89 119.76 96.43 80.52% [SPM] (mg L -1 ) 3.60 40.33 8.40 4.64 55.25% [IPM] (mg L -1 ) 0.90 4.00 2.35 0.51 21.92% [OPM] (mg L -1 ) 1.40 36.33 6.06 4.57 75.43% aphy(482) (m -1 ) 0.16 1.25 0.41 0.22 53.46% aNAP(482) (m -1 ) 0.15 0.55 0.33 0.09 28.52% aCDOM(482) (m -1 ) 0.92 2.74 1.57 0.37 23.73% BAR2 (N = 18) ZSD (m) 1.60 2.50 2.06 0.19 9.31% Turbidity (NTU) 3.48 8.80 5.72 1.25 21.92% pH 6.83 7.28 6.97 0.13 1.89% Wind Speed (m s -1 ) 0.00 6.50 3.06 1.72 56.28% kd(482) (m -1 ) 1.54 2.34 1.79 0.12 6.47% [Chl-a] (μg L -1 ) 3.82 19.11 7.99 3.27 40.90% [SPM] (mg L -1 ) 0.20 2.60 1.59 0.44 27.92% [IPM] (mg L -1 ) 0.24 1.30 0.58 0.24 42.39% [OPM] (mg L -1 ) 0.40 1.60 1.11 0.32 28.81% aphy(482) (m -1 ) 0.02 0.11 0.07 0.02 26.67% aNAP(482) (m -1 ) 0.06 0.19 0.12 0.03 26.39% aCDOM(482) (m -1 ) 0.76 4.38 1.98 0.76 38.17% All data (N = 48) ZSD (m) 0.50 2.50 1.49 0.43 29.03% Turbidity (NTU) 3.48 80.90 12.66 6.34 50.03% pH 6.10 9.90 7.58 0.71 9.39% Wind Speed (m s -1 ) 0.00 8.00 3.32 1.77 53.37% kd(482) (m -1 ) 1.55 3.98 2.42 0.51 21.10% [Chl-a] (μg L -1 ) 3.82 709.89 77.84 77.15 99.10% [SPM] (mg L -1 ) 0.20 40.33 5.79 4.05 69.82% [IPM] (mg L -1 ) 0.20 4.00 1.96 0.79 40.25% [OPM] (mg L -1 ) 0.40 36.33 4.99 3.96 79.52% aphy(482) (m -1 ) 0.02 1.25 0.28 0.19 67.72% aNAP(482) (m -1 ) 0.06 0.55 0.25 0.12 45.70% aCDOM(482) (m -1 ) 0.76 4.38 1.72 0.52 30.36% 29 The water collected in BAR1 was green in most of the samples, which is explained by the presence of phytoplankton pigments due to high [Chl-a] averaging 119.76 ± 96.43 μg L -1 , and ranging between 25.67 and 709.89 μg L -1 . In BAR2, the [Chl-a] variability reduced significantly with average of 7.99 3.27 μg L -1 , varying from 3.82 to 19.11 μg L -1 . Therefore, the mixed data resulted in a range from 3.82 to 709.89 μg L -1 of [Chl-a]. The SPM concentration (average of 8.40 4.64 mg L -1 ) showed predominance of organic particles with average of 6.06 4.57 mg L -1 in BAR1 as well as in BAR2, where the organic compounds predominated the SPM (1.59 0.44 mg L -1 ) with average of 1.11 0.32 mg L -1 . The SPM presented low concentration in BAR2. In all dataset, the organic particle dominated the SPM concentration, with an average of 4.99 3.96 mg L -1 . In respect to absorption, aNAP(482) was higher in BAR1 (average of 0.33 0.09 m -1 ) than BAR2 (average of 0.12 0.03 m -1 ). The aCDOM(482) averaged 1.57 0.37 m -1 in BAR1 and 1.98 0.76 m -1 in BAR2. The aφ(482) averaged 0.41 0.22 m -1 in BAR1 and 0.07 0.02 m -1 in BAR2. The Figure 2.3 shows the average spectra of IOPs for BAR1 and BAR2 dataset. Figure 2.3. Absorption spectra of NAP, phytoplankton, CDOM and pure water (aw) in BAR1 (a) and BAR2 (b). The aCDOM spectra presented the highest values in all spectral range in both field campaigns. On the other hand, the aφ spectra was higher for BAR1 than BAR2; it is possible to see two peaks of absorption at 450 and 670 nm in BAR1 whereas in BAR2, the characteristic aφ spectra was not evident. In BAR2, the aφ and aNAP spectra were similar with light absorption by NAP slightly higher than the absorption by phytoplankton, with the absorption curve decay at 550 nm. The absorption curve of 0.00 1.00 2.00 3.00 4.00 5.00 400 500 600 700 A b so rp ti o n ( m -1 ) Wavelength (nm) aNAP aφ aCDOM aw 0.00 1.00 2.00 3.00 4.00 5.00 400 500 600 700 A b so rp ti o n ( m -1 ) Wavelength (nm) (a) (b) 30 NAP at blue portion of the visible spectrum represents the typical spectra of mineral and/or detrital particles, verified in BAR1 and BAR2 absorptions spectra. The atypical aφ spectra in BAR2, with the exponentially increase at wavelengths shorter than 500 nm, indicate an environment with elevated CDOM absorption (Binding et al., 2008). Among the three absorption coefficients already mentioned, aCDOM(482) showed the highest values in both field campaigns. Zhang et al. (2009) proved that high [Chl-a] and high aCDOM indicate that the accumulation and degradation of phytoplankton were a source of CDOM in eutrophic waters. It suggests that the organic matter is possibly originated by the phytoplankton degradation in BAR1. In BAR2, the aCDOM is also predominant with significant reduction of [Chl-a] which can suggest an allochthonous source of CDOM. The Bariri reservoir has been undergoing impacts due to the sugarcane production and the input of urban and industrial wastewaters (Pamplin, 2004). The proportional contribution of the IOPs in BAR1 and BAR2 among the total absorption (at) budget in Bariri aquatic environment was computed considering the OLI bands (443, 482, 561 and 655 nm) (Figure 2.4a-d). The light absorption in BAR1 was predominant by CDOM in all wavelengths, with 63.49% 11.59%; 67.74% 9.98% and 69.82% 12.56% at 443, 482 and 655 nm, respectively. At 561 nm, the CDOM contribution was the highest with 76.08% 8.84%. The same was verified for BAR2, where the CDOM predominance was even higher with 88.28% 5.43% at 443 nm; 88.85% 5.70% at 482 nm; 92.07% 5.24% at 561 nm but with the highest contribution at 655 nm with 94.02% .6.03%. In BAR1, the phytoplankton contributed with 22.40% 13.06%, 17.40% 10.84%, 13.84% 8.36% and 25.10% 12.16% at 443, 482, 561 and 655 nm, respectively. The highest proportions at the blue and red spectral regions confirm the absorption peaks by Chl-a pigment. In sequence, the NAP presented the lowest proportion of the absorption budget with 14.11% 4.80% at 443 nm; 14.86% 5.27% at 482 nm; 10.07% 4.40% at 561 nm and 5.08% 2.99% at 655 nm. The peak of NAP absorption in the blue spectral region is typical of detrital or mineral particle which was also verified in