PRESIDENTE PRUDENTE 2018 UNIVERSIDADE ESTADUAL PAULISTA “JÚLIO DE MESQUITA FILHO” CAMPUS PRESIDENTE PRUDENTE FACULDADE DE CIÊNCIAS E TECNOLOGIA Programa de Pós-Graduação em Ciências Cartográficas CAROLINE PIFFER DE ANDRADE P REMOTE SENSING OF CHLOROPHYLL-A CONCENTRATION BASED ON ABSORPTION COEFFICIENTS IN IBITINGA RESERVOIR CAROLINE PIFFER DE ANDRADE REMOTE SENSING OF CHLOROPHYLL-A CONCENTRATION BASED ON ABSORPTION COEFFICIENTS IN IBITINGA RESERVOIR A dissertation submitted to the Faculty of Science and Technology of São Paulo State University in partial fulfillment of the requirements for the degree of Master of Cartographic Sciences. Advisor: Prof. Dr. Enner Alcântara Co-advisor: Dr. Milton Kampel PRESIDENTE PRUDENTE 2018 Ficha catalográfica elaborada pela Seção Técnica de Aquisição e Tratamento da Informação - Diretoria Técnica de Biblioteca e Documentação - UNESP, Campus de Presidente Prudente Piffer de Andrade, Caroline. A566r Remote sensing of chlorophyll-a concentration based on absorption coefficients in Ibitinga reservoir / Caroline Piffer de Andrade. - 2018 65 f. : il. Orientador: Enner Herenio de Alcântara Dissertação (mestrado) - Universidade Estadual Paulista. Faculdade de Ciências e Tecnologia, Presidente Prudente, 2018 Inclui bibliografia 1. Qualidade de águas interiores. 2. Algoritmo quase-analítico. 3. Propriedades óticas inerentes. I. Herenio de Alcântara, Enner. II. Universidade Estadual Paulista. Faculdade de Ciências e Tecnologia. III. Título. Alessandra Kuba Oshiro Assunção CRB-8/9013 To my parents, Alessandra e Paulo AGRADECIMENTOS Agradeço a todos que, de alguma forma, colaboraram com este trabalho. Meus agradecimentos especiais: À Força Divina, que tem me guiado ao longo de todas as etapas de minha vida. Aos meus pais, Alessandra e Paulo, por todo apoio, incentivo, amor e pelos sacrifícios que sempre fizeram para que eu pudesse seguir os meus sonhos. Ao meu padrasto e madrasta, Fábio e Renata, pelo incentivo e amizade. Aos meus irmãos, João Pedro e Tiago, pela amizade de sempre. Aos meus avós, Maria Aparecida, Darcy e Líbera, que mesmo diante de situações desfavoráveis sempre nos ensinaram que a educação é prioridade. Ao meu companheiro da vida, Henrique, por todo amor, apoio e por ser o meu maior incentivador. Ao meu orientador, Prof. Enner Alcântara, agradeço pelos ensinamentos e por sempre exigir o meu melhor. Muito obrigada por todo o apoio e principalmente pela confiança. Ao meu co-orientador, Dr. Milton Kampel, por toda ajuda e confiança neste trabalho. Aos professores do Programa de Pós-Graduação em Ciências Cartográficas, pelo conhecimento compartilhado. Aos amigos e colegas do Programa de Pós-Graduação, especialmente: Carol Campos, Bruno Faga, Carol Ambrósio, Sarah Martins, Nariane Bernando, Alisson Carmo, Fernanda Watanabe, Luiz Rotta e Thanan Guimarães. Aos membros da banca, por aceitarem avaliar e contribuir com este trabalho. À FCT-UNESP pela infraestrutura que possibilitou a realização desse trabalho. Aos funcionários da UNESP, em especial da Seção de Pós-Graduação e à Cidinha e Zilda, por todas as palavras de carinho e incentivo. Ao Prof. Edivaldo Velini (FCA/UNESP) e funcionários do Laboratório de Matologia, pela disponibilização do laboratório. À FAPESP (Fundação de Amparo à Pesquisa do Estado de São Paulo) pelo financiamento de projetos (Processos N° 2012/19821-1 e 2015/21586-9). À CAPES/MEC (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior – Ministério da Educação) pela bolsa de mestrado. A todos, meu muito obrigado! “You cannot get through a single day without having an impact on the world around you. What you do makes a difference, and you have to decide what kind of difference you want to make.” Jane Goodall RESUMO O presente estudo objetivou estimar as concentrações de clorofila-a (Chl-a) no reservatório da usina hidroelétrica de Ibitinga (RHI), localizado no Rio Tietê, estado de São Paulo, Brasil, por meio de coeficientes de absorção obtidos via algoritmos quase-analíticos (QAAs). Para isso, realizou-se uma caracterização bio-ótica e biogeoquímica do RHI, por meio de dados espectrais e de qualidade da água coletados em dois trabalhos de campo, conduzidos em Julho de 2016 e Junho de 2017. Os desempenhos das versões originais QAAV5 e QAAV6 em estimar as propriedades óticas inerentes (POIs) no RHI foram avaliados. Versões re- parametrizadas para dois reservatórios localizados no sistema em cascata do Rio Tietê, QAABBHR e QAAOMW, foram também testadas para a área de estudo. Além disso, foram avaliadas as performances de esquemas compostos pelas versões do QAA já mencionadas, seguidas por quatro modelos para estimativa de Chl-a, os quais utilizam coeficientes de absorção como dados de entrada. A distribuição espacial das concentrações de Chl-a foi analisada por meio da aplicação desses esquemas em uma imagem do sensor Ocean and Land Colour Instrument (OLCI) instalado a bordo do satélite Sentinel-3A, com aquisição coincidente com o segundo trabalho de campo realizado na área de estudo. A caracterização bio-ótica demonstrou variabilidade espacial e temporal dos constituintes oticamente significativos (COSs) no RHI, com predominância da absorção pelo material orgânico colorido dissolvido (CDOM). As versões do QAA testadas para o primeiro conjunto de dados não se mostraram completamente adequadas na obtenção de coeficientes de absorção em todos os comprimentos de onda. Com relação aos esquemas para estimativa de concentração de Chl-a, apenas aqueles baseados no QAAV5 foram capazes de obter resultados razoáveis - Raiz do Erro Médio Quadrático Normalizado (REMQN) < 47.50 % - para os dados da imagem OLCI. Todos os quatro modelos para estimativa de Chl-a testados apresentaram resultados similares para os dados de saída do QAAV5. Esses resultados enfatizam o desafio gerado pela grande variabilidade ótica dos sistemas em cascata, com relação à modelagem bio-óptica. Os resultados obtidos dão suporte a futuros trabalhos, os quais podem resultar em aplicações como o monitoramento do estado trófico na área de estudo a partir de dados de satélite, com maior acurácia proveniente do uso de modelos que possam estimar consistentemente suas POIs. Palavras-chave: Qualidade de águas interiores, Algoritmo Quase-analítico, Propriedades Óticas Inerentes, Fitoplâncton, OLCI/Sentinel-3A. ABSTRACT This research was aimed at retrieving chlorophyll-a (Chl-a) concentrations in Ibitinga Hydroelectric Reservoir (IHR), located at Tietê River, São Paulo State, Brazil, using absorption coefficients obtained via Quasi-analytical algorithms (QAAs). For this purpose, a bio-optical and bio-geochemical characterization of IHR was carried out, through spectral and water quality data collected in two field campaigns conducted in July, 2016 and June, 2017. The suitability of two QAA native forms (QAAV5 and QAAV6) in retrieving inherent optical properties (IOPs) in IHR was assessed. Versions re-parameterized for two reservoirs also located in the Tietê River cascading system, QAABBHR and QAAOMW, were also tested for the study area. Besides that, the performances of schemes composed by the QAA versions already mentioned followed by four models that use absorption coefficients as inputs for estimating Chl-a concentration in Ibitinga Reservoir were evaluated. Spatial distribution of Chl-a in the reservoir was analyzed, since these schemes were applied in an image of the Ocean and Land Colour Instrument (OLCI) sensor onboard Sentinel-3A satellite, with acquisition date coincident with the second field campaign. The bio-optical characterization showed spatial and temporal variability of optically significant constituent (OSC) in IHR and colored dissolved organic matter (CDOM) predominance in its absorption budget. None of the QAA versions tested for the first dataset was completely satisfactory in retrieving absorption coefficients for IHR in all wavelengths. Regarding the schemes for Chl- a concentration estimates, only the ones based on QAAV5 were able to obtain reasonable results - Normalized Root Mean Square Error (nRMSE) < 47.50 % - for the OLCI image data. All four models for Chl-a estimation tested presented similar results for QAAV5 outputs. These results highlight the challenge of copying with high optical variability in cascading systems. The results obtained support further works, which can, prospectively, lead to many practical applications, as monitoring of trophic state in the study area from satellite data, with higher accuracy provided by the use of models that can consistently retrieve the IOPs for this specific water system. Keywords: Inland water quality, Quasi-Analytical Algorithm, Inherent Optical Properties, Phytoplankton, OLCI/Sentinel-3A. LIST OF FIGURES Figure 1. (a) São Paulo State location in Brazil; (b) Cascading reservoirs in Tietê River. Please note Ibitinga Hydroelectric Reservoir (black arrow) upstream from Nova Avanhandava reservoir and downstream from Barra Bonita reservoir (see text for more details); (c) Distribution of sampling stations from IBI1 (July, 2016) and IBI2 (June, 2017) field campaigns in the study area………………..............................................................................21 Figure 2. Area-averaged of precipitation rate from June, 2016 to June, 2017. Source: NASA/GIOVANNI (https://giovanni.gsfc.nasa.gov/giovanni/)..............................................22 Figure 3. (a) Acquisition geometry proposed by Mobley (1999). Adapted from Rodrigues (2017); (b) Radiometric sensors arrangement in field works………………………………...25 Figure 4. (a) In situ Rrs spectra; (b) Average absorption coefficients by phytoplankton (aϕ), CDOM (aCDOM), detritus (ad) and pure water (aw); n=29..........................................................29 Figure 5. Ternary plots presenting the relative contribution of detritus, phytoplankton and CDOM to the total absorption (without water fraction) at three different wavelengths for IHR (red dots), BBHR (black dots) and NHR (brown dots): (a) 443 nm; (b) 560 nm; (c) 665nm………………………………………………………………………………………....30 Figure 6. Comparison between estimated and measured aϕ using: (a) QAAV5; (b) QAAV6; (c) QAABBHR; and (d) QAAOMW.....................................................................................................32 Figure 7. Comparison between estimated and measured aCDM using: (a) QAAV5; (b) QAAV6; (c) QAABBHR; and (d) QAAOMW...............................................................................................34 Figure 8. Flowcharts of (a) calibration process of sixteen Chl-a retrieval models for the study area; (b) image pre-processing and application of schemes for Chl-a concentration retrieval.....................................................................................................................................44 Figure 9. Average phytoplankton (aϕ), CDOM (aCDOM), detritus (ad) and pure water (aw) absorption coefficients in Ibitinga reservoir in (a) IBI1 (July, 2016; n=29); (b) IBI2 (June, 2017; n=6). ...............................................................................................................................47 Figure 10. Rrs spectra obtained from in situ measurements in: (a) IBI1 field campaign (n=29); (b) IBI2 field campaign (n=6)…................................................................................47 Figure 11. Rrs spectra from IBI2 (June, 2017, n=6) simulated for OLCI first twelve bands.........................................................................................................................................48 Figure 12. (a) Errors between OLCI-derived Rrs spectra after Empirical Line (EL) atmospheric correction versus Rrs spectra from field campaign (IBI2) measurements resampled for OLCI bands (average for all sampling stations). (b) Comparison of OLCI- derived Rrs spectrum after atmospheric correction and Rrs spectrum from IBI2 measurements resampled for OLCI bands at sampling station 6. First twelve OLCI bands are shown in both (a) and (b)...................................................................................................49 Figure 13. Comparison per band of in situ Rrs and OLCI-derived Rrs after Empirical Line atmospheric correction; six satellite-in situ matchups are considered......................................50 Figure 14. Models configured from (a) γ1 (mγ1); (b) γ2 (mγ2); (c) γ3 (mγ3); and (d) γNDCI (mγNDCI). Indexes used QAAV5 outputs...................................................................................51 Figure 15. Spatial distribution of Chl-a concentrations in Ibitinga Reservoir using OLCI image acquired on June 21 st , 2017. Chl-a retrieved via (a) QAAV5 plus mγ1 scheme; (b) QAAV5 plus mγ2 scheme; (c) QAAV5 plus mγ3 scheme; (d) QAAV5 plus mγNDCI scheme......................................................................................................................................52 LIST OF TABLES Table 1. Descriptive statistics of water quality parameters in IHR; SD: standard deviation; CV: coefficient of variation; n = 29..........................................................................................25 Table 2. QAAs performances for aϕ retrieval, according to each band, based on NRMSE (%) and MAPE (m -1 )........................................................................................................................28 Table 3. QAAs performances for aCDM retrieval, according to each band, based on NRMSE (%) and MAPE (m -1 ).................................................................................................................30 Table 4. Descriptive statistics of water quality data from IBI1 and IBI2 field campaigns. Min: minimum; Max: maximum; SD: standard deviation; CV: coefficient of variation; n = 29 for IBI; n = 6 for IBI2...............................................................................................................43 Table 5. Descriptive statistics of absorption coefficients from IBI1 and IBI2 field campaigns. Min: minimum; Max: maximum; SD: standard deviation; CV: coefficient of variation; n = 29 for IBI; n = 6 for IBI2…………………………………...............................44 Table 6. Errors between fits from cross-validated models and Chl-a concentrations from IBI1……………………………………………………………………..................................46 Table 7. Models performances for indexes estimated via QAAV5 applied in the image………………………………………………………………………….........................48 LIST OF ABBREVIATIONS AND ACRONYMS BBHR Barra Bonita Hydroelectric Reservoir CDOM Colored Dissolved Organic Matter CETESB Companhia Ambiental do Estado de São Paulo Chl-a Chlorophyll-a CRCC Cascading Reservoir Continuum Concept C2RCC Case-2 Regional/Coast Colour CV Coefficient of variation EL Empirical Line ESA European Space Agency IHR Ibitinga Hydroelectric Reservoir IOPs Inherent optical properties ISM Inorganic suspended matter MAPE Mean absolute percentage error NDCI Normalized Difference Chlorophyll Index NHR Nova Avanhandava Hydroelectric Reservoir NIR Near infrared nRMSE Normalized root mean square error NTU Nephelometric Turbidity Units OD Optical density OLCI Ocean Land Colour Instrument OSC Optically Significant Constituents OSM Organic suspended matter QAA Quasi-analytical algorithm S3 Sentinel-3A SD Standard deviation SPM Suspended particulate matter TOA Top of atmosphere LIST OF SYMBOLS at Total absorption coefficient aCDOM Colored dissolved organic matter absorption coefficient aCDM Colored dissolved organic matter plus detritus absorption coefficient ad Detritus absorption coefficient aϕ Absorption coefficient of phytoplankton pigments ap Particulate absorption coefficient aw Pure water absorption coefficient bbp Backscattering coefficient of suspended particles bbw Backscattering coefficient of pure water Ed (λ) Downwelling irradiance F(θ) Surface Fresnel reflectance Lr (λ) Surface-reflected radiance Ls (λ) Atmospheric diffuse radiance Lt (λ) Total upwelling radiance R Irradiance reflectance Rrs Remote sensing reflectance Srs Sky remote sensing reflectance Trs Total remote sensing reflectance ε Absolute percentage difference θ Zenith angle ϕ Azimuth angle λ Wavelength λ0 Reference wavelength CONTENTS CHAPTER 1: INTRODUCTION ................................................................................................. 14 1.1. Contextualization ..................................................................................................... 14 1.2. Hypothesis ................................................................................................................. 17 1.3. Objectives ................................................................................................................. 17 1.3.1. Specific objectives............................................................................................... 18 1.4. Dissertation Structure ............................................................................................ 18 CHAPTER 2: STUDY AREA........................................................................................................ 20 2.1. Characterization .......................................................................................................... 20 2.2. Field Campaigns .......................................................................................................... 21 CHAPTER 3: POTENTIALS AND LIMITATIONS OF A QUASI-ANALYTICAL ALGORITHM FOR RETRIEVING THE WATER ABSORPTION PROPERTIES IN A CASCADING RESERVOIR SYSTEM ...................................................................................... 23 3.1. Introduction .................................................................................................................. 23 3.2. Material and Methods ................................................................................................. 25 3.2.1. Water quality data ................................................................................................... 25 3.2.2. Radiometric Data .................................................................................................... 25 3.2.3. Phytoplankton absorption coefficient (aϕ) and Detritus plus CDOM absorption coefficient (aCDM) ............................................................................................................... 26 3.2.4. QAA ......................................................................................................................... 26 3.3. Results and Discussion ................................................................................................. 28 3.3.1. Water quality characterization................................................................................ 28 3.3.2. Bio-Optical Characterization .................................................................................. 29 3.3.3. QAA performances .................................................................................................. 30 3.3.4. Implications for OSC monitoring ............................................................................ 34 3.4. Conclusion .................................................................................................................... 36 CHAPTER 4: QUASI-ANALYTICAL ALGORITHM BASED SCHEMES FOR CHLOROPHYLL-A RETRIEVAL IN A TROPICAL RESERVOIR VIA OLCI – SENTINEL-3A IMAGE .................................................................................................................. 37 4.1. Introduction .................................................................................................................. 37 4.2. Methods ......................................................................................................................... 39 4.2.1. Water quality data ................................................................................................... 39 4.2.2. Absorption coefficients ............................................................................................ 39 4.2.3. Radiometric Data .................................................................................................... 40 4.2.4. QAA context............................................................................................................. 41 4.2.5. Calibration and validation of absorption coefficients based models ...................... 42 4.2.6. Chl-a retrieval via OLCI image .............................................................................. 43 4.2.7. Accuracy assessment ............................................................................................... 44 4.3. Results ........................................................................................................................... 45 4.3.1. Water quality characterization................................................................................ 45 4.3.2. Bio-optical characterization ................................................................................... 46 4.3.3. Validation of Chl-a concentration retrieval schemes .............................................. 48 4.3.4. Atmospheric Correction Assessment ....................................................................... 49 4.3.5. Assessment of Chl-a Retrieval via OLCI image ...................................................... 50 4.4. Discussion ..................................................................................................................... 53 4.5. Conclusions ................................................................................................................... 57 CHAPTER 5: FINAL CONSIDERATIONS AND RECOMMENDATIONS ................... 58 REFERENCES ................................................................................................................................. 