JEFFERSON FERREIRA-FERREIRA INFLUENCE OF FLOOD DYNAMICS ON FOREST CARBON STOCKS, LITTERFALL SEASONALITY AND NET PRIMARY PRODUCTIVITY IN AMAZONIAN VÁRZEA FORESTS RIO CLARO / SP DEZEMBRO - 2018 UNIVERSIDADE ESTADUAL PAULISTA “Júlio de Mesquita Filho” Instituto de Geociências e Ciências Exatas Campus de Rio Claro JEFFERSON FERREIRA-FERREIRA Influence of flood dynamics on forest carbon stocks, litterfall seasonality and net primary productivity in amazonian várzea forests Tese de Doutorado apresentada ao Instituto de Geociências e Ciências Exatas do Câmpus de Rio Claro, da Unversidade Estadual Paulista “Júlio de Mesquita Filho”, como parte dos requisitos para obtenção do título de Doutor em Geografia. Orientador: Prof. Dr. Thiago Sanna Freire Silva Rio Claro - SP 2018 F383i Ferreira-Ferreira, Jefferson INFLUENCE OF FLOOD DYNAMICS ON FOREST CARBON STOCKS, LITTERFALL SEASONALITY AND NET PRIMARY PRODUCTIVITY / Jefferson Ferreira-Ferreira. -- Rio Claro, 2018 111 p. : il., tabs., fotos, mapas Tese (doutorado) - Universidade Estadual Paulista (Unesp), Faculdade de Ciências Farmacêuticas, Araraquara, Rio Claro Orientador: Thiago Sanna Freire Silva 1. Várzeas amazônicas. 2. Sensoriamento remoto por radar. 3. Estoques de carbono. 4. Produtividade primária líquida. 5. Serrapilheira. I. Título. Sistema de geração automática de fichas catalográficas da Unesp. Biblioteca da Faculdade de Ciências Farmacêuticas, Araraquara. Dados fornecidos pelo autor(a). Essa ficha não pode ser modificada. JEFFERSON FERREIRA-FERREIRA Influence of flood dynamics on forest carbon stocks, litterfall seasonality and net primary productivity in amazonian várzea fores ts Tese de Doutorado apresentada ao Instituto de Geociências e Ciências Exatas do Câmpus de Rio Claro, da Unversidade Estadual Paulista “Júlio de Mesquita Filho”, como parte dos requisitos para obtenção do título de Doutor em Geografia. Comissão Examinadora Prof. Dr. Thiago Sanna Freire Silva - Orientador IGCE/UNESP/Rio Claro (SP) Dr. Luiz Eduardo de Oliveira e Cruz de Aragão INPE/São José dos Campos (SP) Dra. Evlyn Márcia Leão de Moraes Novo INPE/São José dos Campos (SP) Dr. Conrado de Moraes Rudorff CEMADEN/São José dos Campos (SP) Profa. Dra. Leonor Patricia Cerdeira Morellato IB/UNESP/Rio Claro (SP) Conceito: APROVADO Rio Claro/SP, 29 de Outubro de 2018 v Resumo Dois macro-ambientes podem ser distinguidos entre os tipos de vegetacionais da Amazônia: áreas de terra firme, predominantemente florestadas e não suscetíveis a inundações, e as áreas úmidas. A extensão total das áres úmidas na Amazônia é de cerca de 30% da bacia amazônica, das quais mais de 25% são constituídas por planícies fluviais. Na Amazônia, a amplitude das variações sazonais do nível da água pode atingir até 16 m na Amazônia Ocidental, 10 m na Amazônia Cen- tral e 6 m na Amazônia Oriental, com extensão e duração da inundação local dependendo da interação entre precipitação, descarga fluvial e geomorfologia. Os processos ecológicos e ambientais nessas planícies são amplamente controla- dos pelo pulso de inundação - um conceito teórico que postula que a amplitude, duração, frequência e periodicidade (previsibilidade) dos pulsos de inundação são os principais fatores que mantêm o equilíbrio ambiental dinâmico. As planícies fluviais amazônicas desempenham um papel importante nos ciclos biogeoquími- cos regionais e na manutenção da biodiversidade, além de fornecer importantes serviços ecossistêmicos para as sociedades humanas. No entanto, esses am- bientes permanecem largamente negligenciados em relação às estimativas dos ciclos biogeoquímicos na escala da bacia hidrográfica. Assim, o presente estudo tem como objetivo avançar nossa compreensão de como a heterogeneidade es- pacial em termos de diferentes fitofisionomias florestais e padrões de inundação nas florestas de várzea controla três aspectos ecológicos principais: estoques de carbono, produção e dinâmica de serapilheira e produtividade primária líquida. Usando imagens multitemporais de sensoriamento remoto de radar de abertura sintética (SAR) combinadas com dados de campo, derivamos mapas de fitofi- sionomias e de duração da inundação na área de estudo. Também desenvolvemos um método empírico de modelagem de inundação baseado em modelagem logís- tica para nos permitir estender as capacidades atuais de prever padrões espaciais de inundação. Em seguida, exploramos como a heterogeneidade espacial gera padrões de estoques de carbono acima do solo usando métodos de inventário florestal. Monitorando um ano de queda de serapilheira, analisamos como a produção e a dinâmica da serapilheira variam entre comunidades florestais dis- tintas sob diferentes regimes de inundação e também estimamos a produtividade primária líquida e discutimos como ela varia no espaço. Palavras-chave: Várzeas Amazônicas; Sensoriamento remoto por radar; Es- toques de carbono; produtividade primária líquida; serrapilheira. Abstract Two macro-environments can be distinguished among Amazonian vegetation types: upland areas, predominantly forested and not susceptible to flooding, and wetland areas. The full extent of wetlands in the Amazon is about c.a. 30% of the Amazon basin, where floodplains comprise more than 25% of wet- land areas. In the Amazon, the amplitude of seasonal water level variations can reach up to 16 m in Western Amazon, 10 m in Central Amazon, and 6 m in Eastern Amazon, with local flood extent and duration depending on the inter- action among precipitation, river discharge and geomorphology. Ecological and environmental processes in floodplains are largely controlled by the flood pulse - a theoretical concept which postulates that amplitude, duration, frequency and periodicity (predictability) of flood pulses are the major factor maintaining the dynamic environmental equilibrium in floodplains. Amazonian floodplains play an important role in the regional biogeochemical cycles and biodiversity mainte- nance, and provide important ecosystem services to human societies. However, these environments remain greatly overlooked regarding basin-scale estimates of biogeochemical cycles. Therefore, the present study aims to advance our un- derstanding of how spatial heterogeneity in terms of different forest subtypes and flood patterns of várzea forests controls three key ecological aspects: car- bon stocks, litterfall production and dynamics, and net primary productivity. Using multitemporal synthetic aperture radar (SAR) remote sensing imagery combined with field data, we derived vegetation structure and flood duration classes in the study area. We also developed an empirical flood modeling method based on logistic modeling to allow us to extend the current capabilities in pre- dict spatial patterns of flooding. Then we explored how spatial heterogeneity drives patterns of aboveground carbon stocks using standard forest inventory methods. Monitoring one full year of litterfall, we asked how litter produc- tion and dynamics vary among distinct forest communities under different flood regimes and we also estimated net primary productivity and discussed how it varies in space. Keywords: Amazonian várzeas; Radar remote sensing; Carbon stocks; Net primary productivity; litterfall. List of Figures 2.1 Conceptual cross-section diagram showing the three main forest subtypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.2 Location of the study area . . . . . . . . . . . . . . . . . . . . . . 8 2.3 Visual interpretation key used to select training and validation samples for vegetation mapping . . . . . . . . . . . . . . . . . . . 13 2.4 Mean water stage and shaded 95% confidence interval for the 1991-2011 period, measured at the Mamirauá Lake gauge . . . . 15 2.5 Major vegetation types and habitats of the southeastern portion of the Mamirauá Sustainable Development Reserve . . . . . . . . 17 2.6 Temporal variation of ALOS/PALSAR backscattering coefficients for the three main woody vegetation classes . . . . . . . . . . . . 18 2.7 Extent of estimated flood duration classes (in days), based on ALOS/PALSAR imagery . . . . . . . . . . . . . . . . . . . . . . 19 2.8 Estimated flood duration based on a time series of ALOS/ PAL- SAR image data . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.9 Relative area for each combination of land cover and flood dura- tion classes derived from ALOS/PALSAR image time series . . . 21 3.1 Localization of study area . . . . . . . . . . . . . . . . . . . . . . 29 3.2 Installed leveloggers to flood monitoring in the Southeastern por- tion of Mamirauá Sustainable Development Reserve . . . . . . . 32 3.3 Root mean squared errors (RMSE) of the 4-fold cross validation 35 3.4 Flood modeling results . . . . . . . . . . . . . . . . . . . . . . . . 36 3.5 Histograms of flood duration as predicted by the logistic model, by year. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 4.1 Association between flood heights and flood duration recorded at leveloggers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 4.2 Aboveground forest carbon by forest type . . . . . . . . . . . . . 49 4.3 Aboveground forest carbon proportion by diametric classes and forest subtype . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 4.4 Relationship between aboveground forest carbon and flood duration 50 4.5 Relationship between input parameters of allometric models used to estimate aboveground forest carbon and flood duration . . . . 51 4.6 Relationship between metrics related aboveground forest carbon distribution at plot-level and flood duration . . . . . . . . . . . . 52 5.1 Total litterfall and water level heights . . . . . . . . . . . . . . . 60 5.2 Total litterfall and water level heights by vegetation type . . . . 61 vii LIST OF FIGURES viii 5.3 Relationship between litterfall seasonality and water level variation 62 5.4 Relationship between mean annul litter production and flood du- ration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 5.5 Relative contribution of bi-weekly production of different litterfall components and water level heights . . . . . . . . . . . . . . . . . 64 5.6 Relationship between bi-weekly production of litterfall compo- nents and water stage . . . . . . . . . . . . . . . . . . . . . . . . 65 5.7 Relative contribution of bi-weekly production of different litterfall components to total fine litterfall by forest subtypes . . . . . . . 66 5.8 Photosynthetic and reproductive components of fine litterfall by vegetation types . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 5.9 Relative investment into reproductive and photosynthetic com- ponents of fine litterfall by forest subtypes . . . . . . . . . . . . . 68 5.10 Mean annual production of fine litter components by forest subtypes 70 5.11 Association between mean annual production of fine litterfall components and flood duration . . . . . . . . . . . . . . . . . . . 71 5.12 Association between mean annual production of fine litterfall components and flood duration controlling for forest subtype . . 72 5.13 Relationship between investment in photosynthetic organs rela- tive to reproductive organs (RL ratio), flood duration and forest subtypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 5.14 Relationship between net primary productivity and flood duration 75 6.1 Spatial distribution of net primary productivity as a linear func- tion of flood duration. . . . . . . . . . . . . . . . . . . . . . . . . 82 6.2 Spatial distribution of net primary productivity as a function of flood duration controlling for forest subtype. . . . . . . . . . . . . 83 List of Tables 1.1 Above-ground carbon and structural metrics of inventoried forest sample plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2.1 ALOS/PALSAR synthetic aperture radar images acquired . . . . 11 2.2 Water stage for each map of inundation extent, and corresponding flood duration classes . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.3 Confusion matrix and accuracy indices for the classification of vegetation types . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.4 Comparison of flood duration estimates as determined in situ using temperature gauges and as derived from ALOS/PALSAR images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 3.1 ALOS-1/PALSAR images acquired for different water levels to map flood extent in the Mamirauá Sustainable Development Re- serve (Central Amazon, Brazil). The range of captured water stages comprises 70% of the maximum historical amplitude for this location. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 3.2 Best logistic models to estimate the flood duration. . . . . . . . . 34 3.3 Comparison of historical and 2016 water levels. . . . . . . . . . . 35 4.1 Sample design of forest sample plots . . . . . . . . . . . . . . . . 43 4.2 Allometric models to estimate the aboveground biomass . . . . . 44 4.3 Floristic summary of inventoried forest plots . . . . . . . . . . . . 47 4.4 Above-ground carbon and structural metrics of inventoried forest sample plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.5 Paired differences in mean aboveground carbon between forest subtypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 5.1 Bi-weekly maximum (Max), minimum (Min), mean and mean annual fine litterfall . . . . . . . . . . . . . . . . . . . . . . . . . 61 5.2 Paired differences in mean annual fine litter production between forest subtypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 5.3 Paired differences in mean annual production of fine litterfall components between forest subtypes . . . . . . . . . . . . . . . . 69 5.4 Estimates of total Net Primary Productivity (NPP) by forest type 74 5.5 Paired differences in net primary productivity between forest sub- types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 5.6 Comparison of fine litterfall production of different Amazonian forests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 ix Contents 1 INTRODUCTION 1 2 COMBINING ALOS/PALSAR DERIVED VEGETATION STRUC- TURE AND INUNDATION PATTERNS TO CHARACTER- IZE MAJOR VEGETATION TYPES IN THE MAMIRAUÁ SUSTAINABLE DEVELOPMENT RESERVE, CENTRAL AMA- ZON FLOODPLAIN, BRAZIL 5 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2.1 Study Area . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2.2 Remote sensing data and processing . . . . . . . . . . . . 10 2.2.3 Image segmentation and classification . . . . . . . . . . . 11 2.2.4 Flood extent mapping . . . . . . . . . . . . . . . . . . . . 14 2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.4.1 Management implications . . . . . . . . . . . . . . . . . . 23 2.4.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 24 3 MODELING THE SPATIAL AND TEMPORAL DYNAMICS OF INUNDATION IN AMAZONIAN FLOODPLAINS US- ING RADAR REMOTE SENSING AND LOGISTIC MODEL 26 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 3.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 3.2.1 Remote sensing data and flood mapping . . . . . . . . . . 28 3.2.2 Flood monitoring . . . . . . . . . . . . . . . . . . . . . . . 31 3.2.3 Flood modeling with logistic regression . . . . . . . . . . 32 3.3 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . 34 4 ABOVEGROUNDCARBON STOCKS IN AMAZONIAN VÁRZEA FORESTS: THE ROLE OF FOREST SUBTYPES AND FLOOD REGIME 41 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 4.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 4.2.1 Sample Design . . . . . . . . . . . . . . . . . . . . . . . . 42 4.2.2 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 43 4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 4.3.1 Forest inventory . . . . . . . . . . . . . . . . . . . . . . . 45 4.3.2 Aboveground carbon stocks (AGC) . . . . . . . . . . . . . 48 4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 x CONTENTS xi 5 LITTERFALL AND NET PRIMARY PRODUCTIVITY IN AMAZONIAN VÁRZEA FORESTS: THE ROLE OF FOREST SUBTYPES AND FLOOD REGIME 55 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 5.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 5.2.1 Sample Design . . . . . . . . . . . . . . . . . . . . . . . . 57 5.2.2 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 58 5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 5.3.1 Fine litterfall production and dynamics . . . . . . . . . . 60 5.3.2 Production and dynamics of fine litterfall components . . 64 5.3.3 NPP estimates from fine litterfall . . . . . . . . . . . . . . 74 5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 6 SYNTHESIS AND CONCLUSION 80 Chapter 1 INTRODUCTION Throughout the relatively long Amazonian natural history, the changing cli- matic, biotic, and landscape configurations (Hoorn and Wesselingh 2011; Mertes and Dunne 2007) have established complex environmental mosaics. As result, the uneven spatial distribution of current geologic, climatic, edaphic and eco- logical conditions (Anderson 2012; Mertes and Dunne 2007; Saatchi et al. 2007; Silva et al. 2013a) create an heterogeneous mosaic of plant communities. Two macro-environments can be distinguished among Amazonian vegeta- tion types: upland areas, predominantly forested and not susceptible to flood- ing, and wetland areas. Amazonian wetlands play an important role in the regional biogeochemical cycles and biodiversity maintenance, and provide im- portant ecosystem services to human societies (Luize et al. 2018; Melack and Forsberg 2001; Mitsch and Gosselink 2000; Pangala et al. 2017; Wittmann et al. 2012). However, Amazonian wetlands remain greatly overlooked regarding basin- scale estimates of biogeochemical cycles. For instance, it has been estimated that the flooded forests of the Amazon basin can contribute to carbon accumu- lations of about 0.04±0.01 Pg1 C year-1 in living aboveground biomass (Aragão et al. 2014). Nonetheless this value is still considered highly uncertain (Aragão et al. 2014) and depends on better estimates based on wetland habitat and veg- etation maps such as those conducted by Silva et al. 2010 and Ferreira-Ferreira et al. 2015. Such estimates do not consider, for instance, the ecogeographical differences between várzea and igapó forests, determined by the different nutri- tional status of these floodplains, nor the average time that each forested region remains flooded annually. Junk et al. 2011 developed a comprehensive classification of Amazonian wet- lands, using climatic, hydrological, botanical, and limnological parameters. The authors estimate that, including riparian forests along lower-order rivers (1 to 5 in river order) and floodable interfluves (white-sand vegetation known as camp- inas and campinaranas), the full extent of wetlands in the Amazon is about c.a. 30% of the Amazon basin, where floodplains comprise more than 25% of wetland areas (Table 1.1). The oligotrophic river floodplains of black- and clear-water rivers are called igapó, being poor in nutrients, while eutrophic white-water river floodplains are rich in nutrients and locally and scientifically known as várzea (Junk et al. 2012; 11 Petagram (Pg) = 1 Gigatonne (Gt) = 103 Teragram (Tg) = 106 Megagram (Mg) = 106 grams (g). 1 CHAPTER 1. GENERAL INTRODUCTION 2 Table 1.1: Classification and coverage of Amazonian wetland environments according to Junk et al. 2011. Wetland Environments Area (km2) % of Basin % of wetlands % of wetlands (excluded *) Várzea Floodplains† 275,000 3.93 13.62 27.00 Blackwater Floodplains‡ 118,000 1.69 5.85 11.59 Clearwater Floodplains 126,100 1.80 6.25 12.38 Interfluvial Wetlands 488,374 6.98 24.20 47.96 Mangroves§ 10,900 0.16 0.54 1.07 Riparian Wetlands of Low-Order Rivers* 1,000,000 14.29 49.54 Total 2,018,374 28.83 Basin area (approx) 7,000,000 † excluding paleo-várzea environments ≈ 125, 000km2). ‡ only Negro River catchment considered. § Brazilian Mangroves = 10,000 km2 + 900 km2 in Guiana Sioli 1954). Sioli 1954 developed the first scientific classification of Amazonian rivers to explain their limnological characteristics and relate them to the geological and geomorphological properties of their respective catchments. According to Sioli’s classification, the black-water rivers (e.g. Negro, Jutaí, and Tefé rivers) drain the Precambrian Guyana Shield, bordering the northern Amazon basin. Their waters have transparencies of 60-120cm (Secchi disk depth) with low amounts of suspended matter and moderate acidity (pH<5), and both the water and the substrate are relatively nutrient-poor, with large amounts of humic acids that result in reddish-brown water colors. Clear-water rivers (e.g. Branco, Tapajós, and Xingu rivers) mostly drain the Brazilian Central Plateau which borders the Amazon basin to the south. Their waters have transparencies of >150 cm, low amounts of suspended sediment and dissolved solids, variable pH (between 5 and 8), and water and substrate of low to moderate fertility, with transparent to greenish shades. Finally, white-water rivers (e.g. Solimões/Amazonas, Juruá, and Madeira rivers) have their headwaters in the Andeans foothills, transporting large volumes of suspended and dissolved sediments, and thus being rich in nutrient content. Water transparencies are low, between 20 and 60cm, and pH nearly neutral. The large sediment loads give them a characteristic whitish- brown color. Sioli’s classification was later supported by botanical studies such as from Prance 1979 and Kubitzki 1989, which attributed the ecogeographic differenti- ation among floodplain forests to water chemistry. Nutritional conditions im- posed by the different water types lead to ecological and structural differences between várzea and igapó forests. In both forest types, tree species richness tends to increase with decreasing flood height and duration, i.e. with reduced intensity of the environmental filter imposed by the flood (Wittmann et al. 2010b). However, in várzea forests tree densities decrease as flood duration and height decrease along the flood gradient, while in igapó forests the opposite occurs - tree densities increase with the decreasing stress caused by flooding CHAPTER 1. GENERAL INTRODUCTION 3 (Wittmann et al. 2010b). The oligotrophic status of waters and soils in igapó floodplains are also responsible for much slower tree radial growth. For exam- ple, the same tree species (Macrolobium acaciifolium, Leguminosae), subject to the same flooding regime in different environments, had an average diameter increment of 3.04±0.76 mm year-1 in the igapó, while its radial growth was on average 5.32±1.34 mm year-1 in the várzea (Schöngart et al. 2005). Above- ground woody biomass can be as large in igapó forests (227-304 Mg ha-1) as in várzea forests (230-270 Mg ha-1) (Schöngart et al. 2010), but the annual pro- ductivity of várzeas is approximately three times higher (Schöngart et al. 2010; Worbes 1997). About 75% of these várzeas are covered by forests, while 25% of the remain- ing areas correspond to water bodies (channels, rivers and lakes), herbaceous vegetation, and non-vegetated sandy bars (Melack and Hess 2010). The combi- nation of annual variation in water levels along the Solimões River (7-13 meters) and the essentially flat topography results in an active floodplain that can span from 20 to 100 km away from the main channel, favoring the establishment of várzea ecosystems (Goulding et al. 1995; Melack and Hess 2010). The flood pulse is a theoretical concept relating the flood regime of river- floodplain systems with the distribution and dynamics of the associated biota and biogeochemical processes. Junk et al. 1989 define the regular annual “pulse” of river discharge as the major force controlling these systems, governing the lateral exchange and recycling of nutrients and causing biotic responses that ultimately lead to morphological, anatomical, physiological, and phenological/ behavioral adaptations. Each location in a floodplain is positioned along a flood gradient, and distinct plant species will have their “optimum” along the gradient, so that local plant communities are assembled according to the flood regimes. Although this “optimum” can be modulated by other factors such as stability, structure, and fertility of the substrate, groundwater table height, and biotic processes (Junk et al. 1989), the spatial heterogeneity of flood duration is the most important factor for forest differentiation in várzeas, having a crucial influence on floristic composition and species diversity (Assis et al. 2015; Luize et al. 2015; Wittmann et al. 2010b, 2006a). Prolonged flooding imposes constraints on gas exchanges in plants. The diffusion resistance of most gases in the water is about 10,000 times greater than in the air and there is a c.a. 30-fold reduction in