BAR2, with 6.85% 3.31% and 6.91% 3.89% at 443 and 482 nm, respectively, and 5.03% 3.77% at 561 nm and 2.99% 3.01% at 655 nm. The phytoplankton achieved 4.86% 2.30% at 443 nm and 4.24% 2.05% at 482 nm, 31 being higher in the blue spectral region. At 561 nm, the percentage was 2.90% 1.65% and at 655 nm, 2.99% 3.01%. Figure 2.4. Ternary plots depicting BAR1 and BAR2 for center OLI bands (a) 443 nm, (b) 482 nm, (c) 561 nm and (d) 655 nm. Considering both dataset, the [Chl-a] had an average of 77.84 77.15 μg L -1 and the organic particles had the highest concentration (4.99 3.96 mg L -1 ), with the greater part of light absorbed by CDOM. The average of kd(490) in BAR1 and BAR2 was 2.80 0.30 m -1 (ranging from 1.87 to 3.98 m -1 ) and 1.79 0.12 m -1 (ranging from 1.54 to 2.34 m -1 ), respectively. The water column was more attenuated in August/2016 than June/2017. Both months are in the dry season; however, 2016 presented the highest mean rate of precipitation from January to June when compared with 2017 over the Bariri reservoir area (Figure 2.1d). The rain carries organic matter, mainly from industrial effluents; wastewater and particulate matter to the Bariri reservoir, as well as facilitates the increasing of organic matter fluxes along the Tietê River from upstream to downstream. The Tietê River is one of the most industrialized basin of São Paulo State and the reservoirs construction promoted the rapid transformation in the land use and land cover, facilitating the pollution problems and accelerating the sedimentation process (Prado et al., 2007). The 32 concentration of water constituents was higher in August, increasing the process of absorption and backscattering of light and, consequently, showing the highest kd(490) values. 2.3.3. Assessment of the vertical attenuation coefficient models Due to the lack of data from turbid inland waters used in the calibration of kd(490) algorithms, the adjustment for Bariri reservoir can be compromised. The algorithm coefficients developed for a wide range of waters tries to minimize this limitation although it is not a guarantee of success (Lee et al., 2005). For all kd(490) algorithms, the bands were defined according to the center bands from OLI/Landsat-8, including the blue band at 490 nm that was replaced by 482 nm. The Table 2.3 summarized the statistical results yielded after the application of kd(490) models. Table 2.3. Error assessment of kd(490) models developed for clear, slightly turbid, turbid ocean waters, coastal waters and global waters applied to all dataset of Bariri. Type Model MAPE (%) RMSE (m -1 ) bias (m -1 ) Morel et al.(2007) I 60.96 1.48 -1.21 Empirical Morel et al.(2007) II 61.03 1.50 -1.19 Mueller (2000) 86.46 2.19 -2.12 Mueller and Tress (1997) 86.15 2.19 -2.11 Chauhan et al. (2003) 82.35 2.04 -1.86 Werdell (2005) 86.05 2.18 -2.11 Kratzer et al. (2008) 60.62 1.63 -1.52 Zhang and Fell (2007) 42.90 1.22 -1.07 Wang et al. (2009) 54.04 1.39 -1.15 Semi- analytical Wang et al. (2009) I 46.34 1.31 -1.18 Wang et al. (2009) II 63.68 1.69 -1.59 Lee et al. (2013) 41.04 1.07 -0.90 The empirical models of kd(490) based on [Chl-a] underestimated the results for almost the entire dataset, except for two samples with high values of [Chl-a] which were overestimated. The coefficients of the model from Morel et al. (2007) were developed for Case I waters, with the optical properties dominated by phytoplankton, therefore, the accuracy of the model in aquatic environments with high rates of CDOM and/or suspended solids is expected to be fail (Zhao et al., 2013). The models from Morel et al. (2007) were calibrated with open ocean waters with values of [Chl-a] < 2.4 33 μg L -1 , highly discrepant of the values found in Bariri reservoir, resulting in kd(490) errors of MAPE 61% and RMSE 1.50 m -1 . The empirical models using nLw spectral ratio developed for ocean open and global waters presented the worst results with MAPE around 80% and RMSE 2 m -1 . Mueller (2000); Mueller and Tress (1997), Chauhan et al. (2003) and Werdell (2005) models underestimated the results of kd(490) with bias around 2 m -1 . The calibration dataset involved waters with values of kd(490) up to 0.61 m -1 whereas in Bariri all dataset varied from 1.55 to 3.98 m -1 . The blue-green ratio nLw(490)/nLw(555) or Rrs(490)/Rrs(555) has a large uncertainties when kd(490) is greater than 0.25 m -1 . The spectral ratio presents an asymptotic value with increasing OSS concentration (Mueller, 2000). Thus, this spectral ratio is not sensitive for the variations in OSS that occurred in turbid inland waters, resulting in significant underestimation of kd(490). In addition, this spectral ratio does not consider the effects of sun angle changes that decrease the accuracy of empirical algorithms for estimating kd(490) (Lee et al., 2005). Aiming to investigate the estimation of kd(490) from MERIS, Kratzer et al. (2008) developed the empirical algorithm from the ratio of 490 nm and 620 nm for a Case 2 water optically dominated by CDOM which showed inversely relation to salinity in Baltic Sea. The Kratzer et al. (2008) model applied to all dataset from Bariri presented MAPE of 60.62% and RMSE of 1.63 m -1 . The empirical algorithm developed by Wang et al. (2009) combined open ocean and turbid coastal waters (from 0.3 to 0.6 m -1 ) through a linear regression equation of Rrs(670)/Rrs(490) ratio that showed better matching of kd(490) data when the kd(490) values of calibrated data were higher than 0.3 m -1 . The application of this model in Bariri reservoir data resulted in a MAPE of 54.04% and