60 14 CHAPTER 1: INTRODUCTION 1.1.Contextualization Inland waters play a crucial role in hydrological, carbon and nutrient cycles, offer numerous ecosystem services and provide resources for multiple uses (AYRES et al., 1996; MOSS, 2012). Threatens originating from human interventions, as nutrient enrichment and other types of pollution, modifications due to land-use and climate change effects, lead to degradation of inland waters quality resulting in several environmental, sanitary and economic consequences (BRÖNMARK and HANSSON, 2002; CARPENTER et al., 2011). To support the assessment of water quality, as well as to assist its improvement, frequent, long-term sustained and spatially distributed data collection and monitoring are required (DÖRNHÖFER and OPPELT, 2016). Hydroelectric reservoirs are, in special, aquatic systems of notable economic relevance, which also supports the importance of its monitoring. Among the many modifications that occur in these dammed environments, drainage alterations that increase water residence time can be highlighted (TUNDISI and MATSUMURA-TUNDISI, 2008). Greater nutrient availability – related to pollution sources – associated with drainage changes, facilitates excessive phytoplankton proliferation, what makes many reservoirs to become eutrophic environments (TUNDISI, 2008). Combining the great potential of remote sensing for monitoring water systems, with the significant amount of aquatic color radiometry sensors recently launched and with expanding policies for open data access, remotely sensed data are becoming an even more accessible and cost-effective approach for water quality assessment in inland waters (ZHENG and DIGIACOMO, 2017a). Among the variety of parameters used for evaluating inland water quality, phytoplankton biomass is a fundamental one, which is used for addressing the trophic status of a water system (ZHENG and DIGIACOMO, 2017b). Phytoplankton biomass is frequently estimated via chlorophyll-a (Chl-a) concentrations measurements, since it is a photosynthetic pigment that occurs in all phytoplankton species. Through remotely sensed data it is possible to retrieve Chl-a and other optically significant constituents (OSCs) concentrations in a aquatic system. Empirical and semi-analytical approaches are common for water constituent retrieval (MOREL, 1980; CARDER et al., 1999) and have been extensively investigated and developed (MATTHEWS, 2011; ODERMATT et al., 2012). Empirical algorithms directly 15 relate remote sensed measurements to the constituent of interest, usually through statistical regression, while semi-analytical models are based on radiative transfer inverse modeling. In this context, Lee et al. (2002) developed a quasi-analytical algorithm (QAA) that obtain from remote sensing reflectance (Rrs), the inherent optical properties (IOPs) of a water body by combining empirical, semi-analytical and analytical steps. QAA outputs are the backscattering coefficient of suspended particles (bbp) and total absorption coefficient (at), and in sequence, from at, the absorption coefficient of phytoplankton (aϕ) and of colored dissolved organic matter (CDOM) plus detritus (aCDM) are derived. It was originally developed for ocean waters and has been updated since then – currently in its sixth version (LEE, 2014). Due to the direct relationship between OSCs and IOPs, it is possible to obtain Chl-a concentrations in a water body from models based on its IOPs. As reported by Odermatt et al. (2012), Chl-a retrieval algorithms are commonly presented as band arithmetic based on the Chl-a absorption maximums at 442 nm and at 665 nm (BRICAUD et al., 1995), reflectance peak at ~700 nm (GITELSON, 1992) and fluorescence emission region at 681 nm (GOWER et al., 1999). Therefore, the main configurations of these algorithms, besides the empirical ones, are blue-green – common for ocean waters – (GORDON and MOREL, 1983) and red-near infrared (NIR) – widely used for turbid waters – (GITELSON, 1992; GONS, 1999) band ratio using two, three or four spectral bands. As other water constituents, as CDOM and detritus, present significant absorption in these spectral regions, blue-green band algorithms are not suitable for inland waters, especially because these OSCs do not covary with Chl-a in these environments (SUN et al., 2012). Besides presenting OSCs independently variable, inland waters usually presents phytoplankton biomass, CDOM and detritus in higher concentrations than the ones found in open ocean and thus, are considered optically complex systems (PALMER et al., 2015). In addition, OSCs concentrations are significantly variable in different inland water bodies. In this sense, the potential of QAA for addressing IOPs variability is noticeable, as through re- parameterization and calibration of some steps it is also able to retrieve IOPs in different inland water bodies (LE et al., 2009a; YANG et al., 2013; MISHRA et al., 2013, 2014; LI et al., 2013, 2015; WATANABE et al., 2016; RODRIGUES, 2017). The complex constitution of inland aquatic systems, in addition to the variation of its optical properties in different water bodies and even within the same one (PALMER et al., 2015), explains why bio-optical modeling for these systems is a pertinent investigation subject. 16 QAA versions re-parameterized for inland waters comprise aquatic environments with different bio-optical properties and most of them are only proved to be suitable for a unique water body. Overall, turbid and eutrophic waters were mostly considered in these investigations. For these environments, main alterations proposed include, fundamentally, shifting the reference wavelength (λ0) and calibration of empirical steps (YANG et al., 2013). The alteration of λ0 toward longer wavelengths (λ), where pure water absorption is dominant, aims to avoid the influence of OSCs in estimating at (LEE et al., 2002; 2009); adjustment of empirical steps of the algorithm intend to adequate to the IOPs of each water environment. Specifically regarding reservoirs in tropical regions, the first QAA re-parameterized version was developed based on Funil and Itumbiara reservoirs and it is referred as QAACDOM (OGASHAWARA et al., 2016). This version aimed to be suitable to water systems with greater proportion of CDOM in its OSCs, since further developed versions, to date, were aimed at phytoplankton dominated waters. Watanabe et al. (2016) and Rodrigues (2017) versions were configured based on two different reservoirs located in Tietê River cascading system, which is formed by six cascading reservoirs, all built for hydroelectric generation purposes. For Barra Bonita reservoir (BBHR), the first of them, Watanabe et al. (2016) developed QAABBHR version. The re-parameterization conducted by Rodrigues (2017) was designed for Nova Avanhandava reservoir (NHR), the fifth reservoir of the system, and was named QAAOMW. The most recent QAA original versions – QAAV5 (LEE et al., 2009) and QAAV6 (LEE, 2014) – showed to be unsuitable for both cited reservoirs, presenting errors in magnitudes which make the algorithm, in this form, improper. Ibitinga reservoir (IHR) is also part of the Tietê River cascading system, and it is located downstream from BBHR e upstream from NHR. The distinct influences arising from the drainage basin (SMITH et al., 2014), as well as the variations caused by cascading effect (BARBOSA et al., 1999), make the reservoirs along Tietê River to have different compositions. Ecological processes in a cascading reservoir system and, consequently, its water quality parameters, are subject to changes that are considered continuous, as stated by Barbosa et al. (1999) through the Cascading Reservoir Continuum Concept (CRCC). The main changes observed by these authors in the Tietê River cascading system were the continuous decrease of turbidity along the system – especially the inorganic fraction – and reduction of nutrients concentration downstream; phytoplankton biomass presented significant decrease only in the latest reservoirs. The hydrodynamic along these systems are also changeable according to the different influences of tributary rivers and 17 by floodgates opening and closing mechanisms (SMITH et al., 2014). These water systems are also subject to several impacts originating from, as examples, the lack of riparian forest and abundant pollution sources (SMITH et al., 2014). All these processes contribute to the differences in the reservoirs constitutions and, therefore, to the variability in their bio-optical properties. BBHR and NHR are clear examples of this variability. On one hand, BBHR is classified as a turbid and highly productive aquatic environment, (DELLAMANO- OLIVEIRA et al., 2007), presenting a wide range of Chl-a concentrations and organic matter predominance in its SPM (WATANABE et al., 2016); this reservoir receives huge pollutant discharges from point and diffuse sources. In contrast, NHR presents, according to seasonal and geographic variations, oligo-to-mesotrophic waters, with low nutrients concentrations and higher proportion of inorganic matter in its SPM (RODRIGUES, 2017). In IHR, in turn, it has already been observed organic fraction predominance in its SPM and broad spatial and temporal variability of Chl-a concentrations (CAIRO et al., 2017). According to Novo et al. (2013), the trophic state of the reservoir fluctuates from mesotrophic to hypereutrophic, what reinforces the spatial variability characteristic reported by Cairo et al. (2017). OSCs ranges observed by Cairo et al. (2016) in IHR are intermediaries when comparing with the intervals found in BBHR by Watanabe et al. (2016) and in NHR, by Rodrigues (2017). However, some particularities, such as organic predominance in the SPM and higher Chl-a concentrations, make IHR characteristics comparable to BBHR ones. 1.2. Hypothesis Considering that the bio-optical characteristics in IHR are possibly similar to those observed in BBHR, the hypothesis proposed is that, among the assessed algorithms, only QAABBHR will be able to retrieve accurate values of absorption coefficients of phytoplankton (aϕ) for this reservoir and, thus, it will be possible to accurately retrieve Chl-a concentrations in IHR from these coefficients. It is still hypothesized that the same QAA version will not be capable of retrieving CDOM plus detritus absorption coefficient (aCDM), since it was re- parameterized considering a phytoplankton dominated water system. 1.3. Objectives The main objective of this study is to estimate Chl-a concentrations in Ibitinga Reservoir via semi-analytical schemes. 18 1.3.1. Specific objectives In order to support this central aim, specific objectives are: (a) To characterize Ibitinga reservoir bio-optically and bio-geochemically; (b) To assess QAAV5, QAAV6, QAABBHR e QAAOMW performances for estimating absorption coefficients in the study area and to assess Chl-a concentration estimate models for the reservoir; (c) To analyze Chl-a concentrations spatial distribution in the IHR via Ocean and Land Colour Instrument (OLCI) – Sentinel-3A (S3) image, as well as to verify the potential of this sensor for inland water monitoring; 1.4. Dissertation Structure The dissertation was organized in four chapters. This first chapter (Chapter 1) presents the contextualization of the theme addressed in this research, as well as the hypothesis tested and the objectives proposed. Chapter 2 presents the characterization of the study area and the main information about the field campaigns carried out. Chapters 3 and 4 are structured as scientific papers, composed by introduction of the specific theme, methods – including data acquisition and processing – results, discussions and conclusion. Last chapter (Chapter 5) shows the conclusions about the present study and the recommendations for future work. The two next paragraphs in this subsection briefly describe what is comprised in Chapters 3 and 4, which correspond to scientific papers already submitted for publication. CHAPTER 3: POTENTIALS AND LIMITATIONS OF A QUASI-ANALYTICAL ALGORITHM FOR RETRIEVING THE WATER ABSORPTION PROPERTIES IN A CASCADING RESERVOIR SYSTEM As a first step in the direction of addressing the hypothesis proposed, the first scientific paper was aimed at assessing the suitability of QAA existing versions in retrieving absorption coefficients in the study area. This was the first step in order to verify the possibility of obtaining Chl-a concentrations in IHR – which was the final objective – since the following step was dependent of accurate retrieval of the IOPs. Chapter 3 presents the results of the assessment of QAAV5, QAAV6, QAABBHR and QAAOMW performances in the study area, in terms of aϕ and aCDM. The data used was obtained in a dedicated field campaign with 29 sampling stations. Chapter 3 also addresses the bio-optical characterization of IHR. 19 CHAPTER 4: QUASI-ANALYTICAL ALGORITHM BASED SCHEMES FOR CHLOROPHYLL-A RETRIEVAL IN A TROPICAL RESERVOIR VIA OLCI – SENTINEL-3A IMAGE Chapter 4 was aimed at assessing schemes for Chl-a retrieval in IHR; these schemes are formed by QAA plus models that use absorption coefficients as inputs for obtaining Chl-a concentrations. In this chapter, an OLCI – S3 image was used for analyzing the spatial distribution of Chl-a in the