oxygen concentration between the gaseous and dissolved states (Mitsch and Gosselink 2015). Oxygen depletion under inundation may be even more severe in the tropics, as higher temperatures enhance the microbial oxygen demand and reduces its solubility, which is of about 14 mg l-1 at 0°C and about 7 mg l-1 at 35°C (Parolin et al. 2010). In Amazonian várzea forests, sedimentation rates are an additional cause for oxygen depletion in the rhizosphere. In areas close to the river mainstems, the annual floods can lay up to 19.8 cm year-1 of sediments (Wittmann et al. 2004), rapidly leading to anoxia in the root system and, consequently, to a decline in the energy supply of roots, eventually interrupting CO2assimilation (Junk and Piedade 2010; Parolin 2000). The modulating effect of flooding on carbon assimilation in várzea forests is known for many tree species (Parolin 2000; Parolin et al. 2010; Parolin and Wittmann 2010; Schöngart et al. 2010). However, landscape-scale estimates of carbon stocks and net primary productivity that take into account the spa- tial heterogeneity of várzea forests and its different hydrological regimes are CHAPTER 1. GENERAL INTRODUCTION 4 still missing. This spatial heterogeneity of várzea forests in carbon stocks and productivity has not been considered in the estimates of carbon fluxes for the Amazon basin, nor in dynamic global vegetation models that are coupled to biogeochemical and climate models (Cox 2001; Galbraith et al. 2010; Joetzjer et al. 2013; Marengo et al. 2012; Prentice et al. 2007; Sitch et al. 2013; Zhang et al. 2015). Therefore, the present study aims to advance our understanding of how spatial heterogeneity in terms of different forest subtypes and flood patterns of várzea forests controls three key ecological aspects: carbon stocks, litterfall production and dynamics, and net primary productivity. The present study is divided in four main chapters, each one addressing specific questions of the above goal. Chapter 2, published in Ferreira-Ferreira et al. 2015, uses multitemporal synthetic aperture radar (SAR) remote sensing imagery combined with object-based image analysis, and field data to derive vegetation structure and flood duration classes in the study area. This chap- ter provides the basis to assess the spatial heterogeneity in terms of distribu- tion of vegetation and spatial patterns of inundation. In Chapter 3, we go a step further in spatial inundation characterization, moving from flood duration classes to temporally continuous and pixel-based estimation of flood duration. In this chapter we used the same SAR imagery as the previous chapter and a logistic model to develop an empirical flood modeling method which allow us to extend current capabilities in predict spatial patterns of flooding. Chapter 4 address two questions: (i) How much and in which direction aboveground carbon stocks vary depending on forest subtypes and flood regimes? And (ii) which of the factors influencing AGC are most affected by flood duration? We answered these questions using field inventory methods combined with results from previous chapters. Monitoring one full year of litterfall, in Chapter 5 we ask three scientific questions: (i) Does fine litterfall production and dynamics differ among forest subtypes and flood regimes? (ii) Do distinct forest types and flood regimes reveal different community-level signals of varying plant in- vestment strategies into reproductive organs versus photosynthetic organs? (iii) Assuming that litterfall production can be used as a proxy for forest total NPP, what we can infer about how the primary productivity of várzea forest could respond to hydrological variability and changes? In this chapter, we estimate net primary productivity from litterfall production and discuss how it varies in space. Finally, in Chapter 6 we emphasize the main remarks of the complete study and relate them to form a more comprehensive picture of how the present work advances our understanding of the ecology and carbon biogeochemistry of várzea forests. Chapter 2 COMBINING ALOS/PALSAR DERIVED VEGETATION STRUCTURE AND INUNDATION PATTERNS TO CHARACTERIZE MAJOR VEGETATION TYPES IN THE MAMIRAUÁ SUSTAINABLE DEVELOPMENT RESERVE, CENTRAL AMAZON FLOODPLAIN, BRAZIL1 2.1 Introduction The Amazon várzea comprises the eutrophic floodplains influenced by the sediment-rich white-water rivers of the Amazon basin (Junk et al. 2012; Prance 1979; Sioli 1954). These environments cover an area of about 275,000 km2, or between 13% and 27% of all wetlands in the basin, contributing significantly to the regional carbon balance and biodiversity (Junk et al. 2011; Melack and Hess 2010; Melack et al. 2009a). The Solimões/Amazon River várzeas are character- ized by an annual flood regime described as the “flood pulse” (Junk et al. 1989). Average maximum flooding depths can reach up to 16 m in Western Amazon, 10 m in Central Amazon, and 6 m in Eastern Amazon, and local flooding extent and duration depends on the interaction between precipitation, river discharge and topography (Bonnet et al. 2008; Junk et al. 1989; Lesack and Melack 1995; Ramalho et al. 2009). The flood pulse is the main ecological forcing in the floodplain, control- ling the occurrence and distribution of plants and animals, life-history traits, primary and secondary production, and also influencing carbon respiration, de- composition and nutrient cycles in water and soils (Junk 1997b). Together with geomorphological characteristics, the flood pulse is also directly related to ero- 1Published as FERREIRA-FERREIRA, Jefferson et al. Combining ALOS/PALSAR de- rived vegetation structure and inundation patterns to characterize major vegetation types in the Mamirauá Sustainable Development Reserve, Central Amazon floodplain, Brazil. Wet- lands Ecology and Management, v. 23, n. 1, p. 41–59, 2015. 5 CHAPTER 2. FLOOD AND VEGETATION MAPPING 6 sion, transport and deposition processes (Irion et al. 1997). Most floodplain environments have a flooding gradient from aquatic to terrestrial conditions, re- sulting in a complex mosaic of habitats (Junk 1997b). Hydrogeomorphological dynamics such as migrating channels and evolving lakes are also an important feature of the várzea landscape, influencing habitat characteristics and vegeta- tion distribution (Peixoto et al. 2009; Wittmann et al. 2004). Ducke and Black 1954, Rodrigues 1961, and Takeuchi 1962 identified dif- ferent habitat types, flooding regimes, nutrient availability and biogeographi- cal history as factors influencing the composition, distribution, and diversity of species. Prance 1979 offered the first classification of Amazonian wetland forests, based on hydrological and hydrochemical parameters, and Pires and Koury 1959 and Hueck 1966 described a zonation of plant communities along the flooding gradient in eastern and central Amazonn várzeas. Plants subject to waterlogging have a variety of evolutionary adaptation strategies for coping with the anaerobic soil conditions, and flooding is considered to be the major driver of local-scale habitat zonation, a selective force influencing evolutionary processes (Parolin and Wittmann 2010; Wittmann et al. 2010b). Junk 1989 reported associations between tree species and topographic heights, with inundations lasting 140 or less days per year, 140-230 days per year, and 230-270 days per year. Applying the nomenclature used by the local popula- tion, Ayres 1993 described different várzea forest types according to the mean inundation depth along the lower Japurá River. He described the chavascal as a vegetation community of dense shrubs with small trees occurring in areas where the water column depth ranged between 5 and 7 m; the low restinga as the veg- etation community of forest occupying low land where the seasonal maximum inundation depth ranged between 2.5 and 5 m; high restinga as the vegetation community of forest occupying higher land where water column depth ranged from 1 to 2.5 m. Wittmann et al. 2002 updated this classification, modifying the terminology to várzea alta (high várzea) and várzea baixa (low várzea), to avoid confusion with the term restinga as used for coastal vegetation in the Brazilian literature (Figure 2.1). Figure 2.1: Conceptual cross-section diagram showing the three main forest subtypes present in Mamirauá Sustainable Development Reserve (Central Amazon, Brazil). Adapted from Ayres 1993 Recently, Junk et al. 2012 have proposed a comprehensive classification of CHAPTER 2. FLOOD AND VEGETATION MAPPING 7 várzea habitats, based on a combination of hydrological, geomorphological and botanical characteristics. However, given the large extent and heterogeneity of várzea landscapes, little is currently known about the distribution, extent and relative proportion of each of these habitats within the Amazon Basin. Remote sensing methods have been successfully used to study vegetation cover and hydrologic dynamics in wetland environments, and recent advances have allowed the characterization and quantification of multiple wetland ecolog- ical processes (Costa et al. 2013; Ozesmi and Bauer 2002). A few contributions to the understanding of ecological and anthropogenic processes in várzea habi- tats have been derived from optical remote sensing studies (Mertes et al. 1995; Renó et al. 2011; Wittmann et al. 2002), but most advances have been based on the use of synthetic aperture radar (SAR) sensors, given their ability to detect flooding under plant canopies, and its capacity to acquire images even under cloudy conditions (Henderson and Lewis 2008; Kasischke et al. 1997). Early SAR studies in the floodplain were supported by the SIR-C/X-SAR mission (Hess et al. 1995), and the launch of the Japanese JERS-1 L-band orbital sen- sor fostered studies on flooding dynamics, habitat mapping, water level height, and biomass estimation (Alsdorf et al. 2007a; Martinez and Letoan 2007; Rosen- qvist et al. 2002). More recently, the combination of new processing methods such as object-based image analysis (OBIA) and imagery provided by the new crop of polarimetric SAR systems (ALOS/PALSAR, Radarsat-2, TerraSAR-X and Cosmo/Skymed) has allowed researchers to assess vegetation properties and ecological processes in the Amazon floodplain at the landscape scale (Arnesen et al. 2013; Hawes et al. 2012; Silva et al. 2013a). Still, most of these studies are limited either temporally or spatially, and the distribution and spatial configuration of várzea habitats remains poorly known, limiting management, monitoring, and conservation initiatives and leaving these areas open to anthropogenic degradation and overexploitation. In this sense, remote sensing monitoring may not only offer a valuable scientific tool, but also contribute directly to the identification of priority areas for protection and con- servation, as well as contributing to the proper management of these areas. For this reason, the present study demonstrates how L-band ALOS/ PALSAR im- agery can be utilized to (i) identify the distribution and relative proportion of different vegetation types, and (ii) produce landscape-scale estimates of flood- ing extent and duration, within the context of management and conservation needs of the Mamirauá Sustainable Development Reserve protected area, in the Central Amazon floodplain. 