RMSE of 1.39 m -1 , yielding underestimated values (bias of -1.15) of kd(490). The semi-analytical algorithms of Wang et al. (2009) were developed considering the MODIS satellite data and the absorption and backscattering coefficients derived from Lee et al. (2002). The algorithm based on the ratio R(667)/R(488) presented MAPE of 46.34% and RSME of 1.31 m -1 , while the ratio R(645)/R(488) showed MAPE around 63.68% and RMSE of 1.69 m -1 . The OLI sensor does not have a spectral band centered at 488 neither at 667 nm or 645 nm, but at 482 and 655 nm, respectively. Therefore, in both semi-analytical equations, the R(482) and R(655) were used and the 34 differences found in the results above are explained by the coefficients, which in the first model presented better adjustment than the second model in Bariri reservoir. The Zhang and Fell (2007) empirical algorithm and the semi-analytical algorithm developed by Lee et al. (2013) presented similar performances corresponding to MAPE of 42.90%; RMSE of 1.22 m -1 and MAPE of 41.04%; RMSE of 1.07 m -1 , respectively. Calibrated with a wide range of environments (0.016 – 4.6 m -1 ), the Zhang and Fell (2007) model have showed greater correlation for clear waters whereas the Lee et al. (2013) algorithm showed the greatest application for turbid waters such as happens for this study site. The difference in errors among these empirical and semi-analytical algorithms could be explained by the use of 655 nm wavelength in Zhang and Fell (2007) model which introduce noise due to the high absorption rate of water in the red wavelength even after corrections in data processing. The Lee et al. (2013) semi-analytical algorithm showed the lowest errors (MAPE = 41.04%; RMSE = 1.07 m -1 ) for kd(490) estimation in Bariri reservoir. In inland waters, the regional variability of the OSSs generates significant changes in the attenuation of light process, therefore, the Lee et al. (2013) algorithm in considering a( ), bb( ) and Sun angle resulted in a fewer uncertainties regarding data matching. Besides, the algorithms coefficients derived from numerical simulations reduce the dependence of a specific data range. Empirical algorithms were developed (results not shown here) from Bariri using all dataset, however the results were not satisfactory, therefore Lee et al. (2013) model was chosen for mapping the variability of kd(490) in the study site. 2.3.4. Assessment of the OLI atmospheric correction product The evaluation of LaSRC accuracy consisted in the comparison between the OLI- derived Rrs product from August/2016 and the Rrs calculated from TriOS data in BAR1. The relative percentage errors were 50.09% at 443 nm, 20.68% at 482 nm; 17.24% at 561 nm and 29.64% at 655 nm. The highest errors at the coastal and blue regions are related to the increase of scattering in those regions. The lowest error was verified at green region with 17.24% and, in sequence at red region, the error increased again to 29.64%. The LaSRC accuracy results encountered here are compatible with the atmospheric correction analysis presented in Rodrigues et al. (2017) and Bernardo et al. (2017). 35 2.3.5. Application of kd(490) on OLI/Landsat-8 images The variability of kd(490) in Bariri reservoir expressed in OLI/Landsat-8 was presented in Figure 2.5. Figure 2.5. Spatial distribution of kd(490) in Bariri reservoir using the semi-analytical model from Lee et al. (2013) considering the months from February to November of 2016. The arrows A and B highlight the reservoir zones near urban centers. The spatial distribution of Lee et al. (2013) in Bariri reservoir was validated from the kd(490) estimations obtained from the image of 15 th August and presented MAPE of 32.54% and RMSE of 0.95 m -1 . The errors of kd(490) estimations from Lee et al. (2013) model applied for the all dataset, using the in situ Rrs, were 41.04% of MAPE and 1.07 m -1 of RMSE. The differences between the kd(490) estimations from in situ Rrs and Rrs from LaSRC are related to the decrease of the variability among the sampling points through the sample size reduction, that in the former case was 48 (all dataset) and in the latter case was 26 (4 sampling points with missing information in the image from August in region highlighted as zone B in Figure 2.5) that presented the lowest errors. The increase of the percentage error between the OLI-derived Rrs at 482 nm used as input of kd(490) estimation and the kd(490) estimations for BAR1 was 11.86% and A B 36 could be related to IOPs estimations via QAA_v5. The application with no modifications in some empirical steps can affect the QAA_v5 performance in complex optically waters (Watanabe et al., 2016; Li et al., 2016). The maximum value of kd(490) was 5.60 m -1 in April and the minimum value was 0.89 m -1 in September, both in the dry season. In the image from April, specifically in the zone B highlighted by the arrow, an expressive increase of the kd(490) was observed, which is probably related to the reduction of water velocity favoring the increase of [Chl-a] and, in a drastically way, an algal bloom, intensifying the attenuation of light in the water. Along the reservoir, excepting for zone B, the kd(490) values remained below the mean (Figure 2.6), which was expected for a period of low precipitation, confirming an atypical event in April. In September, the kd(490) was uniformly distributed around the mean, with values closer the maximum (4.40 m -1 ) in the northeast region of the reservoir which presents agricultural activities and bare soil areas that facilities the