study area. The four absorption coefficients based models tested were calibrated for IHR using in situ data from the first fieldwork, and cross-validated; each one of the four models were calibrated using absorption coefficients originated from QAAV5, QAAV6, QAABBHR and QAAOMW, resulting in sixteen models. After that, all the schemes (QAA plus model calibrated from its outputs) were applied to the OLCI image atmospherically corrected, since all of them presented reasonable accuracy after validation. The performances of the sixteen schemes in obtaining Chl-a concentrations in IHR were assessed through new in situ data from a second fieldwork, coincident with the image acquisition date. 20 CHAPTER 2: STUDY AREA 2.1. Characterization Ibitinga reservoir was built for hydroelectric generation purpose and operates as a run- of-river system. It lies in a transitional region between tropical and subtropical climate, specifically in a humid subtropical climate, characterized by well-defined dry and wet seasons. IHR is located in the middle course of Tietê River, central area of São Paulo State, Brazil (21º45' S and 48º 59' W). It is the third of six cascading reservoirs, with Barra Bonita and Bariri reservoirs located upstream and Promissão, Nova Avanhandava and Três Irmãos reservoirs downstream (Figure 1 (b)). Its flooded area is around 114 km², with approximately 9 m of average depth and an average flow of 525 m³ s - ¹. The two main tributary rivers are Jacaré-Guaçu and Jacaré-Pepira, and the reservoir extends over 70 km in Tietê River plus 25 km over each one of its main tributaries (LONDE et al., 2016). The surrounding area presents grazing land, sugarcane crops and minor areas of reforestation and secondary vegetation (GUIMARÃES et al., 1998). The river receives wastewater discharges from domestic and industrial sources along its course, in special huge amounts originating from São Paulo city, located upstream of the reservoir. The catchment area also comprises diffuse sources of pollution, as agriculture – notably sugarcane crops – and cattle breeding activities; IHR is inserted in a region that presented significant expansion of sugarcane cultivation in the past ten years (RUDORFF et al., 2010). The trophic state of the reservoir is highly spatially and temporally variable, comprising mesotrophic to eutrophic regions (NOVO et al., 2013; CAIRO et al., 2017). Londe et al. (2016) show, via remote sensing, how the changes in water residence time over the hydrological year affect the phytoplankton proliferation and its spatial distribution in IHR, reaffirming the evidences of fluctuation in its trophic state. The authors found a link between increasing of water residence time and water surface coverage by phytoplankton blooms. The Water Resources Management Unit in which IHR is inserted (UGRHI 13, Tietê- Jacaré) supplies water resources for multiple uses, as domestic and industrial supply, irrigation, navigation, domestic and industrial wastewater discharge, as well as hydroelectric energy generation (CETESB, 2005). This management unit has presenting increasing demand for urban uses and the availability of water per capita presents a reduction trend over the years, what can be associated to population growth (CBH – TJ, 2013). Also, Tietê-Jacaré 21 watershed has one of the lowest rates of natural vegetation remaining in São Paulo State, presenting only 2.2% of the vegetation cover in the State. 2.2. Field Campaigns Two field campaigns, hereafter referred to as IBI1 and IBI2, were conducted respectively between July 19 and 23, 2016 and on June 21, 2017 – austral winter – in order to obtain water quality data, spectral data and water samples in 29 sampling stations in IBI1 and 6 stations in IBI2 (Figure 1 (c)). IBI1 data were already described by Andrade et al. (2017, submitted). Figure 1. (a) São Paulo State location in Brazil; (b) Cascading reservoirs in Tietê River. Please note Ibitinga Hydroelectric Reservoir (black arrow) upstream from Nova Avanhandava reservoir and downstream from Barra Bonita reservoir (see text for more details); (c) Distribution of sampling stations from IBI1 (July, 2016) and IBI2 (June, 2017) field campaigns in the study area. Figure 2 presents the average of daily precipitation rate (millimeters per day) in the reservoir and surrounding area, from June, 2016 to June, 2017. It can be observed that the first field (July, 2016) campaign occurred after a significant period of reduced precipitation, what was expected since it happened during the dry season in the region. However, the second period of data collection (June, 2017), despite of also performed in the dry season, was preceded by days of intense rainfall. (a) (b) (c) 22 Figure 2. Area-averaged of precipitation rate from June, 2016 to June, 2017. Source: NASA/GIOVANNI (https://giovanni.gsfc.nasa.gov/giovanni/). 23 CHAPTER 3: POTENTIALS AND LIMITATIONS OF A QUASI-ANALYTICAL ALGORITHM FOR RETRIEVING THE WATER ABSORPTION PROPERTIES IN A CASCADING RESERVOIR SYSTEM 3.1. Introduction The applicability of Quasi-Analytical Algorithm (QAA) (LEE et al., 2002) in optically complex inland waters is certainly more challenging than when working on Case 1 waters (ODERMATT et al., 2012). In order to make QAA suitable for more complex waters, calibration and re-parameterization of the original versions are required, as demonstrated by several authors (see for example, LE et al., 2009a; YANG et al., 2013; MISHRA et al., 2013, 2014). When it comes to cascading reservoirs systems, variability trends are also influenced by cascading effects (BARBOSA et al., 1999), as well as by the different contributions arising from the drainage basin, tributary rivers and floodgates mechanisms (SMITH et al. 2014). In such aquatic systems there are widely differing optical properties and the efficiency of parameterized versions is uncertain. Tietê River cascading system is located in São Paulo state, a densely populated region of Brazil where water quality is an urgent matter. Recently, two QAA versions were parameterized for two reservoirs from this system: QAABBHR (WATANABE et al., 2016) and QAAOMW (RODRIGUES, 2017). The QAABBHR was parameterized for the eutrophic Barra Bonita Hydroelectric Reservoir (BBHR) (DELLAMANO-OLIVEIRA et al., 2008), which is the first reservoir of the cascading system, receiving large amounts of pollutants from the metropolis of São Paulo and also from agriculture and cattle raising (BARBOSA et al., 1999). Suspended particulate matter (SPM) in this reservoir is dominated by the organic fraction (SMITH et al., 2014; WATANABE et al., 2016). The QAAOMW version was parameterized for the oligo-to-mesotrophic Nova Avanhandava Reservoir (NHR), located further downstream, presenting relatively lower Chlorophyll-a (Chl-a) and SPM concentrations, and inorganic predominance in its SPM (SMITH et al., 2014). Ibitinga Hydroelectric Reservoir (IHR) is also situated in the Tietê River, downstream from BBHR and upstream from NHR. Chl-a and SPM concentrations values in IHR are typically within the ranges found in BBHR and NHR (CAIRO et al., 2017). However, due to some particularities, as organic matter predominance and slightly higher Chl-a concentrations, IHR is possibly more similar to BBHR bio-optical characteristics. Considering that IHR bio- optical status is comparable to BBHR one, the hypothesis here tested is that QAABBHR could be able to consistently retrieve absorption coefficients of phytoplankton pigments (aϕ) in IHR, 24 but may be not capable of accurately retrieve the colored dissolved organic matter (CDOM) plus detritus (aCDM) absorption coefficient, as QAABBHR was parameterized considering phytoplankton features of the BBHR reservoir (WATANABE et al., 2016). To test the aforementioned hypothesis we aim to assess the performance of QAABBHR and QAAOMW, besides two latest native forms of QAA - QAAV5 and QAAV6 (LEE et al., 2009 and LEE, 2014, respectively) - in retrieving aϕ and aCDM for IHR. We decided to also test the QAAV5 and v6 in order to compare the results from the original and re-parameterized versions. This investigation relates to the purpose of monitoring all cascading system in an integrated manner, through a unique QAA version properly adjusted for capturing the bio- optical variability occurring along the system. We also addressed water quality and bio- optical characterization of IHR, intending to associate it with algorithms efficiency, as this was not previously reported in the literature. 25 3.2. Material and Methods 3.2.1. Water quality data Secchi disk depth (m), turbidity (NTU) and electric conductivity (µs/cm) data were obtained in all 29 sampling stations (Figure 1(c)). Water samples were collected in order to estimate Chl-a concentration, through acetone extraction method (GOLTERMAN et al., 1978; LORENZEN et al., 1967). The SPM concentrations, as well as organic (OSM) and inorganic suspended matter (ISM) fractions were estimated according to the American Public Health Association protocol (APHA, 1998). 3.2.2. Radiometric Data Total upwelling radiance (Lt(λ); W m -2 sr -1 nm -1 ) and atmospheric diffuse radiance (Ls(λ); W m -2 sr -1 nm -1 ) were measured using two RAMSES-ARC hyperspectral radiometers. Downwelling irradiance (Ed(λ); W m -2 nm -1 ) was measured using a RAMSES-ACC sensor (TriOS, Oldenburg, Germany). All radiometric sensors used operate in the spectral range between 350 and 900 nm, with a spectral resolution of 3.3 nm. The acquisition geometry followed Mobley (1999), with Lt(λ) measured at a zenith angle (θ) of 140°, Ls(λ) at θ=40° and Ed(λ) measured with the sensor aligned to the zenith, i.e., θ=0°; all measurements were taken at an azimuth angle of 90° (Figure 3). Figure 3. (a) Acquisition geometry proposed by Mobley (1999). Adapted from Rodrigues (2017); (b) Radiometric sensors arrangement in field works. These radiometric quantities were then applied to estimate remote sensing reflectance (Rrs, sr -1 ) spectra, using the spectral optimization approach proposed by Lee et al. (1999), in order to remove surface-reflected radiance (Lr(λ)) (LEE et al., 2010). Rrs spectra, which were estimated through in situ hyperspectral measurements, were then resampled to simulate (a) (b) 26 satellite data. The spectral response functions of each Ocean and Land Colour Instrument (OLCI) – Sentinel-3A bands were used to derive band-weighted data. 3.2.3. Phytoplankton absorption coefficient (aϕ) and Detritus plus CDOM absorption coefficient (aCDM) To estimate CDOM absorption coefficient (aCDOM, m -1 ), water samples previously filtered through GF/F Whatman fiberglass filters – 0.7 μm porosity – were again filtered through Whatman nylon membrane – 0.22 μm porosity and 47 mm diameter. The absorbance of the filtrates was read using a 2600 UV-VIS spectrophotometer (Shimadzu, Japan), and the results were applied to calculate aCDOM as proposed by Bricaud et al. (1981) (Eq.1). r OD a CDOM CDOM )( 3.2   (Eq.1) where ODCDOM (λ) is the optical density of CDOM and r is the cuvette path length (0,1m). To determine the total particulate (algal and detritus) absorption coefficient (ap, m -1 ), water samples were filtered through fiberglass GF/F Whatman – 0.7 μm porosity and 47 mm diameter. Then, Transmittance-Reflectance (T-R) method described by Tassan and Ferrari (1995, 1998) was employed using a double-beam 2600 UV-VIS spectrophotometer equipped with an integrating sphere. The spectral sampling ranged from 280 nm to 800 nm, with a spectral resolution of 1 nm. In order to bleach pigments from the filters, a 10% sodium hypochlorite (NaCLO) solution was used and T-R measurements were taken again. By eliminating pigments influence, it was