2.2 Methods 2.2.1 Study Area First of its kind, the Mamirauá Sustainable Development Reserve (MSDR) is located on a floodplain region at the confluence of the Solimões and Japurá Rivers, near the town of Tefé and approximately 600 km upstream from the city of Manaus, in the Central Amazon floodplain (Figure 2.2). Covering approximately 11,240 km2, the MSDR is one of the largest Brazil- ian protected areas dedicated to wetland environmental conservation, and one of the few functional protected areas in the Brazilian várzea forests (Queiroz CHAPTER 2. FLOOD AND VEGETATION MAPPING 8 Figure 2.2: Mamirauá Sustainable Development Reserve (MSDR) location, in the Central Amazon floodplain, Brazil. On the right, the southeastern portion of MSDR, located between the Solimões and Japurá rivers, and its adjacent wetlands. The underlying image is a temporal average from a set of 13 HH ALOS/PALSAR image mosaics. The wetlands mask shown was derived from Hess et al. 2003, geometrically corrected by Rennó et al. 2013 and manually edited by Ferreira et al. 2013. and Peralta 2006). Recognized as a World Heritage site by UNESCO (http: //whc.unesco.org/en/list/998), the MSDR is also currently the only RAM- SAR site representing Amazon wetlands (Secretariat 2013). Since 1992, active research programs and community-based management projects have been devel- oped in the reserve to understand the biology and conservation of IUCN listed species, such as the Amazonian manatee (Trichechus inunguis - vulnerable), http://whc.unesco.org/en/list/998 http://whc.unesco.org/en/list/998 CHAPTER 2. FLOOD AND VEGETATION MAPPING 9 the jaguar (Panthera onca - near threatened), the Black Caiman (Melanosuchus niger - lower risk/conservation dependent), and the white uakari monkey (Ca- cajao calvus - vulnerable), while promoting sustainable use and protecting the traditional livelihoods of riverine communities. Supporting a range of socio- economic and biological studies on forestry, agriculture, fisheries and ecotourism, the reserve is a key center for research on sustainable development and conser- vation in Amazonian environments (Queiroz and Peralta 2006; Schöngart and Queiroz 2010). Flooding dynamics at the MSDR are characterized by a large monomodal flood pulse, reaching about 10 m in amplitude (Ramalho et al. 2009). The high water phase (cheia) starts in May, extending to mid-July, followed by a receding water phase (vazante) that lasts until September. The low water phase (seca) occurs from September to November, when the rising water phase (enchente) phase starts, lasting from December to May (Ramalho et al. 2009). The main vegetation types observed in the reserve are the chavascal, low várzea, and high várzea, in addition to herbaceous vegetation stands. The chavascal name is given to poorly drained alluvial relicts developing in old depressions, abandoned channels, and shallow lakes, filled with large propor- tions of clay deposited during the aquatic phase and covered by a dense and species-poor shrub/tree community (Wittmann et al. 2010b). The flood dura- tion in these habitats is reported as lasting about 180-240 days per year, with water heights varying between 5 and 7 m. Individual density may exceed 600 individuals ha-1, with characteristic species such as Symmeria spp., Calyptran- thes multiflora, Eugenia ochrophloea, Buchenavia oxycarpa and Pseudobombax munguba (Wittmann et al. 2010b). Low and high várzea habitats are differentiated by floristic and structural features induced by the hydroperiod, sharing between them only 12% of the tree species (Wittmann et al. 2002). Due to the stronger flooding pressure, low várzea areas have the fewest and smallest species, and higher individual density. Early successional stages are usually formed by dense and often monospecific stands of Cecropia latiloba, which decrease hydrodynamic energy, induce sed- imentation, and provide the necessary shading to support the establishment of other species (Wittmann et al. 2010a,b). Late-successional stages usually contain 70-90 species per ha, such as Piranhea trifoliata, Tabebuia barbata, He- vea spp., Pouteria spp., Oxandra spp. and Duroia duckei. These forests toler- ate flood durations of 120-180 days every year, with a water level of 2.5-5 m (Wittmann et al. 2010b). On high várzea communities, population dynamics and canopy architecture are more complex, with higher biomass, species richness and diversity values than low várzea. In a survey conducted by Wittmann et al. 2002 in the same study area as ours, 177 species were found in a single 1 ha plot, where 101 species were represented by a single individual. These communities occur in the highest elevations, with a geomorphological context of relative stability, such as scrollbars and levees, and have a distinctive vertical stratification, with an upper canopy height of 30-35 m and emergent trees reaching heights of up to 45 m. Some representative species are Pouteria procera, Malouetia tamaquarina, Aspi- dosperma riedelii, Guatteriopsis paraensis, Gustavia augusta and Pseudoxandra polyphleba. Flood duration varies between 60 and 120 days per year, with maxi- mum depths of 1-2.5 m. Depending on their position along the flooding gradient, some of these forests may experience less than 50 flooding days per year, and not CHAPTER 2. FLOOD AND VEGETATION MAPPING 10 experience any flooding during exceptionally dry years (Schöngart et al. 2004). In addition to woody vegetation, areas sometimes referred to as várzea fields (campos de várzea) are composed of low lying areas and shallow lakes that alternate seasonally between free water surface, exposed sediments and herbaceous vegetation (macrophytes), occupying areas with the longest flood- ing durations. These communities comprise several aquatic or palustrine grass species (e.g. Echinochloa polystachya, Hymenachne amplexicaulis, Paspalum spp., Oryza spp.), as well as floating herbs such as Eichhornia spp., Pistia spp., Salvinia spp., Ludwigia spp., Neptunia spp., Nymphoides spp. and Victoria ama- zonica. Most of these plants have very high growth rates, and can rapidly occupy available substratum, showing significant seasonal and interannual variation in distribution and coverage (Silva et al. 2013b). 2.2.2 Remote sensing data and processing We acquired L-band (23.6 cm wavelength) SAR imagery from the Phased Array type L-band Synthetic Aperture Radar sensor on-board the Advanced Land Observation Satellite-1 (ALOS-1/PALSAR), operated by the Japanese Aerospace Exploration Agency (JAXA). Launched in January 2006 and decom- missioned in April 2011, this was the first SAR mission to provide a systematic global image acquisition strategy, and it has since 2014 been followed by the ALOS-2 mission (Rosenqvist et al. 2014, 2010). Imagery for this study was pro- vided by JAXA through the ALOS Kyoto and Carbon Initiative (K&C), later made freely available at different sources, e.g. NASA’s Earth Observing System Data and Information System (http://reverb.echo.nasa.gov/reverb/). Images were acquired in two polarization modes: fine beam single (FBS), with HH polarization, and fine beam dual (FBD), with HH and HV polariza- tions. The Fine Beam mode is characterized by a high resolution strip with a ≈70 km swath, a 38.7°of incidence angle (at the scene center) and 12.5 m pixel spacing (Rosenqvist et al. 2007). To better capture the flood pulse dynamics, a set of 26 scenes (Path 85, Frames 7120 and 7130) were acquired for several dates, chosen to provide the largest and most uniform range of water level conditions within the available imagery for the area. Water stage data was recorded at the Mamirauá Lake gauging station, located in the southern part of the study region (see Figure 2.2; IDSM 2013; Ramalho et al. 2009). For each date, the corresponding adjacent scene pairs were mosaicked to provide complete coverage of the study area, resulting in a final set of 13 images (Table 2.1). This acquisition pattern encompassed the southeastern portion of the MSDR, forming a seemingly triangular shape upstream of the confluence between the Japurá and Solimões rivers, where most management activities and research studies take place, and also encompassed adjacent wetlands contained within the selected scenes. The total area mapped comprises approximately 4,680 km2. All images were acquired at the 1.5 processing level, which includes range and azimuth compression, multilooking, slant to ground range conversion and radiometric and geometric corrections (Japan Space Systems 2012), and con- verted to linear backscattering coefficients (σ0) for statistical summarization. Final results were expressed in dB units, to allow comparisons with the previ- ous literature. For PALSAR Fine Beam products, the conversion to σ0 follows http://reverb.echo.nasa.gov/reverb/ CHAPTER 2. FLOOD AND VEGETATION MAPPING 11 Table 2.1: ALOS/PALSAR synthetic aperture radar images acquired at different dates and water stage levels, for mapping vegetation types and inundation extent within the Mamirauá Sustainable Development Reserve (Central Amazon, Brazil). FBS - fine beam single (HH polarization), FBD - fine beam dual (HH/HV polariza- tion). Water level heights were obtained from Mamirauá gauge station. (IDSM 2013; Ramalho et al. 2009). PALSAR acquisition date Mode Water stage height (masl) 2010-09-22 FBD 24.71 2007-10-30 FBS 27.00 2010-12-23 FBS 27.38 2008-12-17 FBS 30.02 2007-12-15 FBS 31.07 2010-03-22 FBS 32.72 2008-08-01 FBD 32.85 2007-07-30 FBD 33.37 2008-05-01 FBD 35.12 2010-05-07 FBD 35.65 2007-06-14 FBD 36.06 2010-06-22 FBD 36.28 2009-06-19 FBD 38.32 Equation 2.1: σ0 = 10× log10(DN2)− 83 (2.1) where DN is the backscattering amplitude expressed in digital numbers, and -83 is the calibration coefficient for PALSAR standard products (Shimada et al. 2009). In addition to PALSAR images, georeferenced multispectral sensor mosaics with 5 m spatial resolution of RapidEye (multiple dates between 2009 and 2011) and 2.5 m spatial resolution of SPOT-5 (acquisition (acquisition on 2012-11-08) images were utilized as aid for visual interpretation of the land cover classes in the study area. These images were acquired and provided by the Mamirauá Sustainable Development Institute, who manages the MSDR (www.mamiraua.o rg.br). 2.2.3 Image segmentation and classification Our study follows a similar approach to Silva et al. 2010 and Arnesen et al. 2013 to map land cover and inundation status in várzea environments, com- bining multitemporal SAR imagery and OBIA techniques. Standard image classification techniques work solely on a pixel-by-pixel basis, ignoring both the spatial context and the multi-scale information (texture) contained within the image elements, and are overly susceptible to SAR speckle. OBIA methods start by segmenting the image into homogeneous groups of pixels (objects), ideally www.mamiraua.org.br www.mamiraua.org.br CHAPTER 2. FLOOD AND VEGETATION MAPPING 12 corresponding to homogeneous land cover features, and allow the use of mul- tiple descriptive statistics and contextual information during the classification process (Blaschke 2010). Prior to image segmentation, temporal composite images were produced, fol- lowing Arnesen et al. 2013: temporal average backscattering (TAB), comprising the average backscattering of the entire image series; temporal standard devia- tion (TSD), comprising the per-pixel standard deviation for all observed values in the series, and lowest water level backscattering (LWB), simply defined as the scene with the lowest observed water stage level (2010-09-22). These sea- sonal descriptors allow the segmentation and classification to identify groups of pixels with similar time series of PALSAR backscatter coefficients. The mea- sures chosen enable vegetation communities to be defined as a combination of vegetation structure and inundation dynamics. These images were filtered us- ing three consecutive passes of a 3 x 3 Gamma filter (Shi and Fung 1994) and converted to an 8-bit radiometric scale, to reduce speckle heterogeneity and increase computational efficiency during segmentation. The TAB, TSD and LWB images were used as inputs for the multi-resolution segmentation algorithm implemented on eCognition 8.0 (Definiens 2009), to- gether with a vector file of the Amazon wetland mask produced by Hess et al. 2003, geometrically corrected by Rennó et al. 2013 and manually edited by Fer- reira et al. 2013. This is a region-merging algorithm that begins with a single pixel and a pairwise comparison with its neighbors, with the goal of minimizing the resulting calculated heterogeneity. After iterative testing, the parameters of scale = 150, shape = 0.1 and compactness = 0.5 were selected. After segmentation, the mean and standard deviation of σ0 were computed for each image object, separately for all 15 available layers (single date im- ages plus TAB and TSD), resulting in a total of 50 object attributes across all dates and polarizations available. For the TAB and TSD seasonal descriptors, the original unfiltered and unscaled images were used to ensure comparability with the single date imagery. Using vegetation type information from 86 survey plots provided by the Mamirauá Institute for Sustainable Development, and sup- ported by Rapid