input of external material into the reservoir even in periods of low precipitation. Other regions of high values of kd(490) were in zone B and another winding points where the water fluxes is slower than in zone A. Figure 2.6. Boxplot of kd(490) spatial distribution regarding the months of February to November of 2016 in Bariri reservoir. The boxplot in Figure 2.6 summarizes the statistical performance of kd(490) for each month. In dry season, the kd(490) values were closer to the maximum in May and June, with the mostly values above the mean and, concomitantly values above 2.50 m -1 in zone A. In July and August, the kd(490) started to decrease with mostly of the values closer to the minimum and below the mean, with also decreasing of kd(490) in zone A. The precipitation rates were quite smaller in July ( 3.99 mm) and August ( 53.83 0.00 1.00 2.00 3.00 4.00 5.00 6.00 M a p p ed k d (4 9 0 ) (m -1 ) 37 mm) than in May ( 125.25 mm) and June ( 125.43 mm) in conjunction with reducing of kd(490) in zone A which could be revealed a point source of wastewater. Unfortunately, the image from August, coincident with the BAR1 field campaign, has no information in the region highlighted as zone B that prevent to show the highest Chl-a concentrations in these sampling points which generated high values of kd(490) as mentioned in section 2.3.1. The mean was 1.48 m -1 , with values closer to the minimum, ranging from 1.30 to 1.73 m -1 , not corresponding with in situ kd(490). The precipitation rate in August was very low, resulting in low kd(490) in zone A. The months of October, November, February and March are representative of wet season, however in October the mean kd(490) almost coincided with the minimum resulting in low values with some peaks in reservoir northeast and in a winding region. In November, the increase of kd(490) was observed with the mean closer to maximum value, presenting some points above 2.50 m -1 . In February, with higher precipitation rates than the months previously mentioned, the kd(490) was around 3.00 m -1 even with approximately 30% of the reservoir covered by clouds. In March, the kd(490) was well distributed around the mean, ranging from 1.69 to 4.87 m -1 , with the highest values in northeast and extremely north of reservoir and values above 2.50 m -1 in some part of zone B and in zone A. 2.4. Conclusions After the application of several tests with the kd(490) estimation models available for a wide range of waters, it was verified that the empiricism does not allowed to estimate the kd(490) in an accurate way in Bariri reservoir. The uncertainty presented by the empirical models ( 80% of MAPE) confirms that the blue/green wavelength ratio is not able to express the kd(490) variability in an environment dominated by CDOM which was verified in the two surveys realized. The [Chl-a] empirical models presented intermediary errors ( 60% of MAPE) when applied with the aggregated data from the two surveys realized. In BAR1, the [Chl-a] were significantly higher than BAR2; the light absorption by CDOM was maintained predominant in both surveys. Among the tested empirical models, the model developed by The Zhang and Fell (2007) was an exception achieving lower errors and statistical results close to those obtained from the semi-analytical models. 38 The semi-analytical model developed by Lee et al. (2013) presented the lowest error (MAPE of 41.04%) among the 12 models analyzed. The coefficients present in Lee et al. (2013) model were calibrated from a wide range of waters, filling the lack of the empirical models that are based on a specific dataset. However, the IOPs estimations through the original QAA_v5 can introduce some errors in inland waters due to the optical complexity. The Lee et al. (2013) model was applied in atmospherically corrected OLI/Landsat- 8 images (LaSRC) with a relative error percentage difference of 8.50% in relation to the application to all dataset from the Rrs obtained in situ. Therefore, the atmospheric correction was appropriate to retrieve the kd(490) and the Lee et al. (2013) model was able to highlights the main variations of kd(490) in an environment dominated by CDOM as the Bariri reservoir, allowing identify the regions with more or less light attenuation and, consequently, the biota modifications that are dependent of the photic conditions in the water column. The application of kd(490) model on OLI/Landsat-8 images exhibited important temporal and spatial variability. The spatial variability was verified in points where the sinuosity of reservoir promoted the reduction of water velocity and facilitates the OSS concentrations, increasing the attenuation of light. The temporal variability was linked to significant activities around the reservoir that increase the runoff and the input of external material into the reservoir mainly in periods of high precipitation rates. Besides, the precipitation in wet season promotes the resuspension and carriage of sediments, increasing the attenuation of light. 