possible to obtain detritus absorption coefficient (ad, m - 1 ); ap and ad are then obtained according to Eq.2: (Eq. 2) where ODp(λ) is the optical density of total particulate, ODd(λ) is the optical density of detritus, V is the filtered volume (m³) and A is the filter clearance area (m²). The phytoplankton absorption coefficients (aϕ, m- 1 ) were calculated by subtracting ad from ap. Finally, aCDM is calculated by adding aCDOM to ad. 3.2.4. QAA QAAV5 and QAAV6 performances were evaluated for IHR. It is important to highlight that QAAV6 was developed in order to improve QAA accuracy in water systems in which Rrs 27 at 670 nm is greater than 0.0015 sr -1 , i.e., coastal regions and areas with high sediments concentration. QAABBHR and QAAOMW, the two re-parameterized versions for Tietê River cascading system were also tested. The dataset considered by Watanabe et al. (2016) for QAABBHR parameterization presented high values and a broad range of Chl-a concentration (from 17.7 to 797.8 mg m -3 ; average 274.5 mg m -3 ) and SPM concentration ranging from 3.6 to 44.0 g m -3 (average 14.6 g m -3 ) , with its organic fraction corresponding to approximately 90% of total SPM in the samples. Differently, the dataset used by Rodrigues (2017) for QAAOMW parameterization showed low average Chl-a concentration (16.15 mg m -3 , ranging from 2.46 to 38.59 mg m -3 ), as well as low average SPM concentration (1.85 g m -3 , varying from 0.10 to 5.30 g m -3 ); inorganic fraction was predominant in total SPM. Statistic metrics used to assess the QAAs performances for IHR were normalized root mean square error (nRMSE) and mean absolute percentage error (MAPE), considering aϕ and aCDM estimated from water samples obtained in the field campaign as reference data. 28 3.3. Results and Discussion 3.3.1. Water quality characterization Chl-a and SPM concentrations varied widely (1.37 - 119.04 mg m - ³ and 1.00 - 8.10 g m - ³, respectively), showing significant spatial variability (Table 1). OSM represented 63% of the SPM, which means organic matter is slightly predominant. Turbidity values also showed some fluctuation, although it was lower than the variation presented by Chl-a and SPM (coefficient of variation 23.28%); in general, turbidity presented a relative low average value (4.24 NTU). Secchi disk presented an average of 2.23 m and the lowest coefficient of variation (15.50%). Table 1. Descriptive statistics of water quality parameters in IHR; SD: standard deviation; CV: coefficient of variation; n = 29. Minimum Maximum Mean SD CV (%) Chl-a (mg m - ³) 1.37 119.04 19.34 24.71 127.79 SPM (g m - ³) 1.00 8.10 2.45 1.40 57.19 OSM/SPM 0.29 0.88 0.63 0.15 23.28 ISM/SPM 0.12 0.71 0.37 0.15 40.11 Depth (m) 9.50 21.60 14.90 4.29 28.77 Turbidity (NTU) 2.82 8.87 4.24 1.19 28.01 Secchi Depth (m) 1.60 3.20 2.23 0.35 15.50 aϕ(443) (m - ¹) 0.06 1.88 0.30 0.36 120.44 aCDM(443) (m - ¹) 1.24 5.93 1.95 0.83 42.41 ad (443) (m - ¹) 0.14 0.62 0.37 0.13 33.54 aCDOM(443) (m - ¹) 0.87 5.31 1.57 0.80 50.85 Regarding the absorption coefficients, ad (443) and aϕ (443) presented similar average values, while aCDOM(443) presented the highest values - greater than ad (443) and aϕ (443). Detritus can be constituted of mineral matter, humus or organic remains; considering that organic fraction of SPM showed to be slightly higher in this dataset and aϕ is not dominant, detritus in the reservoir is probably mainly constituted of humus and organic remains that are not resulting from phytoplankton degradation. 29 3.3.2. Bio-Optical Characterization In situ measured Rrs spectra (Figure 4 (a)) shows absorption features around 675 nm and a reflectance peak around 700 nm that can be associated with Chl-a. However, these Rrs features are not observed in all stations, indicating spatial variability in terms of optically significant constituents (OSC) at IHR. A prominent feature of reflectance in the green spectral region, around 550 nm, can be observed with different magnitudes (Figure 4 (a)). This reflectance peak is usually related with scattering from algal cells (GITELSON, 1992), and also with inorganic suspended matter, which shifts higher reflectance values toward longer wavelengths (HAN and RUNDQUIST, 1997). Average aCDOM was the greater contributor for at until around 660 nm (Figure 4 (b)). Average ad presented higher values than average aϕ through blue and green spectral regions; approximately at 600 nm average aϕ shows slight increase, and it is considerably greater than average ad from 660 nm to 700 nm (Figure 4 (b)). Figure 4. (a) In situ Rrs spectra; (b) Average absorption coefficients by phytoplankton (aϕ), CDOM (aCDOM), detritus (ad) and pure water (aw); n=29. Ternary plots (Figure 5) display the relative contribution of aCDOM, aϕ and ad to the total absorption at IHR. Similar datasets from BBHR (WATANABE et al., 2016) and NHR (RODRIGUES, 2017), collected in May, 2014 and May, 2016, respectively are also plotted for comparison purpose. It is relevant to highlight that all three dataset considered here were collected during the dry season. 0.000 0.002 0.004 0.006 0.008 0.010 0.012 0.014 400 500 600 700 800 R rs ( sr -¹ ) λ (nm) 0.00 0.50 1.00 1.50 2.00 2.50 3.00 400 500 600 700 a ϕ , a C D O M , a d , a w ( m -¹ ) Wavelength (nm) aw aCDOM aϕ a d (a) (b) 30 Figure 5. Ternary plots presenting the relative contribution of detritus, phytoplankton and CDOM to the total absorption (without water fraction) at three different wavelengths for IHR (red dots), BBHR (black dots) and NHR (brown dots): (a) 443 nm; (b) 560 nm; (c) 665 nm. Absorption in IHR dataset is dominated by CDOM in all three wavelengths here considered, corresponding to 70.11 3.94% at 443 nm, 73.63 4.67% at 560 nm and 68.94 6.43% at 665 nm. Detritus contributed to the total absorption budget with 18.13 2.67% at 443 nm and 17.18 3.40% at 560 nm. At 665 nm, phytoplankton contributed with 20.08 5.27% of the total while detritus contributed with only 10.97 2.86%. In BBHR, total absorption was predominately due to phytoplankton in all three wavelengths, especially at 665 nm, with aϕ contributing with 85.72 3.18%. However, at 443 nm and 560 nm the total absorption is more balanced among all OSC, as can be observed in the scatter of dots around the central area of the plots (Figures 5 (a) and 5 (b)). In NHR dataset, otherwise, ad was the major contributor at 443 nm and 560 nm, corresponding to 62.16 4.42% and 72.02 6.52%, respectively. At 665 nm aϕ represented 65.15 5.97% of total absorption, while ad was the second greater contributor with 32.51 5.72%. These differences may be related to factors as the cascading effect (BARBOSA et al., 1999); this concept explains the changes in the water quality parameters occurring along a cascading system as reduction of turbidity, nutrients concentration and phytoplankton biomass in Tietê River downstream reservoirs. It also can be related to the changeable allochthonous contributions arriving in each reservoir. 3.3.3. QAA performances QAAOMW presented the lowest average nRMSE (%) and MAPE (m -1 ) for aϕ retrieval (Table 2). However, if we consider only 665 nm band, which is a diagnosis wavelength for (b) (a) (c) 31 Chl-a, this same version shows the highest errors (nRMSE = 31.74%, MAPE = 6.03 m -1 ). QAAOMW tuning was based on NHR dataset, comprising Chl-a concentration values lower than the ones measured in IHR (NHR average 7.94 mg m -3 ; IHR average 19.34 mg m - ³), and this difference can be related to its poor performance at 665 nm band (RODRIGUES, 2017). QAABBHR, differently, was developed for a highly productive reservoir, and IHR bio-optical status did not show to be similar to BBHR one, justifying why it was not completely suitable for retrieving IOPs in IHR. QAAV6 showed higher values of nRMSE and MAPE than QAAV5 in all wavelengths, although this performance was not initially expected since QAAV6 was proposed for Rrs (670) > 0.0015 sr -1 , which fits IHR dataset (Rrs (670) = 0.00427 sr -1 ). Comparatively, QAAV5 showed to be the most accurate version at 665 nm (NMRSE = 15.35%, MAPE = 1.19 m -1 ). Table 2. QAAs performances for aϕ retrieval, according to each band, based on nRMSE (%) and MAPE (m -1 ). QAAV5 QAAV6 QAABBHR QAAOMW Bands (nm) nRMSE % MAPE (m -1 ) nRMSE % MAPE (m -1 ) nRMSE % MAPE (m -1 ) nRMSE % MAPE (m -1 ) 412 14.04 1.36 17.14 1.93 19.46 0.70 21.94 1.84 443 15.86 1.35 19.82 1.91 13.80 0.86 21.02 1.15 490 30.77 2.44 40.53 3.29 30.86 2.46 26.07 0.98 510 32.46 2.61 44.12 3.58 35.94 2.86 19.92 0.67 560 38.14 3.21 54.41 4.55 49.70 4.03 20.49 0.60 620 14.19 1.38 21.25 3.49 20.01 3.48 16.07 0.84 665 15.35 1.19 20.23 4.09 18.06 5.46 31.74 6.03 681 15.49 0.78 19.17 2.32 14.03 2.56 23.79 1.94 709 112.00 3.78 118.04 10.99 89.48 10.70 30.64 0.45 Average 32.03 2.01 39.41 4.02 32.37 3.68 23.52 1.61 Scatterplots of aϕ estimated by the QAAs and aϕ measured in laboratory are shown in Figure 6. In general, QAAV5 overestimated aϕ in great part of wavelengths (Figure 6 (a)), except for the longer ones – specifically 681 and 709 nm – for which the version underestimated the values. The other versions overestimated most of aϕ values at all wavelengths, excluding QAABBHR at 412 nm. It can be clearly observed that a sampling spot 32 disagrees with the other ones in almost all wavelengths for all four QAA tested; this spot corresponds to the highest Chl-a and SPM concentrations of the dataset. Figure 6. Comparison between estimated and measured aϕ using: (a) QAAV5; (b) QAAV6; (c) QAABBHR; and (d) QAAOMW. For aCDM, all QAAs variants demonstrated similar average nRMSE and MAPE, presenting reasonable accuracy (Table 3). An improvement of nRMSE values can be observed with increasing wavelengths. QAAOMW showed slight better results (nRMSE = 21.74%, MAPE = 0.44 m -1 ); NHR is a non-productive reservoir and it is dominated by aCDM in the blue-green region, just as IHR, and this can be associated with the lower errors presented by QAAOMW. 0.00 0.50 1.00 1.50 2.00 2.50 0.00 0.50 1.00 1.50 2.00 2.50 QAAV5 0.00 0.50 1.00 1.50 2.00 2.50 0.00 0.50 1.00 1.50 2.00 2.50 QAAV6 412 443 490 510 560 620 665 681 709 0.00 0.50 1.00 1.50 2.00 2.50 0.00 0.50 1.00 1.50 2.00 2.50 QAABBHR 0.00 0.50 1.00 1.50 2.00 2.50 0.00 0.50 1.00 1.50 2.00 2.50 QAAOMW 412 443 490 510 560 620 665 681 709 Measured aϕ (m-¹) E st im a te d a ϕ ( m -¹ ) E st im a te d a ϕ ( m -¹ ) Measured aϕ (m-¹) (a) (b) (c) (d) 33 Table 3. QAAs performances for aCDM retrieval, according to each band, based on nRMSE (%) and MAPE (m -1 ). QAAV5 QAAV6 QAABBHR QAAOMW Bands (nm) nRMSE % MAPE (m -1 ) nRMSE % MAPE (m -1 ) nRMSE % MAPE (m -1 ) nRMSE % MAPE (m -1 ) 412 50.41 0.85 49.60 0.84 41.17 0.67 29.49 0.35 443 40.71 0.87 40.26 0.86 34.54 0.70 24.64 0.31 490 32.75 0.91 32.56 0.90 29.54 0.78 21.43 0.33 510 30.46 0.92 30.33 0.91 28.07 0.81 20.69 0.34 560 25.94 0.95 25.89 0.94 24.83 0.86 19.71 0.39 620 23.52 0.97 23.50 0.97 23.07 0.92 19.82 0.49 665 22.42 0.98 22.41 0.98 22.20 0.95 19.98 0.57 681 21.91 0.98 21.91 0.98 21.74 0.96 19.85 0.58 709 21.65 0.99 21.65 0.99 21.54 0.97 20.07 0.64 Average 29.97 0.94 29.79 0.93 27.41 0.85 21.74 0.44 Figure 7 shows that QAAV5, QAAV6 and QAABBHR consistently underestimated aCDM in all wavelengths. It probably occurs due to high CDOM and detritus concentrations and its relative predominance in the absorption budget of IHR (see Figure 5) that is not commonly found in ocean and coastal waters, and also differs from BBHR optical characteristics. In these scatterplots, two sampling spots evidently differ from the others: one of them is discrepant in all wavelengths and also corresponds to the spot with the highest Chl-a and SPM concentrations, while the second one disagrees mostly in the shortest wavelengths - except for QAAOMW, which overestimated aCDM values for all wavelengths - and it corresponds to the second highest SPM concentration in the dataset. 