Eye, SPOT-5 and Google EarthTM high resolution imagery, 360 objects were selected as training samples (72 objects per class) for subsequent radiometric analysis and classification, based on a multi-sensor interpretation key (Figure 2.3). Five land cover classes recognized in the literature (Ayres 1993; Junk et al. 2012; Wittmann et al. 2002) were defined for evaluation: three main arboreal vegetation types (Chavascal, Low Várzea and High várzea), permanently free water surfaces (Water Bodies), and transient areas that alternate seasonally between water, substratum and herbaceous vegetation (Herbaceous/Soil). Non- floodable uplands (terra firme) areas were excluded using the wetland mask from Hess et al. 2003, and not further evaluated. After sample selection, the temporal radiometric response of each class was graphically analyzed using boxplots. To discriminate the defined classes, we used the random forests (RF) al- gorithm, proposed by (Breiman 2001), as implemented in the randomForest package of the R open source statistical programming environment (Liaw and Wiener 2002). A vector file containing all image objects identified as training samples, with the associated attribute table containing class labels and sampled backscattering responses for all images in the series, was submitted as input to the RF algorithm, to derive the classification tree ensemble. This ensemble was CHAPTER 2. FLOOD AND VEGETATION MAPPING 13 Figure 2.3: Visual interpretation key used to select training and validation samples for vegetation mapping at the Mamirauá Sustainable Development Reserve (Central Amazon, Brazil). TAB temporal average backscattering, TSD temporal standard de- viation of backscattering, TAB and TSD were calculated using all available ALOS/ PALSAR images of the 2007-2010 time series. then applied to the entire set of generated image objects, to produce the final classification. The RF algorithm is an ensemble learning method based on classification and regression trees (CART), where instead of a single decision tree, a “forest” (i.e., ensemble) of individual trees is built through randomization of the training data. Final class predictions are based on using a majority voting scheme (consensus) among the trees in the ensemble, improving predictive accuracy. Independent trees are constructed using a bootstrap sample of the data set in a process called “bagging”. In each bootstrap sample, approximately one-third of the reference data are left out. At the end of each iteration, the “out-of-bag” samples are then predicted using the ensemble derived from the bootstrap sample, and later aggregated to produce an out-of-bag (OOB) estimate of classification error for the entire “forest”. The two main parameters of the classifier are the number of independent trees generated (ntree) and the number of predictive variables that are randomly selected for choosing the best split (mtry). Multiple combinations of the two main parameters were tested until an optimal set of parameters was found. The accuracy of the vegetation map was assessed using 142 validation points, randomly distributed within the study area using a Geographic Information System software. These points were manually classified based on the prede- fined interpretation key, using available high-resolution imagery. The resulting manual classification was compared to the RF classification to build a confusion matrix and derive overall accuracy, class accuracy, kappa statistics (Congalton 1991) and quantity and allocation disagreement measures (Pontius and Millones 2011). Quantity disagreement refers to the difference in area proportions be- CHAPTER 2. FLOOD AND VEGETATION MAPPING 14 tween the reference data (training samples derived from field plots) and the classification. Allocation disagreement, on the other hand, is the proportion of misplaced objects from the classified map in comparison with positions in the reference data. A comparison between a reference map with two classes and a classification where every point is misclassified as the opposite class, for instance, would have 0% quantity disagreement and 100% allocation disagree- ment. Although Pontius and Millones 2011 condemn the use of the kappa index of agreement, we include it here to allow comparisons with previous literature. 2.2.4 Flood extent mapping Flood extent maps were generated for all single date images based on the expected increase in SAR signals due to enhanced double-bounce scattering, where the radar beam is specularly reflected by the free water surface under the canopy, and then scattered back to the sensor by the standing vegetation, or vice versa (Hess et al. 1995; Silva et al. 2008). Flooded area for woody vegetation was determined based on simple thresholds, determined by the graphical analysis of backscattering values in each PALSAR scene. Once flooded area was determined for each image in the time series, each flood map was associated to a corresponding water level, according to Table 2.2. Table 2.2: Water stage for each map of inundation extent, and corresponding flood duration classes, in days, for the Mamirauá Sustainable Development Reserve (Central Amazon, Brazil). Water stage height (masl) Flood duration (days) 27.00 > 295 30.02 175-295 31.07 175-295 32.72 125-175 33.37 105-125 35.12 40-105 35.65 < 40 36.06 < 40 38.32 < 40 Water levels considered too similar in terms of flood extent were grouped into a single class, and the image acquired on 2010-09-22 was excluded from the analysis, since it had a very low water level and negligible flood extent outside of permanent water bodies. This resulted in nine different inundation extent maps, corresponding to each water level (27.0, 30.72, 31.07, 32.72, 33.73, 35.12, 35.65, 36.06 and 38.32 m). Average flood duration was estimated by taking the average stage height for all available data from the Mamirauá gauge (1991-2011; meters above sea level), and determining the number of days per year where this average was equal or above the observed stage height at the moment of image acquisition. Flood duration categories were then established by taking this duration and extending it backwards until the previously observed duration category (see example on Figure 2.4). CHAPTER 2. FLOOD AND VEGETATION MAPPING 15 Day of the Year Rising High DescendingLow Low 126 days 175 days 2010-03-22 2008-12-17 Figure 2.4: Mean water stage and shaded 95% confidence interval for the 1991-2011 period, measured at the Mamirauá Lake gauge, Mamirauá Sustainable Development Reserve (Central Amazon, Brazil) (IDSM 2013; Ramalho et al. 2009). The ordering of the Julian dates is offset, starting at the lowest mean water level to indicate the average onset date for the rising water period. The four stages of the flood pulse are indicated as low, rising, high and descending. The dashed line indicates the abnormally high water levels for the 2008-2009 hydrological year, when iButton inundation data was recorded (Affonso et al. 2011; Hess et al. 2011). Horizontal lines show the stage levels for two hydrologically consecutive image acquisition dates. According to our criteria, flood duration for the 2010-03-22 image is assumed to be between 175 and 125 days Stage heights with very similar flood durations were again grouped into the same category, resulting in six flood duration classes. The single final map was derived by successively overlaying inundation maps of consecutive stage heights, and labelling all pairwise non-overlapping mapped areas as flooded between the levels observed for the first and second maps. Finally, all mapped Herbaceous/Soil areas were added to the map as belonging to the “> 295 flooding days” class, and the mapped Water Bodies class was appended as a “365 flooding days” class. Flood mapping validation was performed using temperature-based inunda- tion data from Affonso et al. 2011; Hess et al. 2011, based on thermistor chains (Thermocron iButtons). These authors have recovered inundation data from 18 sites within the focal research area of the MSDR, for the 2008-2009 hydro- logi- cal year. The location of each site was identified on the flood duration map, and the in situ duration of inundation was then compared to the estimated duration derived from the PALSAR time series. Once both maps were validated, the co-occurrence of vegetation and inun- dation classes was determined by an overlay of both maps, quantifying the total area for each combination of vegetation type and flood duration. CHAPTER 2. FLOOD AND VEGETATION MAPPING 16 2.3 Results The best parameterization of the RF algorithm consisted of an ensemble of ntree = 5000 decision trees, with mtry = 20 out of the 50 available predictor variables as candidates for a split. This resulted in an overall estimated OOB error of 10.6%. The highest prediction errors were observed for Chavascal, with a 20% prediction error composed mainly of misclassification with low várzea (12.5%) and high várzea (5.5%), and for low várzea with 13% prediction error, equally distributed between Chavascal and High Várzea. The remaining classes had prediction errors of 8% (Herbaceous/Soil), 6% (high várzea) and 2% (water bodies). The resulting classification revealed a dominance of low várzea environments, and an overall complex mosaic of habitats resulting from the dynamic hydro- geomorphological characteristics of the area (Figure 2.5). Vegetative cover was distributed as 1,753 km2 (37.7%) of low várzea, 873 km2 (18.7%) of high várzea, 832 km2 (18%) of Chavascal, 711 km2 (15.3%) of Water Bodies, and 480 km2 (10.3%) of herbaceous/soil. Independent validation of the vegetation cover clas- sification based on comparison with high-resolution optical imagery yielded an overall accuracy of 83%, with a kappa index of agreement of 0.8 (Table 2.3). The worst results were again observed for the Chavascal and Low várzea classes, while the Herbaceous/Soil class had the largest spread of misclassification. Table 2.3: Confusion matrix and accuracy indices for the classification of vegeta- tion types using ALOS/PALSAR image time series and the random forests classifica- tion algorithm, for the Mamirauá Sustainable Development Reserve (Central Amazon, Brazil). Water Bodies Várzea Fields High Várzea Low Várzea Chavascal Water Bodies 24 3 0 0 0 Várzea Fields 0 28 0 0 0 High Várzea 0 1 21 6 0 Low Várzea 0 3 1 18 7 Chavascal 0 0 1 2 27 N 24 35 23 26 34 % Error 0% 20% 8.7% 30.7% 20.5% Overall accuracy 83% Kappa: 0.8 Quantity disagreement 5% Allocation disagreement 10% Overall disagreement rates were of 5% for quantity and 10% for allocation. The highest allocation disagreement was observed for the Low várzea class, with 9% of misplaced objects, followed by Chavascal, with 8% and High várzea with 4%. Both Herbaceous/Soil and Water Bodies did not have a significant CHAPTER 2. FLOOD AND VEGETATION MAPPING 17 Figure 2.5: Major vegetation types and habitats of the southeastern portion of the Mamirauá Sustainable Development Reserve (Central Amazon, Brazil) and its sur- roundings, mapped using ALOS/PALSAR image time series, and the random forests classification algorithm. amount of misplaced objects. Quantity disagreement indicated underestimation of Herbaceous/Soil (4%), resulting from confusion with Water Bodies, with negligible errors for the remaining classes (1-2%). Graphical analysis of backscattering values showed the effect of flooding on the radar signal, emphasizing the different patterns of radar backscattering evolution that reflect different flood regimes for chavascal, low várzea and high várzea environments (Figure 2.6). The final image-specific thresholds selected CHAPTER 2. FLOOD AND VEGETATION MAPPING 18 for inundation mapping varied between -6.57 and -5.85 dB, resulting in estimated flooding durations between less than 40 days and more than 295 days, in addition to the permanently flooded areas (Figure 2.8, Table 2.2). Overall, most areas were inundated for less than 40 days or for 125 to 175 days (Figure 2.7). Figure 2.6: Temporal variation of ALOS/PALSAR backscattering coefficients for the three main woody vegetation classes occurring on the Mamirauá Sustainable Devel- opment Reserve (Central Amazon, Brazil). The arrows indicate the period where flooding begins. The different flooding onset and flooding patterns for each