39 CHAPTER 3: Remotely sensed estimation of euphotic zone and Secchi disk depths in a CDOM dominated inland waters 3.1. Introduction The penetration and availability of underwater light are controlling factors of biological (phytoplankton photosynthesis), chemical (nutrient cycling) and physical (heat transfer) processes (Kirk, 1994). The water transparency is an important parameter for environmental monitoring and water quality (Al Kaabi et al., 2016; Alikas & Kratzer, 2017). The freshwater environments provide support for diverse ecosystems and habitats, however they suffer with anthropogenic interference with waste discharge and runoff which drives to changes of the water quality characterized as algal proliferation, eutrophication and increasing of turbidity and reducing of water transparency (Wetzel, 2001; Calijuri et al., 2002; Lira et al., 2009). The water transparency is affected by the optical significant substances (OSS) such as phytoplankton, chromophoric dissolved organic matter (CDOM), non-algal particles (NAP) and the pure water that absorb and/or scatter the downwelling irradiance which diminish with the depth. The parameters used to quantify the water transparency are the attenuation coefficient [kd(PAR)] of the photosynthetically active radiation (PAR), the euphotic zone depth (Zeu) and the Secchi disk depth (ZSD). The kd(PAR) is the exponential decrease of downwelling irradiance with the depth and it is related to Zeu. The Zeu is defined as the depth where the downwelling irradiance of the PAR [Ed(PAR)] achieved 1% of the Ed(PAR) measured at the surface water; assuming that the water column is homogenous, kd(PAR) and Zeu are related through the equation (Kirk, 1994). In the euphotic zone, there is sufficient light intensity for significant photosynthesis, therefore, the Zeu had been use to provide information of primary production in water bodies and the Zeu changes can depict environmental patterns and anthropogenic impacts (Kirk, 1994; Majozi et al., 2014; Yang et al., 2015; Ma et al., 2016). The most common and easy way to quantify the water transparency is the ZSD, taken when the black-and-white disk, lowered in water, is no longer viewable by an observer; because the simple and universal method, the ZSD is routinely used for turbidity monitoring, water transparency, estimate the order of magnitude of optical 40 substances in water and indicate the eutrophic state in water bodies (Binding et al., 2007; Lee et al., 2015; Shang et al., 2016; Alikas & Kratzer, 2017). The ZSD is able to provide in first-hand the water transparency whereas Zeu measures the water clarity more rigorously and generate more reliable results (Lee et al., 2007; Majozi et al., 2014). In order to estimate the water clarity using remote sensing data, some empirical and semi-analytical models were developed. Empirical models explore the relation between the diffuse attenuation coefficient at 490 nm [kd(490)] and kd(PAR) to estimate Zeu in coastal waters and turbid lake waters (Zhao et al., 2013; Zhang et al., 2012; Wang et al., 2009). Zhao et al. (2013) used MODIS/Acqua/Terra and SeaWiFS satellite-data to understand the light environment in SW Florida coastal waters. Zhang et al. (2012) calibrated a simple model to estimate kd and, further, Zeu in a shallow, turbid Taihu Lake, China using MODIS data and Wang et al. (2009) proposed to use a combination of kd(490) models to improve the accuracy of estimation of kd(490) and kd(PAR) products for both turbid and clear waters using MODIS data in Chesapeake Bay. Lee et al. (2007), aiming an approach that avoided parameterizations regarding the wide seasonal and regional bio-optical variations, developed a semi-analytical Zeu model based on kd(PAR) and the inherent optical properties (IOPs) derived from the quasi-analytical algorithm (QAA; Lee et al., 2002) using SeaWiFS and MODIS satellite data in ocean and coastal waters. Still using the QAA scheme and a mechanistic model to overcome the empirical limitations, Lee et al. (2015) proposed a semi-analytical equation to estimate ZSD as an inverse relation of kd, derived from the IOPs calculated via QAA, at the transparent window of the water body within the visible domain. This mechanist model of ZSD is a new approach of the underwater visibility theory (Preisendorfer, 1986) that aimed to interpret the exactly sighting of Secchi disk in water by the human eye. The ZSD model achieved good performance (R² = 0.96 and absolute percentage difference of 16.7%) for a wide range of water clarities (ZSD 0.1-30 m) in configuration of OLI/Landsat-8 visible bands (Lee et al., 2016). From our understanding there are no investigation of water clarity through the Zeu and ZSD parameters using satellite data in inland waters dominated by CDOM, which is an OSS of dissolved organic matter and it is considered an important water quality indicator which affect the primary productivity due to the interference in penetration of PAR into water column of water bodies, reducing the accuracy of water transparency estimation (Zhang et al., 2009). The main goals of this paper were to assess the 41 performance of models to estimate Zeu and the new ZSD model developed by Lee et al. (2015) in CDOM dominant inland waters; investigate how the estimations are affected by this optical water type and comprehend the influences of the seasonality in rainfall terms on Zeu and ZSD. The intermediary product to obtain Zeu and ZSD through the available models, kd(490), was derived from the semi-analytical equation of Lee et al. (2013) for rich-CDOM inland waters through the QAA using the OLI/Landsat-8 with mean absolute percentage error of 41%. Six Zeu algorithms were tested and the temporal and spatial distributions of the euphotic zone and Secchi disk depths were evaluated. 