34 Figure 7. Comparison between estimated and measured aCDM using: (a) QAAV5; (b) QAAV6; (c) QAABBHR; and (d) QAAOMW. On the other hand, QAAOMW was able to retrieve, in general, more consistent values of aCDM (Figure 7 (d)). It is important here to highlight that QAA approach integrates the absorption by detritus and CDOM as aCDM. Taking this into account, and also considering that NHR absorption occurs mostly due to detritus, while in IHR the CDOM is predominant, it corroborates the capability of QAAOMW in retrieving aCDM in non-phytoplankton dominated waters and, consequently, its potential in determining carbon content in aquatic systems. 3.3.4. Implications for OSC monitoring Neither originals nor re-parameterized QAA versions here tested were capable of estimating accurately the absorption coefficients in all wavelengths. This shows a challenge in coping with high optical variability in cascading system. Although QAABBHR was parameterized for a high productive aquatic system, the model could not accurately retrieve the aϕ at 665 nm (aϕ (665)), which is used as a proxy to estimate Chl-a concentration. This could have happened because Chl-a concentration values in BBHR are very high (~ 797 mg m -3 ), and due to the fact that the Rrs spectra is masked by the package effect, as reported by Alcântara et al. (2016). For aϕ (665) the QAAV5 presented the lowest error. On the other hand, 0.00 2.00 4.00 6.00 8.00 0.00 2.00 4.00 6.00 8.00 QAAV5 0.00 2.00 4.00 6.00 8.00 0.00 2.00 4.00 6.00 8.00 QAAV6 412 443 490 510 560 620 665 681 709 0.00 2.00 4.00 6.00 8.00 0.00 2.00 4.00 6.00 8.00 QAABBHR 0.00 2.00 4.00 6.00 8.00 0.00 2.00 4.00 6.00 8.00 QAAOMW 412 443 490 510 560 620 665 681 709 Measured aCDM (m-¹) Measured aCDM (m-¹) E st im a te d a C D M ( m -¹ ) E st im a te d a C D M ( m -¹ ) (a) (b) (c) (d) 35 to estimate the aCDM at 443 nm (aCDM (443)), which is a proxy for carbon content, the QAAOMW presented the lowest error. These results highlighted the limitation of such quasi- analytical scheme in order to monitor the spatial-temporal OSCs in the cascading system. However, the obtained results also represents an opportunity to better understand the complexity of these aquatic systems and to figure out how to improve the QAA in order to use only one version to estimate the OSCs in the entire cascade. Up to now the monitoring of OSCs from space operationally is still a challenge. 36 3.4. Conclusion Since QAA native versions, QAAV5 and QAAV6, were designed for ocean and coastal waters, a poor performance for optically complex inland waters was already expected and it was here confirmed. However, QAABBHR did not perform satisfactorily as supposed, demonstrating that even presenting some similarity regarding the OSCs, the bio-optical status of BBHR and IHR are substantially different and, thus, IHR IOPs cannot be derived through this re-parameterized QAA. Although QAAOMW was able to retrieve relatively accurate average values for aCDM and aϕ in IHR, it presented an unsuitable performance for aϕ at 665nm, indicating that this version also has its limitations for deriving IHR IOPs. Therefore, our hypothesis can be rejected, since QAABBHR was not suitable for retrieving neither aϕ nor aCDM. Variability of bio-optical characteristics along the cascading system was confirmed, since none re-parameterized version for reservoirs in the same system was completely suitable for IHR. The re-parameterized versions showed not to be sensitive to the variations occurring in the cascade. Further research is necessary, aiming to achieve a QAA re-parameterization appropriate for IHR and, prospectively, for the whole Tietê River cascading system. Acknowledgements The authors thank FAPESP Projects (Process N° 2012/19821-1 and 2015/21586-9) and Professor Edivaldo Velini and staffs from FCA/UNESP for allowing the use of their laboratory facilities. 37 CHAPTER 4: QUASI-ANALYTICAL ALGORITHM BASED SCHEMES FOR CHLOROPHYLL-A RETRIEVAL IN A TROPICAL RESERVOIR VIA OLCI – SENTINEL-3A IMAGE 4.1. Introduction Inland waters are sensitive to anthropogenic disturbances and have been affected by environmental changes resulting from these pressures, what clearly explains the demand for assessing and monitoring water quality parameters (PALMER et al., 2015). Besides the known potential of remote sensing for monitoring lakes, reservoirs and rivers (BUKATA, 2013), it is still challenging to retrieve optical and biogeochemical properties in these commonly optically complex water systems (MOUW et al., 2015; PALMER et al., 2015). Water optically significant constituents such as phytoplankton pigments, suspended particulate matter (SPM) and colored dissolved organic matter (CDOM), vary widely and independently to each other in inland waters. These constituents, called optically significant constituents (OSC), are directly related to water quality. OSC variation is also substantial between different inland water bodies, increasing the obstacles for the development of a universally applicable model for quantifying their concentrations from remote sensing measurements. Chlorophyll-a (Chl-a), in particular, is a photosynthesizing pigment that occurs in all phytoplankton species and, thus, is a key parameter for water quality. Besides its role in carbon sequestration and primary productivity (ROESLER et al., 1989), phytoplankton excessive proliferation can lead to eutrophication processes (AYRES et al., 1996). Therefore, Chl-a concentration is considered the main variable to indicate the trophic state of aquatic environments, serving as a proxy of their ecological health. While relatively simple blue-to-green band-ratio algorithms presents acceptable results for Chl-a retrieval in open ocean (O'REILLY et al., 1998), they are inadequate for inland waters, due to a greater influence of other OSC, such as CDOM and detritus, which present strong contribution in the absorption budget at these spectral regions. Aiming to quantify Chl- a and other OSC, several algorithms based on remote sensing measurements were developed (ODERMATT et al., 2012). These algorithms can be classified in empirical, which relate remote sensing reflectance (Rrs) or irradiance reflectance (R) to Chl-a through statistical regression, and semi-analytical, based on the radiative transfer theory (MOREL, 1980; CARDER et al. 1999). While the first present limitations for widening temporal and spatial application, the second often require input parameters and their performance depend on the 38 representativeness of the model used for estimating the inherent optical properties (IOPs) of the environment in which it is being applied. Quasi-analytical algorithm (QAA), a special case which includes empirical and analytical steps for retrieving absorption, a, and particle backscattering, bbp, coefficients (LEE et al., 2002), represents a great contribution in this sense. It is broadly applicable for ocean waters, for which it was originally parameterized. Re-parameterizations and calibrations can make QAA suitable for inland waters. Nonetheless, until now, each parameterization is only proven to be suitable for a unique water body (LE et al., 2009a; YANG et al., 2013; MISHRA et al., 2013, 2014; LI et al., 2013, 2015; WATANABE et al., 2016; RODRIGUES, 2017, among others). These versions were mainly developed for turbid and eutrophic waters. In order to adapt QAA for these waters, the main enhancements proposed by the authors were the shift of reference wavelength to near-infrared bands, adjustment of spectral slope for particle backscattering and calibration of other empirical steps. Changing reference wavelength to longer wavelengths, where pure water absorption is dominant, aims to avoid OSC influences in a(λ) estimation (LEE et al., 2002); other changes intend to adequate the algorithm to the IOPs of each water system. Further assessment of how QAAs versions perform in different inland water bodies, goes in the direction of expanding its applicability. By integrating QAA with models that use absorption coefficients – QAA outputs – as inputs, it is possible to compose schemes for Chl- a concentration retrieval. Chl-a monitoring in a reservoir is addressed in this study, which is aimed at assessing the suitability of schemes composed by QAA plus absorption coefficients based models, in order to obtain Chl-a concentration in the Ibitinga Reservoir, Tietê River, Brazil, via Ocean and Land Colour Instrument (OLCI) image. Originals and re- parameterized QAA versions were assessed, as well as four absorption coefficients based models. In order to support this main objective, specific aims were: to carry out optical and biogeochemical characterization of the study area; to analyze spatial distribution of Chl -a concentrations in Ibitinga Reservoir; and to assess OLCI image performance for Chl-a retrieval in an inland water body. Since the sensor recently became operational – first data released in October, 2016 (ESA, 2016) – its use for inland water monitoring still has great potential to be investigated. 39 4.2. Methods 4.2.1. Water quality data Water samples collected just below air-water interface in the sampling stations were used for estimating Chl-a and suspended particulate matter (SPM) concentrations. Secchi disk depth (m), turbidity (NTU) and electric conductivity (µs/cm) data were also acquired in all IBI1 and IBI2 stations, using respectively a Secchi disk, a turbidimeter and a conductivity meter. Chl-a concentration is determined via acetone extraction method (GOLTERMAN et al., 1978), through which samples are filtered using Whatman fiberglass filters – 0.7 μm porosity and 47 mm diameter – on the day of collection and frozen until laboratory analysis. In sequence, the filters are macerated with 10 ml of acetone solution (90%) and centrifuged. Absorbances at wavelengths of 665 and 750 nm of the resulting liquid phase are determined by spectrophotometry, before and after acidification with hydrochloric acid solution (0.1 N) and, using the absorbance values, Chl-a concentrations are finally calculated (LORENZEN et al., 1967). SPM concentrations, as well as organic (OSM) and inorganic suspended matter (ISM) fractions were obtained according to the American Public Health Association protocol (APHA, 1998). Samples were filtered on the day of collection and filters were kept refrigerated until further procedures. The filters were then dried in the oven at 100°C for 12 hours, weighted using an analytical balance and, after that, taken to muffle furnace at 550° for 30 minutes and again weighted. The first weights are used for determining SPM concentration, by subtracting the value from the filter weight before the filtration; the second ones determine ISM, also by subtracting from the original weight before filtration. OSM concentrations are obtained by the difference between SPM and ISM weights. 4.2.2. Absorption coefficients Absorbance of water samples filtered through GF/F Whatman fiberglass filters – 0.7 μm porosity – and sequentially filtered through Whatman nylon membrane – 0.22 μm porosity and 47 mm diameter – were read using a 2600 UV-VIS spectrophotometer (Shimadzu, Japan). The results were applied to calculate CDOM absorption coefficients (aCDOM, m -1 ) as proposed by Bricaud et al. (1981) (Eq. 3). 