class are visible The agreement between estimated flood durations and ground data from Affonso et al. 2011 was variable; from the 18 records, eight corresponded to the actual range of estimated flood duration, nine were off by one estimated class, and one was off by two classes (Table 2.4). The intersection of vegetation types and flood duration classes showed that chavascal areas had the most varied inundation pattern, covering the entire range of estimated flood duration classes, with a higher frequency (49%) in the range of 105-125 days of flooding per year (Figure 2.9). Low várzea areas, on the other hand, occurred predominantly on the class labeled as 175-295 days of flooding, followed by the 125-175 days of flooding class. High várzea agreed more closely with the expected distribution across the flood duration classes, with the highest frequency being observed at less than 40 days of flooding per year. Moreover, about 181 km2 of this class occurred in areas that were never mapped as flooded, considering the existing range of images (and which were added to the “<40 days of flooding class”). CHAPTER 2. FLOOD AND VEGETATION MAPPING 19 Figure 2.7: Extent of estimated flood duration classes (in days), based on ALOS/PALSAR imagery, for the Mamirauá Sustainable Development Reserve (Cen- tral Amazon, Brazil). Table 2.4: Comparison of flood duration estimates as determined in situ using tem- perature gauges by Affonso et al. 2011 and Hess et al. 2011 and as derived from ALOS/PALSAR images for the Mamirauá Sustainable Development Reserve (Central Amazon, Brazil). Station Flood duration (days) Flood duration PALSAR (days) 1 176 125-175 2 233 105-125 3 231 >295 4 217 125-175 5 241 >295 6 291 >295 7 219 175-295 8 237 175-295 9 239 175-295 11 238 175-295 13 240 >295 14 241 125-175 18 239 >295 19 176 175-295 20 238 125-175 21 244 175-295 22 249 >295 23 243 >295 CHAPTER 2. FLOOD AND VEGETATION MAPPING 20 Figure 2.8: Estimated flood duration based on a time series of ALOS/PALSAR image data, for the southeastern portion of the Mamirauá Sustainable Development Reserve and surroundings, Central Amazon floodplain, Brazil. CHAPTER 2. FLOOD AND VEGETATION MAPPING 21 Figure 2.9: Relative area for each combination of land cover and flood duration classes derived from ALOS/PALSAR image time series for the Mamirauá Sustainable Development Reserve (Central Amazon, Brazil). CHAPTER 2. FLOOD AND VEGETATION MAPPING 22 2.4 Discussion The RF algorithm was a robust classification technique for várzea vegetation, when coupled with reliable and sufficient ground data and with OBIA methods. The availability of multitemporal information was paramount to obtain accu- rate class discrimination, as already shown by Martinez and Letoan 2007 and Silva et al. 2010. Due to their higher structural similarity, woody vegetation classes tended to share most of the misclassification errors, which may have led to the overestimation and erroneous allocation of Low várzea objects, com- pared to Chavascal and High várzea. While chavascal areas tend to form more homogeneous and densely packed stands and high várzea forests will often cor- respond to complex, relatively stable vegetation assemblies, low várzea regions can display a wide range of community composition and structural characteris- tics, depending on relative age and position along the flooding gradient. This variability was translated into well- defined class attributes for High várzea and Chavascal, while the larger variability of Low várzea samples increased classifi- cation error. L-band SAR data also led to confusion between Herbaceous/soil and water bodies, due the relative similar and smooth surface of these targets at this wavelength. The complementary use of shorter SAR wavelengths, such as C or X bands, could lower this error. The dynamic geomorphological nature of the study area explains the natu- rally fragmented landscape mosaic evidenced by our mapping. In this context, high várzea forests can be seen as the narrowest and most disjoint landscape elements, while chavascal formations tend to form more aggregated, continuous fields. Noteworthy spatial associations were also observed, such as the interre- lated distribution of water channels and flood duration, versus chavascal and high várzea distributions. As the chavascal occurs in poorly drained depres- sions or silted-up lakes, it tends to occupy the backswamps behind the levees covered by high várzea. Eventually, the establishment of pioneer vegetation will increase sediment deposition and raise the terrain level, reducing flood duration and allowing the establishment of other species, in a process of ecological and geomorphological co-evolution (Wittmann et al. 2010b). Given the well-known effect of landscape configuration on the conservation of plant and animal species (Lindenmayer et al. 2008), our detailed vegetation map can enable spatially aware decision making for conservation measures in the MSDR. Virtually all of the study area was flooded when water levels were close to the peak of the high water phase. Results also showed that 22% of the evaluated areas were classified as “< 40 days of flooding”, followed by areas flooded for 125-175 days per year representing about 12% of the mapped area. The uneven distribution of flooded areas at different water levels results from the stepped nature of the floodplain terrain, where critical water stage heights result in large expansions of inundation area while other heights in the range have minimum effects. The best example of such as process is the water stage height at which transition from channel to overbank flow occurs, immediately inundating the backswamp depressions. Although the comparison between flood duration and in situ temperature- based observations showed some disagreement, we believe that the maps ob- tained do represent the overall spatial distribution and variability of flooding in the studied system, and are similar to other products developed for flood- plain forest environments in the Amazon (Forsberg et al. 2001; Rosenqvist et al. CHAPTER 2. FLOOD AND VEGETATION MAPPING 23 2002). Three main sources of error can account for the observed disagreements: positional errors between in situ stations and image data; inherent variability of flood duration, as shown by the confidence intervals on Figure 2.3; and the fact that field data was acquired during the 2008-2009 hydrological year, the second largest flood of the last 50 years (Figure 2.4), whereas our imagery spans a broader time period. In response to our results, the Mamirauá Sustainable Development Institute is currently funding the installation of high precision level gauges, tied to surveyed altimetric transects and with accurate geoloca- tion, to properly characterize flooding dynamics in the region. Once such data is available, the current results can be further validated and refined. The intersection between vegetation and flood duration classes showed a wider range of combinations than expected based on the literature. Chavascal areas had shorter inundation periods than the usually recognized hydroperiod of 180-240 days of flooding (e.g. Ayres 1993; Wittmann et al. 2002), while low várzea was distributed between flood duration ranges that were higher than reported by the literature (120-180 days). This apparent inversion of results is likely owed to the higher misclassification errors between these two classes, implying that forests with a highly variable range of structural and taxonomic characteristics are distributed within the range of approximately 50-200 days of flooding. The graphical analysis of training samples based on actual ground data (Figure 2.6) suggests that inundation occurs at similar times for both classes, further adding to the differentiation problem. While some of the unusual combinations of vegetation and flooding observed likely occurred due to classification errors, these results suggest that such com- bined information can be a good indicator of the complex gradient of habitats along the floodplain, including the identification of rare habitats. Further ver- ification of these locations in the field could therefore suggest potential areas for special conservation measures, given their relative rarity in the landscape. For example, shrub-like vegetation occurring in areas flooded for short periods (mislabeled as chavascal) could indicate prevailing soil properties, such as high clay content and/or high phreatic levels, while forest communities growing at sites that were never mapped as flooded could indicate areas that only flood during extreme hydrological events, for short periods. As only 31% of várzea tree species are shared with upland forests, of which 67.5% are restricted to high várzea (Wittmann et al. 2006a), these areas could house species or assemblages that are currently rare in the landscape, but have the potential to become more prevalent under current scenarios of longer dry periods and more frequent ex- treme climatic events predicted for Amazonian environments (Malhi et al. 2008; Melack and Coe 2013). These areas can therefore have an important conserva- tion role as vegetation refugia for maintaining current and future diversity in the floodplain (Ashcroft et al. 2009). 2.4.1 Management implications The Amazonian várzeas are endangered ecosystems that require special pro- tection initiatives (Junk et al. 2011). The composition and abundance of various components of the fauna are also associated to várzea environments and its veg- etation types (Beja et al. 2010; Paim et al. 2013; Pereira et al. 2009). Detailed knowledge of the distribution and abundance of rare, endemic or threatened species in these environments is needed to define sensitive areas or areas that CHAPTER 2. FLOOD AND VEGETATION MAPPING 24 should be addressed with additional efforts or special protection. The distribu- tion and abundance of different endangered species, such as jaguars (P. onca), giant otters (Pteronura brasiliensis) and some primates such as the black squir- rel monkey (Saimiri vanzolinii) or white uacari (C. calvus) are also associated with dominant types of forest formation, and the local regime of flooding (Da Silveira et al. 2010; Lima et al. 2012; Paim et al. 2013). The flood predictability in the várzea environments is a key factor for a wide range of conservation initiatives and extractive activities of the local population that use the resources of these ecosystems. Large scale predictability already includes flood pulse intensity for some places of the Amazon, such as Manaus and Tefé (Schöngart and Junk 2007), but local scale information on flood dy- namics and extent remains unavailable. Such information would allow adequate sustainable use of innumerous natural resources from the várzea ecosystems; for example, access to remote sites within the forest for extractive activities depends on previous information about flood dynamics at these sites. To plan timber exploitation on várzea forests, where there are no roads to transport the logs, local-scale flood predictability is crucial to allow wood transport by raft- ing during high-water periods (Schöngart and Queiroz 2010). Access to lakes and channels of exceptional productivity for Pirarucu (Arapaima gigas) fishing (Viana et al. 2007) and for caiman catching (Botero-Arias et al. 2009), both of which are forms of sustainable resource management in the floodplains at different stages of development, is also related to the predictability of flood- ing and hydrological dynamics in these locations. Therefore,providing habitat and flooding maps for the Amazon floodplain can significantly improve the effi- ciency in developing and managing conservation actions targeted towards these ecosystems. 