3.2. Material and Methods 3.2.1. Study Site The Bariri hydroelectric reservoir (22° 9‟ 49.260” S 48° 44‟ 21.420” W) is in the middle of the São Paulo State, in Tietê River, being part of the Cascading Reservoir System (CRS) as the second one of a total of six reservoirs (Figure 3.1). The Bariri reservoir has been in operation since 1969, presenting the smallest flooded area of 63 km² in an average altitude of 450 meters and has serviced the water supply for human use, irrigation and recreation. By the reduced dimension of the reservoir, the operation type is water line with retention time variation from 7 to 24 days (Barbosa et al., 1999). The reservoir is situated in a tropical climate with a dry period (April-September) and a wet period (October-March) according to Köppen classification. The Tietê River traverses the metropolitan region of São Paulo State bringing an amount of wastewater in the downstream direction. The Cascading Reservoir Continuum Concept (CRCC) allows the gradual decrease of nutrients and pollutants along the reservoirs and location and retention time of Bariri reservoir promotes the heterogeneity of the eutrophication levels in upstream to downstream, revealed by the water transparency of the water. The Bariri reservoir is characterized as highly productive water with high average concentrations of total nitrogen (2750 μg L -1 ), phosphorus (87 μg L -1 ) and Chl-a (55.8 μg L -1 ) with the phytoplankton community dominated by cyanobacteria (Barbosa et al., 1999). 42 Figure 3.1. Maps showing the location of (a) São Paulo State in Brazilian territory and the path/row of OLI/L8 images which the coordinates of Bariri reservoir are contained, highlighted by the red square; (b) CRS of Tietê River in São Paulo State with the six reservoirs and their respective dams (1 – Barra Bonita; 2 – Bariri; 3 – Ibitinga; 4 – Promissão; 5 – Nova Avanhadava; 6 – Três Irmãos) and (c) Bariri reservoir boundaries with the sampling points of the two field campaigns. 3.2.2. Planning of Sampling and Fieldworks The establishment of the samples was based on a random stratified sampling method using an annual cycle (2013) of OLI/L8 images acquired at the USGS website (www.earthexplorer.usgs.gov). In order to analyze the natural and anthropic variations along the reservoir, the same band of each radiometrically calibrated image (each one with 6 spectral bands) were compressed and submitted to calculation of the mean and, sequentially, of the standard deviation (SD). The SD images were used in a Principal Component Analysis (PCA) for selection of the image with the highest variability, which was sliced for the random stratified sampling (Rodrigues et al., 2016). For Bariri reservoir, 30 sampling points were established ensuring the minimal distance of 1 km between them to avoid clusters. Two fieldworks were carried out in the dry period, being the first one (BAR1) realized from 15 to 18 August 2016 with the 30 43 sampling points considered and the second one (BAR2), realized from 23 to 24 June 2017 with 12 samples points less than BAR1, maintaining the points which presented high values of OSS concentrations and avoiding the clusters to ensure the spatial variability (see Figure 3.1 for sampling locations). The days of the field campaigns were determined according to the temporal resolution of OLI/Landsat-8 aiming to match the data collection with the satellite overpass the study site. The months of fieldworks were chosen to avoid rainy periods and facilitate the field data collection. 3.2.3. Field Data Collection During the BAR1 and BAR2, water samples for laboratory analyses, water quality parameters and optical data were collected. The water quality parameters such as depth (m), turbidity (NTU) and Secchi disk depth (ZSD; m) were sampled in both fieldworks. The wind speed (m -1 ) was measured using a portable anemometer. The water samples were collected in the field and stored refrigerated in polyethylene bottles to derive the OSSs and IOPs. The OSSs such as the Chl-a and the suspended particulate matter (SPM) as well as the inorganic (IPM) and organic particulate matter (OPM) concentrations were obtained from 0.2-0.5L of filtered water and replica in each sample point according to Golterman et al. (1978) and the American Public Health Association Protocol (APHA, 1998), respectively. The laboratorial analyses of the IOPs were made from 0.1L of filtered water and replica in each sampling point following the methodology proposed in Bricaud et al. (1981) for the absorption coefficient of the CDOM (aCDOM) while for the absorption coefficients of the non-algal particle (aNAP) and the total particulate (ap), that consists in the sum of aNAP and phytoplankton absorption coefficient (aϕ). The acquisition and analyses were made in accordance with the Transmittance-Reflectance method according to Tassan and Ferrari (1995, 1998, 2002). The radiometric data were obtained from the two hyperspectral RAMSES sensors, the ARC type with an angle-of-view of 7º used for the radiance measurements and the ACC type with a cosine collector for the irradiance measurements (TriOS, Rastede, Germany). The sky and total radiance data (Lsky and Lt, respectively, both in W m -2 sr -1 ) and the downwelling and sky irradiance data (Ed and Esky, respectively, both in W m -2 ) were taken with the sensors fixed by a steel frame in the boat, in a configuration of 40° from nadir (zenith) and to azimuthal angle of 90° in order to minimize the specular 44 reflection (Mobley, 1999). The radiometric quantities were used to calculate the remote sensing reflectance according to Equation (3.1). = (3.1) where + is the downwelling irradiance measured at the water surface. The LSR is the surface-reflected radiance and consists in a multiplicative product of Lsky and ρ - an effective surface reflectance. The ρ value ( 0.02-0.05) depends on the viewing