40 r OD a CDOM CDOM )( 3.2   (Eq.3)) where ODCDOM(λ) is the optical density of CDOM and r is the cuvette path length (0,1m). In order to determine the total particulate (algal and detritus) absorption coefficients (ap, m -1 ), GF/F Whatman fiberglass filters – 0.7 μm porosity and 47 mm diameter – , previously used for filtering water samples, were used for reading optical density of total particulate using a double-beam 2600 UV-VIS spectrophotometer equipped with an integrating sphere. Transmittance-Reflectance method (T-R) described by Tassan and Ferrari (1995, 1998) was employed. For eliminating pigments influence, a 10% sodium hypochlorite (NaCLO) solution was applied in the filters and optical density now corresponding to detritus was measured, also through T-R technique. The spectral sampling ranged from 280 nm to 800 nm, with spectral resolution of 1 nm. From these measurements it was possible to calculate detritus absorption coefficients (ad, m -1 ) and ap according to Eq. 4: (Eq. 4) where ODp(λ) is the optical density of total particulate; ODd(λ) is the optical density of detritus, V is the filtered volume (m³) and A is the filter clearance area (m²). Phytoplankton absorption coefficients (aϕ, m- 1 ) are obtained by subtracting ad from ap. Finally, aCDOM was added to ad in order to obtain aCDM, since QAA approach derives aCDOM and ad as this combined coefficient. 4.2.3. Radiometric Data In situ data collection also included radiance and irradiance measurements using three RAMSES sensors (TriOS, Oldenburg, Germany), operating in a spectral range between 350 and 900 nm, with spectral resolution of 3.3 nm. RAMSES-ACC measured downwelling irradiance (Ed(λ); W m -2 nm -1 ) and two RAMSES-ARC sensors measured total upwelling radiance (Lt(λ); W m -2 sr -1 nm -1 ) and atmospheric diffuse radiance (Ls(λ); W m -2 sr -1 nm -1 ). The acquisition geometry followed Mobley (1999), with Lt(λ) measured at a zenith angle (θ) of 140° and azimuth angle (ϕ) of 90°, Ls(λ) measured at θ of 40° and ϕ of 90°, and Ed(λ) measured at θ of 0°, i.e. sensor aligned to Zenith, and ϕ of 90°. Remote sensing reflectance (Rrs, sr -1 ) spectra was then calculated from these radiometric quantities, using spectral optimization approach proposed by Lee et al. (1999) as follow: 41 (Eq. 5) where is the total remote sensing reflectance, which is the ratio of Lt(λ) to Ed(λ); is the surface Fresnel reflectance based on the viewing geometry (~0.021); is the sky remote sensing reflectance, the ratio of surface-reflected radiance ( ) to Ed(λ); and is a spectrally constant offset for adjusting sun glint effects from Lt(λ) measurements. Although for oceans Δ can be considered as zero in the near-infrared, in turbid waters it is significant. The absorption caused by SPM decreases to almost zero at wavelengths longer than 700 nm, and the increase of SPM concentration create a gradient in the near-infrared of Rrs due to the particle scattering (DEKKER, 1993; GOODIN et al., 1993; YANG et al., 2013). For obtaining Δ, can be modeled as a function of spectral IOPs and compared with calculated from above-surface measurements and, then, Δ is estimated through optimization that minimizes the error between the modeled and the optimized (LEE et al. 2010). Rrs obtained from in situ measurements are inputs for QAAs. Aiming to make its outputs comparable to satellite-derived ones, hyperspectral Rrs was subject to convolution, then becoming correspondent to satellite data. Spectral response functions of the first twelve bands of OLCI, on board Sentinel 3A, were used to derive band-weighted data (GORDON, 1995): ∫ ∫ (Eq. 6) where stands for the remote sensing reflectance convoluted from OLCI spectral bands; and are the lower and upper limit of the band , respectively; and S(λ) is the spectral response function of the ith spectral band of OLCI. It is important to highlight that OLCI was still not operating when the first field campaign (IBI1) was performed and, therefore, we only had OLCI image matching to field campaign available for IBI2. 4.2.4. QAA context QAAV6 (LEE, 2014), a re-design from the previous QAA original version – QAAV5 (LEE et al., 2009) – was developed in order to improve its accuracy in water systems in which Rrs at 670 nm is greater than 0.0015 sr -1 , i.e., coastal regions and areas with high sediment concentration. Both QAAV5 and QAAV6 were used for composing the schemes for Chl-a 42 retrieval. QAABBHR (WATANABE et al., 2016) and QAAOMW (RODRIGUES, 2017), re- parameterized versions for reservoirs located at Tietê River cascading system were also tested for the schemes. Barra Bonita, reservoir for which QAABBHR was parameterized, is a eutrophic environment located upstream from Ibitinga reservoir and presents a broad range of Chl-a concentrations, with OSM corresponding to approximately 90% of total SPM (WATANABE et al., 2016). Nova Avanhandava Reservoir, located downstream from Ibitinga, was the basis for QAAOMW parameterization. Contrasting with Barra Bonita, this reservoir presents oligo to mesotrophic waters, with low average Chl-a concentration, as well as low average SPM concentration in the dataset considered for the parameterization; inorganic fraction is usually predominant in total SPM (RODRIGUES, 2017). 4.2.5. Calibration and validation of absorption coefficients based models Absorption coefficients based models for Chl-a concentration retrieval were calibrated for the study area. Firstly, Rrs spectra from 29 samples obtained in IBI1 field campaign and resampled for OLCI bands, were used as inputs for the four QAA versions aforementioned. Outputs from each QAA - aϕ and aCDM at different wavelengths - were applied to obtain four indexes, which were subsequently used to calibrate the models. Therefore, each QAA version originated four indexes and, consequently, four models. The three first indexes, here labeled γ1, γ2 and γ3 (Eqs. 7, 8, 9) are based on two-band (2B) (DEKKER, 1993), three-band (3B) (DALL‟OLMO and GITELSON, 2005; GITELSON et al., 2008) and four-band (4B) (LE et al., 2009b) approaches, and were adapted by Le et al. (2013) in terms of aϕ and pure water absorption coefficient (aw); the remaining one (γNDCI; Eq. 10) is an adaptation in terms of aϕ, aw and aCDM of the Normalized Difference Chlorophyll Index (NDCI) (MISHRA and MISHRA, 2012) proposed by Watanabe et al. (2016). Original indexes use Rrs at different wavelengths as inputs. (Eq. 7) (Eq. 8) (Eq. 9) (Eq. 10) 43 Using Chl-a concentrations from IBI1 as response variable, and γ1, γ2, γ3 and γNDCI as explanatory variables – one index for each model – four models were adjusted through linear regression. In addition, each index was calculated from absorption coefficients originating from the four QAA versions and, thus, sixteen schemes – four QAA versions plus four models – were configured in total. Leave-one-out cross-validation method was used for checking representativeness of the schemes. After validation was done, accuracy of the models was assessed in order to verify which models could be applied in the image. Figure 8 (a) shows the steps for models calibration and validation. 4.2.6. Chl-a retrieval via OLCI image A cloud free OLCI full resolution Level-1 image acquired in June 21, 2017, matching IBI2 fieldwork date, was selected. Using Rrs spectra collected in IBI2 and convolved for OLCI bands, empirical line (EL) atmospheric correction (KRUSE et al., 1990) was applied in order to obtain bottom of atmosphere Rrs from image top of atmosphere (TOA) radiance. In sequence, the schemes that presented reasonable performances in the cross-validation were applied to the image. Firstly, QAA were applied in the image and, then, the absorption coefficients (aϕ and aCDM) resulting from it, were used to calculate γ1, γ2, γ3 and γNDCI indexes. These calculated indexes, applied to the models previously calibrated for the study area, resulted in Chl-a concentration estimates in the image. Figure 8 (b) shows the steps for image processing. 44 Figure 8. Flowcharts of (a) calibration process of sixteen Chl-a retrieval models for the study area; (b) image pre-processing and application of schemes for Chl-a concentration retrieval. 4.2.7. Accuracy assessment Averaged unbiased absolute percentage difference (ε, Eq. 11) and normalized root mean square error (nRMSE, Eq. 12) were used to (i) evaluate the absorption coefficients based models after validation, (ii) to assess OLCI image atmospheric correction and also (iii) to assess Chl-a retrieval schemes performance for the study area via OLCI image. Reference data for accuracy assessment of each step were, respectively, Chl-a concentrations from IBI1, Rrs from IBI2 resampled for OLCI bands and Chl-a concentrations from IBI2. [ ∑ | | ] (Eq. 11) √ ∑ (Eq. 12) where n is the number of samples; xestimated and xmeasured are the estimated and measured values, respectively; and are respectively the maximum and minimum measured values. (a) (b) 45 4.3. Results 4.3.1. Water quality characterization Chl-a concentrations presented broad variation in IBI1, while in IBI2 both the range and the concentration values were lower (Table 4). SPM concentrations were higher in IBI1, as well as turbidity values. Organic and inorganic fractions of SPM presented the same trend in both datasets, with the organic matter being predominant; the second fieldwork, however, showed greater percentage of OSM. Variation of conductivity values was much higher in the second dataset, while the first one showed greater fluctuation for turbidity and also presented wider ranges of Secchi depth. Table 4 presents descriptive statistics of biogeochemical data from both field campaigns. Table 4. Descriptive statistics of water quality data from IBI1 and IBI2 field campaigns. Min: minimum; Max: maximum; SD: standard deviation; CV: coefficient of variation; n = 29 for IBI; n = 6 for IBI2. C h l- a (m g m - ³) S P M (g m - ³) O S M /S P M IS M /S P M C o n d u ct iv it y (μ s cm -1 ) T u rb id it y (N T U ) S ec ch i D ep th (m ) IBI1 Min 1.37 1.00 0.29 0.12 171.00 2.82 1.6 Max 119.04 8.10 0.88 0.71 198.30 8.87 3.2 Mean 19.34 2.45 0.63 0.37 185.71 4.24 2.23 SD 24.71 1.40 0.15 0.15 8.37 1.19 0.35 CV (%) 127.8 57.19 23.28 40.11 4.51 28.01 15.5 IBI2 Min 2.73 0.30 0.67 0.13 155.30 2.63 1.90 Max 13.38 2.20 0.88 0.33 205.00 3.60 2.80 Mean 9.97 1.25 0.76 0.24 170.53 3.06 2.31 SD 4.35 0.66 0.09 0.09 19.50 0.33 0.33 CV (%) 43.62 52.52 12.33 39.55 11.43 10.78 14.40 46 4.3.2. Bio-optical characterization Descriptive statistics of absorption coefficients from IBI1 and IBI2 are shown in Table 5. In both field campaigns aCDM was the absorption coefficient that presented the highest average value at 443 nm (aCDM(443)) and, although the maximum value was much higher in IBI1, average values were similar in both datasets. aCDOM(443) showed the same tendency as aCDM(443), with substantially lower maximum values in IBI2 and very similar mean values. aϕ(443) and ad(443) average values, as well as their maximum, presented drastic reduction from IBI1 to IBI2. Variation of all absorption coefficients was much higher in the first dataset. Table 5. Descriptive statistics of absorption coefficients from IBI1 and IBI2 field campaigns. Min: minimum; Max: maximum; SD: standard deviation; CV: coefficient of variation; n = 29 for IBI; n = 6 for IBI2. Minimum Maximum Mean SD CV (%) IBI1 aϕ(443) (m - ¹) 0.06 1.88 0.30 0.36 120.44 aCDM(443) (m - ¹) 1.24 5.93 1.95 0.83 42.41 ad (443) (m - ¹) 0.14 0.62 0.37 0.13 33.54 aCDOM(443) (m - ¹) 0.87 5.31 1.57 0.80 50.85 IBI2 aϕ(443) (m - ¹) 0.05 0.13 0.11 0.03 26.92 aCDM(443) (m - ¹) 1.27 1.99 1.69 0.28 27.04 ad (443) (m - ¹) 0.15 0.19 0.18 0.02 9.30 aCDOM(443) (m - ¹) 1.12 1.79 1.51 0.27 17.74 Average aϕ, ad and aCDOM for IBI1 and IBI2 are shown in Figure 9; aw (POPE and FRY, 1997) is also shown in order to assist analysis of total contribution of the absorption budget throughout the spectrum. It is possible to observe aCDOM is dominant in the blue, green and red spectral regions of IBI2 (Figure 9 (b)); this trend was also observed by Andrade et al. (2017, submitted) in IBI1 (Figure 9 (a)), although aw becomes the greater contributor above 660 nm in this dataset. In IBI2, ad contribution is the second highest value in the blue-green region, even though the difference from aϕ is really modest. The difference of ad and aϕ contributions is more distinct in IBI1, but ad also contributes more than aϕ, as already shown by Andrade et al. (2017, submitted). In the red channel, aϕ and ad presented similar contributions in both datasets, with minor aϕ increase from around 660 nm to 700 nm in both cases; this feature is more clearly seen in IBI1. 