2.4.2 Conclusion Our results emphasize the potential contribution of SAR remote sensing to the monitoring and management of wetland environments, providing not only accurate information on spatial landscape configuration and vegetation distri- bution, but also important insights on the ecohydrological processes that ulti- mately determine this distribution. SAR systems are unique in their ability to map both vegetation distribution and flooding extent, and the combination of the two, together with a multitemporal approach, offers unique insight into the functioning of wetland ecosystems. Information derived from the present study also provides a solid basis for the study of plant and animal species distribution and habitat use, as well as an understanding of spatial variability of biogeochem- ical processes, and may ultimately support ecosystem modeling efforts and the forecasting of different ecological scenarios. It also provides an ideal database for testing the spatial implications of the “flood pulse concept” (Junk et al. 1989), a general theory of floodplain ecosystems which relates flood durations and other hydrological characteristics to the distribution and dynamics of aquatic flora and fauna. We believe our method could be successfully replicated for other seasonal wetland environments, using different kinds of SAR and optical image time series and available open-source remote sensing and statistical software. Given the rising availability of SAR sensors operating at multiple frequencies and spatial configurations, with a plethora of new systems planned or already scheduled for launch in the following years, multitemporal SAR studies could CHAPTER 2. FLOOD AND VEGETATION MAPPING 25 become an affordable and reliable method for wetland ecological monitoring in the Amazon. Chapter 3 MODELING THE SPATIAL AND TEMPORAL DYNAMICS OF INUNDATION IN AMAZONIAN FLOODPLAINS USING RADAR REMOTE SENSING AND LOGISTIC MODEL 3.1 Introduction Flood regime and geomorphology are the prime factors modulating ecosys- tem structure and function in floodplain systems (Junk et al. 1989; Lewis et al. 2000). The broader paradigm regarding the function of flooding in river- floodplain systems is given by the flood pulse concept (Junk et al. 1989), which postulates that amplitude, duration, frequency and periodicity (predictability) of flood pulses are the major factor maintaining the dynamic environmental equilibrium in floodplains. Major biogeochemical processes and the exchange of nutrients and organisms between different floodplain habitats are modulated by flooding regime (Jardine et al. 2015; Junk 1997a; Junk et al. 1989; Lake et al. 2006; Melack and Fors- berg 2001; Mitsch and Gosselink 2015; Mitsch et al. 2010; Richey et al. 2002). The seasonal inundation regime also provides environmental connectivity that is critical for keeping long-term gamma diversity (Thomaz et al. 2007; Ward et al. 2002). In the community level of large river floodplains, birds and fishes have more stable communities in environments with rhythmic annual floods (Jardine et al. 2015; Luz-Agostinho et al. 2009). The occurrence, distribution, diversity, and densities of plant species is known to be strongly influenced by flood du- ration (Junk 1989; Luize et al. 2015; Silva et al. 2013b; Wittmann et al. 2002, 2006a). Furthermore, for many plant species living in floodplains, gaseous ex- change rates are controlled by inundation, as it acts as a phenological trigger (Hawes and Peres 2016; Parolin et al. 2010). In the Amazon, the amplitude of seasonal water level variations can reach 26 CHAPTER 3. SPATIO-TEMPORAL FLOOD MODELING 27 up to 16 m in Western Amazon, 10 m in Central Amazon, and 6 m in Eastern Amazonia, with local flood extent and duration depending on the interaction among precipitation, river discharge and geomorphology (Bonnet et al. 2008; Junk 1989; Lesack and Melack 1995; Ramalho et al. 2009). As ecological and biogeochemical processes in these systems are closely linked to climate and hy- drology, they are the first to experience impacts when such conditions change (Mitsch et al. 2010). Results from numerical simulations to the end of the 21st century (2070-2099) predict increased mean (+9%) and maximum (+18.3%) in- undation extent over Peruvian floodplains and the Amazon River in the western Amazon, while decreased river discharges (mainly in the dry season) are pre- dicted for the eastern Amazon, reducing inundation extent during the low water season in the central (-15.9%) and eastern Amazon river (-4.4%) (Sorribas et al. 2016). Guimberteau et al. 2017 included deforestation scenarios when modeling hy- drological changes caused by climate change. While overall results agree with the contrasting regional changes above, the authors show that the expected de- creases in river discharge for some catchments could be attenuated by increased runoff caused by deforestation, with a general consistent increase of 2.2% in runoff under the worst deforestation scenario (34% of forest loss). Climate- deforestation feedbacks in the intensely deforested southeastern Amazon basin have already caused significant hydrological changes in east-southern sub-basins (Coe et al. 2009; Coe et al. 2013). Recent studies investigating hydrological changes in the Amazon basin have also shown that annual mean precipitation (from 1990 to 2015), wet season precipitation (from 1980 to 2015), and maximum river runoff (from 1980 to 2015) have increased, while dry season precipitation and minimum runoff have slightly decreased (Gloor et al. 2015). These changes have led to an increase in the frequency of extreme floods and “drier than usual” dry seasons in the last three decades (Gloor et al. 2015, 2013), agreeing with the expected increase in frequency and intensity of extreme events for the Amazon under climate change (Cook et al. 2012; Cox et al. 2008; Joetzjer et al. 2013; Marengo et al. 2012). The lack of appropriate monitoring and management capacities for the large area comprising Amazonian floodplains makes them even more susceptible to changes in hydrological conditions (Castello et al. 2013). Throughout the Ama- zon basin, areas complying with the international criteria for wetland definition comprise between 14% and 30% of the lowland Amazon basin (Junk et al. 2011; UNEP/CBD 2003). Floodplains represent between 25 and 50% of Amazonian wetlands, or about 519,100 km2 (Junk et al. 2011). This large uncertainty in wetlands and floodplains representativity results from methodological differ- ences in the quantification of total wetland area (Hess et al. 2015; Junk et al. 2011; Melack and Hess 2010), mainly depending on whether small floodplains along the dense network of low-order rivers and streams are considered or not. These areas are difficult to detect either from optical or radar remote sensing, but according to Junk 1993 could reach one million squared kilometers, and were considered in the estimate by Junk et al. 2011. One of the main current anthropogenic threats looming over Amazonian floodplain ecosystems is the construction of dams to satisfy energy demands. Brazilian Amazonian rivers have less hydroelectric plants than other biomes and incur in less compensation costs from permanent inundation of local com- munities and private lands, so that most of the planned hydropower plants for CHAPTER 3. SPATIO-TEMPORAL FLOOD MODELING 28 Brazil are focusing on these rivers (Lees et al. 2016). According to Lees et al. 2016, Brazil will be the most impacted Amazonian country, with 143 dams already operational or under-construction, and 254 planned dams. Recently, Anderson et al. 2018 showed that Marañón and Ucayali river basins already lost about 20% of tributary network connectivity due to hydropower dams and that planned hydropower plants could decrease river connectivity up to 50%, isolating river reaches and imposing serious threatens to freshwater ecosystems. Analyzing the potential impacts of six planned hydropower dams on Andean forelands, Forsberg et al. 2017 estimated a total reduction of 64% in basin-wide sediment supply, 97 and 83 % in phosphorus and nitrogen supply and impacts to fish yeld, greenhouse gas emissions and flood pulse dynamics downstream. Given these threats, and the critical role of the flood pulse as a driver of environmental dynamics in Amazonian floodplains, it is critical that we improve our understanding of the spatial and temporal behavior of flooding at finer scales, and its impacts on ecological systems. Important efforts have been made using hydrological models (Rudorff et al. 2014a,b; Trigg et al. 2009; Wilson et al. 2007). Although these mechanistic approaches offer many advantages in represent high-fidelity flood dynamics, accurate modeling largely depends on high quality terrain models, still unavailable for most of the Amazon basin. Synthetic aperture radar (SAR) remote sensing has been used as an al- ternative to characterize landscape-level inundation extent and behavior over Amazonian floodplains (Alsdorf 2004; Alsdorf et al. 2007b; Arnesen et al. 2013; Chapman et al. 2015; Hawes et al. 2012; Martinez and Letoan 2007; Rosenqvist et al. 2002). Most of these studies, however, are limited by the temporal reso- lution of SAR satellites and/or by the availability of regular time-series in their ability to provide detailed spatially-explicit estimates of flood duration, the key hydrological variable for ecosystems response and adaptation. Ferreira-Ferreira et al. 2015 was one of the first studies to combine SAR imagery with historical records of water level to estimate not only flooding extent, but also flood duration. However, the lack of regular and continuous time series of SAR data limited the predictions to semi-quantitative estimates of flood duration classes. To date, fully quantitative and spatially explicit estimates of flood duration and dynamics based on remote sensing are still lacking. Thus, the present study addresses this issue developing a method to spatially estimate flood duration using a logistic model. 3.2 Methods 3.2.1 Remote sensing data and flood mapping The flood mapping performed in the Section 2.2.4 (Figure 3.1) was used in this study and briefly summarized here. We acquired a set of 26 ALOS- 1/PALSAR scenes (Path 85, Frames 7120 and 7130), and mosaicked consecutive frames of the same date, resulting in 13 images/dates. These dates were chosen to better capture the flood pulse dynamics, providing the largest and most uniform coverage of water level conditions within the available imagery (Table 3.1). The acquired images captured a stage height range from 24.71 to 38.32 meter above sea level (masl), as recorded in the Mamirauá lake gauging station, or about 70% of the historical maximum amplitude for this gauge. CHAPTER 3. SPATIO-TEMPORAL FLOOD MODELING 29 55°W60°W65°W70°W75°W 0° 5°S 10°S #0 #* #*#* #* #* #* #* #* #* #* Level 10 Level 5Level 3 Level 7 Level 2Level 1 Level 9 Level 6 Level 8 Level 4 64°40'W64°50'W65°0'W65°10'W65°20'W 2°1 5'S 2°3 0'S 2°4 5'S 3°0 'S 3°1 5'S 0 10 205 Km 40°W60°W80°W 0° 10 °S 20 °S 30 °S ¯ 0 400 800200 Km MSDR Hydrography Brazil Amazonas State Tefé Manaus Mamirauá Sustainable Development Reserve (MSDR) Water Bodies #* Leveloggers #0 Mamirauá Gauging Station MSDR 27 30.72 31.07 32.72 33.73 35.12 35.65 36.06 38.32 Flood extents for each water stage Figure 3.1: Southeastern portion of Mamirauá Sustainable Development Reserve (MSDR), in the Central Amazon floodplain, Brazil. The underlying map is the flood extent for each water stage recorded in Mamirauá gauging station for the dates of ALOS-1/PALSAR imagery used (see Chapter 2) . CHAPTER 3. SPATIO-TEMPORAL FLOOD MODELING 30 Table 3.1: ALOS-1/PALSAR images acquired for different water levels to map flood extent in the Mamirauá Sustainable Development Reserve (Central Ama- zon, Brazil). The range of captured water stages comprises 70% of the maximum historical amplitude for this location. PALSAR acquisition date Mode Water stage height (masl) 2010-09-22 FBD 24.71 2007-10-30 FBS 27.00 2010-12-23 FBS 27.38 2008-12-17 FBS 30.02 2007-12-15 FBS 31.07 2010-03-22 FBS 32.72 2008-08-01 FBD 32.85 2007-07-30 FBD 33.37 2008-05-01 FBD 35.12 2010-05-07 FBD 35.65 2007-06-14 FBD 36.06 2010-06-22 FBD 36.28 2009-06-19 FBD 38.32 FBS - fine beam single (HH polarization), FBD - fine beam dual (HH/HV polarization). Water level heights were obtained from Mamirauá gauge station. (IDSM 2013; Ramalho et al. 2009). Water stage heights were recorded at the Mamirauá Lake gauging station, located in the southern portion of the study area (see Figure 3.1; IDSM 2013; Ramalho et al. 2009) and paired with image acquisition dates to determine the expected water levels at the time of acquisition of each image. This data is freely available at http://mamiraua.org.br/pt-br/pesquisa-e-monitora