geometry and spectral variation. In agreement with that, Lee et al. (2010) proposed a calculation approach of Rrs as a function of the total remote-sensing reflectance (Trs, ratio of Lt to Ed) and sky remote-sensing reflectance (Srs, ratio of Lsky to Ed) for each Lt and Lsky scan from the Equation (3.2): (3.2) where F refers to Fresnel reflectance and it was set as 0.021 according to the viewing geometry; is a spectrally constant settled before can be derived (Lee et al., 2010). In oceanic waters, the is negligible in the red and near-infrared wavelengths and can be assumed as zero beyond 700 nm, however, for turbid inland waters, the IOPs have a significant influence in , thus, one alternative is to model the spectral based on the IOPs, and then resolve Δ after comparing modeled and derived from Equation (3.2). In sequence, the Rrs used in estimates model of kd and, in sequence, estimates models of ZSD and Zeu, must be simulated to the satellite signal at each spectral channel centered at a wavelength (λ = 443; 482; 561; 655 nm). The band simulation consists in the convolution of the radiation signal of TriOS RAMSES hyperspectral sensor and the spectral response function [Fr(λ)] of the OLI/L8 sensor in wavelength interval of the spectral resolution (Barsi et al., 2014): ∫ ∫ (3.3) where is the remote-sensing reflectance simulated at center wavelength; xmax and xmin are, respectively, the maximum and minimum values of the sensor spectral channel. 45 According to Ed data, the kd(490) in both field campaigns was determined as the slope Ed data simulated to satellite signal at spectral channel centered at 482 nm at subsurface depth (z) (Mobley, 1994): (3.4) In order to eliminate the noise in the due to changes in the sun illumination condition caused by cloud cover during the measurements, a normalization factor was required in all scans. The Ed normalization consists in a division of Esky at first scan t(z1) to Esky at subsequent scans t(zi) as factor normalization of the t(zi) according Mishra et al. (2005) and Mueller (2000): (3.5) where is the normalized The Ed data was still used for Zeu determination in both field campaigns by summing the hyperspectral from 350 nm to 700 nm, obtaining vertical profiles of for each sampling point according to Equation (3.6). ∫ (3.6) In sequence, the Zeu was derived at the depth where the achieves 1% of the available at subsurface water or, shortly, as (Lee et al., 2005, 2007). 3.2.4. Euphotic Zone Depth Models The Table 3.1 summarizes the six Zeu models, which cover the semi-analytical and empirical ones. 46 Table 3.1. Zeu models developed from coastal and ocean waters, slightly turbid ocean waters and turbid lake waters data. The semi-analytical model developed by Lee et al. (2007) (Zeu_Lee) is a cubic- polynomial equation with the constants based on numerical simulations and two parameters, k1 and k2, estimated from the IOPs (a(490) and bb(490)) and sun angle (θs = 40°). The parameters, k1 and k2, are related with kd(PAR)(z), the diffuse attenuation of Ed(PAR)(z) in the visible domain (400-700 nm). Considering that Zeu is the layer within which the Ed(PAR) falls 1% of the surface value and kd(PAR) is approximately constant with the depth (Morel, 1988) the Zeu was related through kd(PAR) to k1 and k2 from the Equation (3.7). (3.7) The model generates three solutions, one negative and two positive, but the smallest positive value is consistent with the radiative transfer theory. The IOPs were obtained from the QAA_v5 proposed by Lee et al. (2009) (the version 5 was used in this study) with in situ as input, simulated at the satellite center channel of reference wavelength ( of 561 nm, the nearest OLI center band of = 55x of QAA_v5. Zhao et al. (2013) aiming to improve the Zeu estimations for the calibration set of coastal waters developed an empirical hyperbolic function based on the relation between in situ measures of Zeu and kd(490), achieving better results for their study site than Zeu_Lee application. Calibration Dataset Coastal and Ocean Waters Coastal Waters Turbid Coastal Waters Turbid Lake Waters Wang et al. (2009) Zhang et al. (2012) Equations Model Lee et al. (2007) Zhao et al. (2013) (I) (II) (III) 47 The Zeu values from the empirical models were derived from the correlation between kd(PAR) and kd(490) for Zeu estimations in Wang et al. (2009) (Zeu_Wang) and in the two first equations of Zhang et al. (2012) (Zeu_Zhang_I and Zeu_Zhang_II, respectively). The equation III of Zhang et al. (2012) (Zeu_Zhang_III) was calibrated with MODIS data with a channel at the near infra-red spectral region of 748 nm and considered the ratio of Rrs( ) and solar zenith angle cosine (μ0), calculated according to the sampling time, latitude and solar declination. Taking into account the attenuation coefficient dependence on the sun angle and the light angular distribution, the ratio Rrs( )/μ0 reduces the effect of sun angle on kd(PAR) estimations. From the kd(PAR) obtained in Zeu_Wang, Zeu_Zhang_I-III, the Zeu was estimated using the Equation (3.7). The relation used to obtain Zeu from kd(PAR) has been commonly applied in water quality studies (Ma et al., 2016; Liu et al., 2016; Majozi et al., 2014). 