47 Figure 9. Average phytoplankton (aϕ), CDOM (aCDOM), detritus (ad) and pure water (aw) absorption coefficients in Ibitinga reservoir in (a) IBI1 (July, 2016; n=29); (b) IBI2 (June, 2017; n=6). Rrs spectra obtained from the two fieldwork measurements (Figure 10) present a peak in the green region, at approximately 550 nm, that can be related to scattering caused by inorganic suspended matter (HAN and RUNDQUIST, 1997) and low Chl-a absorption. Although this feature is noticeable in the spectra, its magnitude is relatively low in IBI2. It is also possible to observe features related to Chl-a, as the absorption feature at around 675 nm and reflectance peak at around 700 nm, which are more evident in IBI1 spectra. A discrete reflectance feature occurs around 750 nm in both datasets, followed by a minor dip at 760 nm that can possible be related to absorption by atmospheric oxygen (GOWER et al., 2005). Rrs spectra from IBI1 (already shown by ANDRADE et al., 2017, submitted) present, in general, slightly higher magnitudes when compared with IBI2. Figure 10. Rrs spectra obtained from in situ measurements in: (a) IBI1 field campaign (n=29); (b) IBI2 field campaign (n=6). 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 400 500 600 700 a ϕ , a C D O M , a d , a w ( m -¹ ) aw aCDOM aϕ 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 400 500 600 700 aCDOM aw ad aϕ 0 0.004 0.008 0.012 0.016 400 500 600 700 800 R rs ( sr -¹ ) 0 0.004 0.008 0.012 0.016 400 500 600 700 800 (b) (a) Wavelengths (nm) (a) (b) Wavelengths (nm) ad 48 Measured Rrs spectra after convolution to match OLCI bands can be visualized in Figure 11. It is possible to observe that the reflectance peak at ~550 nm is preserved after the convolution. Chl-a features, however, were significantly softened. Figure 11. Rrs spectra from IBI2 (June, 2017, n=6) simulated for OLCI first twelve bands. 4.3.3. Validation of Chl-a concentration retrieval schemes The four models calibrated for the study area using Chl-a concentration from IBI1 and γ1, γ2, γ3 and γNDCI indexes, through linear regression, are labeled respectively as mγ1, mγ2, mγ3 and mγNDCI. Each scheme – QAA plus models originated from its output – was cross- validated and the results can be visualized in Table 6; mγ1, mγ2 and mγ3 presented minimal difference between its errors and, thus, are displayed together. QAABBHR plus mγNDCI presented the lowest nRMSE and ε, although all schemes showed reasonable errors (nRMSE < 27.00%; ε < 3.50%) Table 6. Errors between fits from cross-validated models and Chl-a concentrations from IBI1. QAAV5 QAAV6 QAABBHR QAAOMW mγ1,mγ2, mγ3 mγNDCI mγ1,mγ2, mγ3 mγNDCI mγ1,mγ2, mγ3 mγNDCI mγ1,mγ2, mγ3 mγNDCI nRMSE (%) 22.16 18.17 24.25 24.21 11.58 11.46 24.73 26.89 ε (%) 2.36 2.27 3.03 3.03 1.21 1.30 2.81 3.13 Considering that the four models originated from a same QAA version consistently presented similar errors, t-test for paired samples was applied and it was verified their 0 0.003 0.006 0.009 0.012 400 500 600 700 800 R rs ( sr -1 ) Wavelengths (nm) 49 statistically equality (α=0.05; p-value=1). As the models were proven to obtain analogous results, allied to the fact that all schemes demonstrated acceptable errors after cross- validation, all of them were employed in the next processing step. 4.3.4. Atmospheric Correction Assessment OLCI-derived Rrs after EL atmospheric correction were assessed based on Rrs obtained from IBI2 measurements and resampled for OLCI bands (Figure 12 (a)); errors referrers to the average for all IBI2 sampling stations. The highest errors (nRMSE and ε) occurred at 412 nm. Figure 12. (a) Errors between OLCI-derived Rrs spectra after Empirical Line (EL) atmospheric correction versus Rrs spectra from field campaign (IBI2) measurements resampled for OLCI bands (average for all sampling stations). (b) Comparison of OLCI- derived Rrs spectrum after atmospheric correction and Rrs spectrum from IBI2 measurements resampled for OLCI bands at sampling station 6. First twelve OLCI bands are shown in both (a) and (b). On the other hand, the lowest nRMSE occurred at 620 nm (34.77%), while ε was lowest at 443 nm (5.05%). Most bands presented nRMSE ranging from ~35 to ~53% and ε from ~5 to ~16%. Regarding all bands, average nRMSE was 63.41% and average ε was 13.64%. Bias (not shown here) ranged from -0.0013 to 0.0021 m -1 , with negative values in the majority of the wavelengths analyzed, which means that the empirical line method presented a trend of underestimating Rrs values. Figure 12 (b) shows OLCI-derived Rrs spectrum after EL atmospheric correction and Rrs spectrum from IBI2 fieldwork at a unique sampling station (station 6). Overall, it presents the same tendency that can be seen in Figure 12 (a), with major discrepancies occurring in the blue channel, while the red spectral region presents minor ones. 0 50 100 150 200 250 400 500 600 700 800 E rr o rs ( % ) Wavelengths (nm) ε nRMSE 0.0000 0.0025 0.0050 0.0075 0.0100 400 500 600 700 800 Wavelengths (nm) OLCI-derived Rrs Measured Rrs (a) (b) R rs ( sr -1 ) 50 Figure 13. Comparison per band of in situ Rrs and OLCI-derived Rrs after Empirical Line atmospheric correction; six satellite-in situ matchups are considered. Comparing, per band, OLCI-derived Rrs after atmospheric correction and Rrs calculated from in situ measurements (Figure 13), it can be observed that values from 412 nm band were consistently overestimated and are the most discrepant ones. 560 nm band also presents most of values overestimated, while the majority of other bands presents a slight underestimating trend. 4.3.5. Assessment of Chl-a Retrieval via OLCI image After applying all schemes – QAA plus model – in the image already atmospherically corrected, the accuracy of Chl-a concentration estimates retrieved from each scheme was assessed using Chl-a concentrations obtained in IBI2. The QAAV5 plus mγNDCI scheme presented the best results – nRMSE of 42.42% and ε of 37.82% – while the other models (mγ1, mγ2, mγ3) tested with QAAV5 outputs showed slight greater errors (Table 7). Despite that, all QAAV5 schemes presented the lowest errors when compared with schemes using other QAA versions. Only QAAV5 based schemes presented reasonable results, since all schemes based on other QAA versions presented nRMSE greater than 287.29 % and ε greater than 122.96%. Table 7. Models performances for indexes estimated via QAAV5 applied in the image. mγNDCI mγ1 mγ2 mγ3 nRMSE (%) 42.42 47.48 46.36 46.36 ε (%) 37.82 39.59 38.76 38.76 0 0.002 0.004 0.006 0.008 0.01 0 0.002 0.004 0.006 0.008 0.01 O L C I- d er iv ed R r s (s r-1 ) Measured Rrs (sr-1) 400 412 443 490 510 560 620 665 674 681 709 754 51 Models set through linear regression, using indexes obtained via QAAV5 outputs and Chl-a concentrations from IBI1 are displayed in Figure 14. mγNDCI presents the highest R² (0.5592), although other models also show R² around 0.5. Considering all schemes (not shown), QAABBHR plus mγNDCI demonstrated the best fit, with R² of ~0.84; however, as already explained, QAABBHR schemes did not present good results for image data (corresponding with IBI2), despite of the great fits of the models for IBI1 data. Figure 14. Models configured from (a) γ1 (mγ1); (b) γ2 (mγ2); (c) γ3 (mγ3); and (d) γNDCI (mγNDCI). Indexes used QAAV5 outputs. Chl-a concentrations obtained via QAAV5 based schemes are shown in Figure 15, since these schemes were the only ones able to retrieve reasonable accurate estimates (nRMSE < 47.50% and ε < 39.60%). It can be verified in the images that tributaries rivers present low Chl-a concentrations when comparing to concentrations from the main body of the reservoir. Since average Chl-a concentration in the sampling stations was ~ 10 mg m -3 in IBI2 dataset, and in the image these stations are mostly located at yellow areas (~ 13 mg m -3 ), it is possible to infer that the schemes slightly overestimated the concentrations. It is Chl-a = 113.43γ2 - 51.973 R² = 0.5312 0 20 40 60 80 100 120 140 0 0.5 1 1.5 C h l- a ( m g .m -3 ) γ1 mγ1 Chl-a = 407.31γ3 + 61.17 R² = 0.5312 0 20 40 60 80 100 120 140 -0.2 -0.1 0 0.1 C h l- a ( m g .m -3 ) γ2 mγ2 Chl-a = 294.16γ4 + 14.761 R² = 0.5312 0 20 40 60 80 100 120 140 -0.2 0 0.2 0.4 C h l - a ( m g .m -3 ) γ3 mγ3 Chl-a = 191.11γNDCI + 64.169 R² = 0.5592 0 20 40 60 80 100 120 140 -0.5 -0.25 0 0.25 C h l - a ( m g .m -3 ) γNDCI mγNDCI (a) (b) (c) (d) 52 also possible to observe that station 4 is located at a yellow to red transition region and it is actually the higher concentration station. Figure 15. Spatial distribution of Chl-a concentrations in Ibitinga Reservoir using OLCI image acquired on June 21 st , 2017. Chl-a retrieved via (a) QAAV5 plus mγ1 scheme; (b) QAAV5 plus mγ2 scheme; (c) QAAV5 plus mγ3 scheme; (d) QAAV5 plus mγNDCI scheme. (a) (b) (c) (d) 53 4.4. Discussion Water quality data analysis shows not only significant variability within each fieldwork dataset, but also a huge variability of values and amplitudes of most parameters from one field campaign to another. Water quality data analysis shows not only significant variability within each fieldwork dataset, but also a huge variability of values and amplitudes of most parameters from one field campaign to another. These observations may be related to two main factors: the water residence time and precipitation rates preceding the data collection, elements which are also related to each other. Water residence time varies cyclically over time, affecting the nutrients concentration and, consequently, the trophic state in a reservoir (LONDE et al., 2016; VIEIRA et al., 2002); according to Vieira et al. (2002), water residence time in IHR is relatively short, what leads to more frequents nutrients cycling. Considering that the data collections occurred after different precipitation amounts in the region – the first fieldwork was carried out after a period of low precipitation rates, while the second occurred after considerable precipitation amounts (NASA/GIOVANNI, 2016, 2017) –, it possible affected the water residence time, that may have been longer during the first data collection, what triggered higher nutrients concentration and, consequently, may have caused phytoplankton proliferation in more intense rates. In the other hand, the second data collected presented lower values for water quality parameters, what may be related to the precipitations amounts leading to a shorter residence time, and reduced nutrients availability and, then, fewer occurrences of phytoplankton blooms. Vieira et al. (2002) also observed that the limnological variables were influenced by the location of the sampling stations, supporting the fact that spatial variations are standard in IHR. It can be related, for example, to the distinct contribution arriving from the watershed; IHR presents a narrow and prolonged shape, similar to a river channel, extending in 70 km along Tietê River, and this significant spatial distribution may contribute to different inputs