mento/monitoramento/fluviometrico/. Based on backscattering thresholds, Ferreira-Ferreira et al. 2015 mapped flooded areas in each of the 13 images. As some water levels were considered too similar, some flood maps were disregarded resulting in flood extent maps for nine water stages: 27.00, 30.72, 31.07, 32.72, 33.73, 35.12, 35.65, 36.06 and 38.32 masl (see Figure 3.1). To support the empirical inundation modeling, we obtained the Shuttle Radar Topography Mission Digital Elevation Model (SRTM-DEM) version 3.0 with 1 arc second spatial resolution (about 30 meters; Available at http://re verb.echo.nasa.gov/reverb/). From SRTM-DEM we derived , (1) the ter- rain Height Above the nearest drainage (HAND), (2) euclidean distance from the nearest drainage (EDND), (3) slope (SLP), (4) terrain curvature (TC), (5) profile curvature (PC), (6) planform curvature (PLANC) and (7) accumulated precipitation 15 days prior to each image acquisition date (PCP). HAND is a locally and hydrologically coherent terrain descriptor that nor- malizes the topography in respect to the drainage network, applying two se- quential procedures on a DEM. First, the algorithm creates a hydrologically coherent DEM, by defining flow paths and delineating channels and lake basins. http://mamiraua.org.br/pt-br/pesquisa-e-monitoramento/monitoramento/fluviometrico/ http://mamiraua.org.br/pt-br/pesquisa-e-monitoramento/monitoramento/fluviometrico/ http://reverb.echo.nasa.gov/reverb/ http://reverb.echo.nasa.gov/reverb/ CHAPTER 3. SPATIO-TEMPORAL FLOOD MODELING 31 Then, it uses local flow directions and the estimated drainage network to de- rive the nearest drainage map, which will then guide the HAND algorithm in calculating the normalized topography (Nobre et al. 2011; Rennó et al. 2008). The HAND algorithm is implemented in the TerraView GIS Software (http: //www.dpi.inpe.br/terraview/index.php). The resulting HAND model was resampled using nearest neighbor assignment to 12.5 meters to match the flood- ing map resolutions. EDND was determined using the “gdal_proximity.py” python routine imple- mented in Geospatial Data Abstraction Library (GDAL; http://www.gdal.o rg/gdal_proximity.html). The output cell size was also set to 12.5 meters. SLP, TC, PC and PLANC were calculated using algorithms available in ArcMap 10.3, an implementation of the methods by Zevenbergen and Thorne 1987 and whose detailed procedures are described in the official documentation, available at http://desktop.arcgis.com/en/arcmap/10.3/tools/spatial- analyst-toolbox/an-overview-of-the-surface-tools.htm. SLP was pro- cessed to express angles in degrees. TC is the second derivative of the vertical dimension of the raster surface with respect to slope (i.e. “the slope of the slope”). A positive cell value indicates the surface is upwardly convex at that cell, while a negative curvature indicates the surface is upwardly concave. A flat surface is assigned a value of 0. PC is the terrain’s curvature in the direction of maximum slope - i.e. the rate of change in slope - which affects the acceleration and deceleration of water flow. Finally, PLANC is the curvature of the land surface in the perpendicular direction of the slope, influencing flow convergence and divergence. Both TC, PC and PLANC units are in hundredths (1/100) of the vertical unit (meters). Reasonable expected values for all three output variables for an area of moderate relief can vary from -0.5 to 0.5. We also used daily precipitation estimates from the 3B42-v7 product of TRMM (Tropical Rainfall Measuring Mission), at 0.25° spatial resolution, down- loaded from http://mirador.gsfc.nasa.gov/. From this dataset, we ex- tracted the accumulated precipitation 15 days prior to the acquisition date of each SAR image used to map flood extent (see Table 3.1), resulting in nine accumulated precipitation observations to be included in our model. 3.2.2 Flood monitoring We monitored in-situ flood dynamics to calibrate the flooding model, us- ing ten pressure transducers (“leveloggers”, Level Logger Edge Solinst® model M20) and two barometers (Barologger Edge Solinst®). These transducers are equipped with Hastelloy type piezoresistive sensors, measuring the total abso- lute pressure above the sensor, i.e. atmospheric plus hydrostatic pressure when submerged. Expected accuracies are within ±1 cm for levelogers, and ±0.05 kPA for barologgers. All loggers had their internal systems codified with unique identifiers and were programmed to collect data every 8 hours (8, 16 and 24h, GMT -4). Cylin- drical supports of galvanized steel, bored and open at the ends were built to allow free water movement and prevent sediment accumulation on the sensor (Figure 3.2 and Figure 3.1). Logger distribution was planned in a GIS environment, to ensure coverage of all flood duration classes mapped by Ferreira-Ferreira et al. 2015. Following the manufacturer’s recommendations, no leveloggers were installed beyond 30 Km http://www.dpi.inpe.br/terraview/index.php http://www.dpi.inpe.br/terraview/index.php http://www.gdal.org/gdal_proximity.html http://www.gdal.org/gdal_proximity.html http://desktop.arcgis.com/en/arcmap/10.3/tools/spatial-analyst-toolbox/an-overview-of-the-surface-tools.htm http://desktop.arcgis.com/en/arcmap/10.3/tools/spatial-analyst-toolbox/an-overview-of-the-surface-tools.htm http://mirador.gsfc.nasa.gov/ CHAPTER 3. SPATIO-TEMPORAL FLOOD MODELING 32 (a) (b) Figure 3.2: Installed leveloggers to flood monitoring in the Southeastern portion of Mamirauá Sustainable Development Reserve (MSDR), in the Central Amazon floodplain, Brazil. (a) Leveloggers have about 16cm and it was fixed in (b) cylindrical supports of galvanized steel, bored and open at the ends to allow free water movement and prevent sediment accumulation. Leveloggers were attached to the base of large trees, at about 5 cm from the ground. from a barologger, thus preventing compensation errors from horizontal atmo- spheric pressure variations. The devices were installed in two field campaigns during the dry season (December/2013 - January/2014). GIS planned logger locations were loaded in a Trimble® ProXRT double-frequency (L1/L2) GNSS receiver. Each levelogger was attached to the base of a large tree, at about 5 cm from the ground, and the barologgers were positioned above the maximum flood water marks left on the trunks. Logger geographic positions were determined with an average planimetric precision of 90 cm (Figure 3.1). Water level data was retrieved for 4 full hydrological cycles, from Decem- ber/January of 2014 to October 2017. Exceptions are the leveloggers 8 and 9 which could not be accessed due to low water conditions in October 2017 and which data extends until January 2017. The raw measures made by the leveloggers were subsequently compensated for atmospheric pressure using the barologger measurements, using the Levelogger Series Software (http://www. solinst.com/downloads/). Final measures were expressed in centimeters of water column above the sensor. As we had three water stages measurements per day, we filtered the data series to include only the daily maximum water stage. Considering the expected accuracies, leveloggers were labeled as “flooded” when a water column≥2 cm was recorded. After labeling all observations as either “flooded” or “non-flooded”, we then determined flood start and end dates, as well as flood height and duration, for each levelogger and hydrological season. 3.2.3 Flood modeling with logistic regression The basic rationale of our proposed method is to be able to predict per pixel flooding probability based on daily observations of water stage level, and then use probability thresholds to determine if it could be considered flooded for the given water level. Daily model predictions obtained for a full hydrological cycle would then allow us to estimate the spatial distribution of flood durations. To describe the relationship between flood probability and predictor vari- ables we used a logistic regression model. Logistic regression uses the log odds- http://www.solinst.com/downloads/ http://www.solinst.com/downloads/ CHAPTER 3. SPATIO-TEMPORAL FLOOD MODELING 33 ratio as link function (Equation 3.1) and the anti-logit to transform logit-scaled values in probabilities (0 ≥ p ≤ 1; 3.2) (Kleinbaum and Klein 2010). loge ( p 1− p ) = α+ β1x1 + β2x2 + ...+ βixi (3.1) p = eβ0+β1x 1 + eβ0+β1x (3.2) where p would be the flooding probability, βi are the model coefficients, and xi are observed values of predictor variables. We first converted our flood maps into binary rasters representing observa- tions of flooded and non-flooded pixels (1/0) for our nine water stages mapped through ALOS/PALSAR imagery. According to Silvapulle 1981, a certain de- gree of overlap in the response variable is necessary so as there be sufficient conditions for the existence of maximum likelihood estimates for a binomial re- sponse model, otherwise at least one coefficient estimate diverges to ±∞ (Heinze and Schemper 2002). The nature of our analyzed phenomenon (floods) did not show any overlap, i.e. there was always a water stage value that perfectly sep- arates whether a given pixel was flooded or not. To fit the model under this restriction, we used the logistf R package (Heinze et al. 2013), which imple- ments the Firth 1993 solution for bias reduction on coefficient estimation using a penalized maximum likelihood scheme (see Heinze and Schemper 2002). We started by fitting a full model including all predictor variables, and suc- cessively tested all possible variable combinations. For each model, we gener- ated predictions of per-pixel flooding probability using water level data from the Mamirauá gauging station spanning 2014 to 2017, the same observation range of the levelogger water level data. To compute annual flood duration, we converted probability values to binary values (flooded-non/flooded). The probability threshold defining if a pixel was labeled as “flooded” was determined through a 4-fold cross-validation. First, we defined every hydrological year as starting at the minimum water level of a given calendar year, and ending at the next year’s minimum. We looked at the probability values of the first hydrological year (2014) in each of the ten levelogger pixels when the devices were first submerged and calculated its mean value. This one-year mean probability was then used as threshold to define flooded/non-flooded pixels (0/1) of the remaining years for the entire study area. The process was repeated for all years, using one year to define the flood probability threshold and the three remaining years for validation. The pre- dicted annual flood duration of the remaining years were compared to levelogger data to compute the RMSE (root mean squared error), finally resulting in 12 RMSE values (4 years used to thresholds definition x 3 remaining years each of these 4 years). Yearly thresholds defined in the cross validation were then averaged and used to generate models results, i.e. daily flood maps and annual flood duration computation. Final model selection was based on the smallest mean RMSE. CHAPTER 3. SPATIO-TEMPORAL FLOOD MODELING 34 3.3 Results and discussion The best inundation prediction model used only water stage (WS) from the Mamirauá gauging station and the HAND terrain descriptor as predictor variables, with one interaction term (model 1, Table 3.2). Mean probability threshold defining if pixels were flooded or non-flooded was 0.942. The model had a mean RMSE of 44.73 days (range: 28.71 to 84.74, see Figure 3.3a) whereas the next three best models achieved RMSE of 47.55, 61.03 and 70.86 days, respectively. Table 3.2: Best logistic models assessed to predict flood duration in the Mami- rauá Sustainable Development Reserve (Central Amazon, Brazil). Model Average RMSE SE 1 − 0.28 + (5.087WS) + (0.039HAND)+ (−0.120(WS ×HAND)) 44.73 4.48 2 − 0.197 + (5.101WS) + (0.060HAND)+ (−0.0001EDND) + (−0.119(WS ×HAND)) 47.55 5.89 3 0.421 + (4.809WS) + (−0.047HAND) 61.03 10.17 4 0.584 + (4.810WS) + (−0.020HAND)+ (−0.0001EDND) 70.86 12.50 Average RMSE - average root mean square errors among 4-fold cross validation folds; SE - standard errors of the averaged RMSE. Results from 4-fold cross validation showed that flood duration errors were mostly influenced by flood onset e