3.2.5. Secchi Disk Depth Model The ZSD model analyzed in this study was developed by Lee et al. (2015) based on the inversion relation between ZSD and kd. The ZSD estimations require the application of a sequential scheme with three steps, initializing with the IOPs (a(λ) and bb(λ)) estimations from QAA_v5 (λ0 = 561 nm); in sequence, as second step, the kd(561) calculation according to the Lee et al. (2013) from the IOPs previously estimated and, finally, as third step, the ZSD calculation from the Equation (3.8). ( | | ) (3.8) where at 561 nm was the minimal value of attenuation of Ed in the visible domain (443-665 nm) as recommended in Lee et al. (2015). The light attenuation in turbid waters is higher in the blue and red spectral regions because the selective absorption (Cairo et al., 2017; Kirk, 1994). In accordance with that, the maximum transmission of light happens at the green spectral region, called the transparency window in the visible domain, justifying the choice of Rrs(561) among the OLI center wavelengths. 48 3.2.6. OLI/Landsat-8 Data Acquisition and Processing The Landsat 8 Surface Reflectance on-demand data generated by the Landsat Surface Reflectance Code (LaSRC) were obtained in the U.S. Geological Survey platform (http://earthexplorer.usgs.gov/). The LaSRC algorithm for atmospheric correction was developed using the Second Simulation of the Satellite Signal in the Solar Spectrum Vectorial (6SV) model. The algorithm uses the OLI Coastal Aerosol Band (0.433–0.450 μm) which works as cover for shorter wavelength as the blue band in previous Landsat and helpful to retrieve aerosol properties (Vermote et al., 2016). Pahlevan et al. (2017) and Bernardo et al. (2016) showed that LaSRC is a consistent product to derive aquatic estimates. A total of 31 images atmospherically corrected images (Path/Row 221/75 or 220/76) were acquired during the years of 2014, 2015 and 2016. The criteria used for the images choice was attributed to cloud free data over the entire reservoir (images with some parts covered were selected). In 2014, the twelve images were selected; in 2015, nine images were selected with March, November and December images missing; in 2016, ten images were selected with January and December images missing. In order to map the Zeu and ZSD models, surface reflectance images were divided by π to convert them into Rrs and rescaled by a factor scale of 0.0001 to convert to the 0 to 1 range (Moses et al., 2012; USGS, 2018). The Zeu model with the best performance and the ZSD model from Lee et al. (2015) were applied to the time-series of LaSRC images selected to obtain the spatial distribution of depths of the euphotic zone and the Secchi disk. The Figure 3.2 displayed the flowchart of the of Zeu and ZSD spatialization. Figure 3.2. Flowchart showing the sequential steps followed to mapping ZSD and Zeu. 49 The Rrs of first four bands of LaSRC time-series images were used in the equations in original configuration of QAA_v5 (Lee et al., 2009) in order to obtain the a(λ) and bb(λ) images. The a(490) and bb(490) and a(561) and bb(561) images were used to generated the kd(490) and kd(561) images, respectively, from the semi-analytical model proposed by Lee et al. (2013). These two steps enable the mapping of ZSD using the semi-analytical equation by Lee et al. (2015) and the mapping of Zeu from any model tested in this study. The ZSD model and Zeu model best fitted for Bariri reservoir were validated through the satellite image from August 15 th 2016, correspondent to the first day of BAR1. In the same way, the LaSRC accuracy in Zeu and ZSD estimations were evaluated from the analysis between the Rrs converted images with the Rrs calculated from TriOS data in BAR1. 3.2.7. Statistical Analysis and Accuracy Assessment Statistical analyses, including calculations of the maximum, minimum and average values and linear and non-linear regressions were performed. The Zeu and ZSD analyzed models were applied for BAR1 and BAR2 datasets. The Root Mean Square Error (RMSE), Unbiased Absolute Percentage Error (ε) and the bias were used to assess the accuracy of the models. √ ∑ ( ) (3.9) ∑ | | (3.10) ∑ ( ) (3.11) where and the ZSD and Zeu estimated values and is the in situ measures of ZSD and Zeu. 3.3. Results 3.3.1. In situ Measurements The water color in BAR1 was green in most of the sampling points whereas in BAR 2, the water was brownish. The sky during the field campaigns was in most 50 favorable with sunny days with the exception of some period of overcast in BAR1. The wind speed (m s -1 ) ranged from 0 to 8 among the field campaigns with the maximum value in BAR1 that presented average of 3.47 1.80 m s -1 and 3.10 2.00 m s -1 . The in situ Rrs spectral curves are in Figure 3.3. The spectral feature of BAR1 was predominant by Chl-a pigment influence with a significant absorption at 680 nm and two peaks of reflectance at 550-570 nm and 710 nm, corresponding to typical waters with high phytoplankton concentrations. At the green spectral region, the pigments assume the minimal absorption and all the particulate matter play the major role in reflectance. The sampling points with the highest values of Chl-a concentration ([Chl- a]) presented a peak of absorption at 440 nm. The dip verified at 625 nm is associated with the phycocyanin pigment that is present mainly in cyanobacteria (Majozi et al., 2014; Gitelson et al., 2007). The Rrs magnitude in BAR2 was lower than BAR1 due to the OSS concentrations reduction (see Table 2). The CDOM dominated the light absorption process in both field campaigns (67.74% 9.98% in BAR1 and 88.85% 5.70% in BAR2 at 490 nm) and it was verified in BAR1 with the lower Rrs at 400- 500 nm while in BAR2, the Rrs curves assumed a flat feature at 600 nm, resulted of a CDOM-rich water with other substances influence (Binding et al., 2008; Brezonik et al., 2015). Figure 3.3. The Rrs spectra for BAR1 (a) and BAR2 (b) surveys. T