Examining Water Area Changes Accompanying Dam Construction in the Madeira River in the Brazilian Amazon

Two recently constructed run-of-the-river dams (Santo Antônio and Jirau), along the Madeira River in Brazil, have been controversial due to their large unquantified impacts on (1) land use and land cover (LULC) and (2) on the area that would be flooded. Based on annual LULC data from 1985 to 2017, this study integrated intensity analysis and difference components methods to analyze the impacts of the two dams on the annual flooded area in upstream, midstream, and downstream regions of the Madeira River. The dam construction significantly influenced LULC change intensity in the upstream and midstream regions since 2011 and 2010, respectively. An increase of 18.5% of the newly flooded area (462.58 km2) in the post-dam construction period was observed. The water gross gain intensity was active during 2011–2017 and 2011–2014 in upstream and midstream, respectively. The dominant difference components of water change were exchanged in the pre-dam period and became quantity in the post-dam period for both upstream and midstream regions. Forest was the major land category replaced by water; however, the highest gain intensities occurred in other non-vegetated areas in upstream and midstream. This study provided a useful approach for characterizing impacts of dam construction on water area change.


Introduction
The Amazonian basin is the vastest and complex river network on Earth [1]. The river system has great capacity and potential for power generation with possible great economic benefits. At the same time, its natural system supports rich biodiversity with a huge ecological value [2]. Brazil has long embarked on the construction of hydroelectric dams in the Amazon region [3], attracting the attention of scholars, public, environmental protection personnel, and local residents to the associated environmental problems [4][5][6][7]. Quantifying the trends in water area change at different sections of the river is critical to accurately assess the impacts of dam construction and to build policies to protect the river's ecosystem services [8].
The Madeira River is the largest tributary of the Amazon River, with the largest catchment area, water volume, and sediment flow among its tributaries. Recently, the Madeira River's dams of of adjacent land resources [17,24]. LULCC may offset or exacerbate hydropower-induced flow alteration [14,25,26]. Revealing the patterns and processes of LULCC is critical for investigating the complex interactions between human activities and environmental changes [17,27]. The time series of remotely sensed data and products have fostered LULCC investigations worldwide [28,29]. It can regularly monitor water surface area without the need for flow records, which are absent for many rivers [30,31]. A proper representation of water changes over long periods (before and after dam construction) is the foundation for a deep understanding of how changes in water area are related to dam construction. However, using traditional transfer matrix analysis of LULC is difficult to conclude whether the change patterns are more or less intensive compared with a random or uniform change [7,32,33]. Aldwaik and Pontius developed the intensity analysis method to characterize various forms of LULCC, including the areas and intensities of gross change, gains, losses, and transitions among different categories [32]. This method has been used to examine the LULC processes in many regions, reaching results with high confidence levels [34][35][36][37][38]. To supplement intensity analysis and to reflect changes in the components, the gross difference in the LULC matrix has been further dissected into quantity, exchange, and shift by difference components [39]. Previous studies have demonstrated that the integration of intensity analysis with difference components improves the capacity for understanding and interpreting LULC [40]. This method was used to evaluate the effectiveness of cropland requisition-compensation balance policy in maintaining the total cropland area in China [41] and to interpret the LULC during the period of 2000-2010 based on the GlobeLand30 maps of Asia [42]. Previous studies have mainly focused on LULC of two or multiple time points [38][39][40][41], but have not been used to examine the response of a specific land category change process (such as water) to human activities with annual LULC data.
This research aimed to use intensity analysis and difference components to evaluate changes in water surface area and examine the impacts of dam constructions on the change patterns. Specifically, a 33-year LULC dataset was used to quantify the areas, intensities, and components of annual change for LULC and water area from 1985 to 2017 in the Madeira River, the Brazilian Amazon, to answer the following questions: (1) How did the dam constructions affect the LULC process in upstream, midstream, and downstream regions? (2) How did the dam constructions influence changes in water surface area dynamics (dormant or active) and change components? (3) Did the dam constructions significantly change the pattern of water transition size and intensity? (4) Did the distance from dam influence water area gain and loss intensities? This study might provide a new perspective for understanding the impacts of Jirau and Santo Antônio dams on LULC and water area dynamics and for guiding future dam construction plans and management.

Study Area
The Madeira River, which runs a distance of 3352 km from Bolivia to Brazil, is one of the major contributors to the Amazon River in Brazil, and its watershed (i.e., the Madeira River watershed) represents 23% of the Amazon basin [43]. The river flows northward, forming the border between Bolivia and Brazil for approximately 100 km. After receiving the Abuná River, the Madeira River flows northeastward in Brazil through Rondônia and Amazonas states to its junction with the Amazon River. Beginning in the Andean regions of Peru and Bolivia, its altitudinal gradient variation from the Andes towards the Amazon flood plains gives the Madeira River and its watershed's many tributaries high hydroelectric potential. With two large hydroelectric power plants (Jirau (construction was completed in 2012) and Santo Antônio (construction was completed in 2011) [4,6]) (Figure 1), the watershed has the potential to add more dams for hydroelectric power generation [5,10]. The study area was determined based on the locations of the two dams in Rondônia state. The region has a humid tropical climate with annual precipitation around 1900-2200 mm [44]. The Madeira's mean annual discharge is about 18,500 m 3 /s but can reach 45,000 m 3 /s during strong flood periods [45]. In an annual cycle, there are four hydrological periods: low water (August to November), flood (December to January), high water (February to May), and ebb (June to July) [43].

Collection of Annual Land Use and Land Cover Data
MapBiomas provides annual LULC maps for entire Brazil from 1985 to the present (MapBiomas Project, Collection Version 3.0 of the Annual Land Use Land Cover Maps of Brazil, accessed on July-2019 through the link: https:// mapbiomas.org). The MapBiomas united the experts from different fields (ecology, remote sensing, geographic information science, computing science, etc.) and relied on the Google Earth Engine platform and automated classifiers to generate Brazil's annual LULC. In this study, the annual LULC data were used from Collection 3 products covering the period 1985-2017 The Jirau dam is about 136 km from Porto Velho (upstream), while Santo Antônio is located only 7 km from the city [9] (Figure 1). The Jirau and Santo Antônio hydroelectric power plants are two run-of-the-river dams and, therefore, without the need for large water reservoirs [9,43,46]. However, they have been considered mega-dams as their power generation amount to 3750 MW and 3150 MW, respectively [2]. The run-of-the-river dams-the Jirau dam and Santo Antônio dam-have the flooded area of 258 km 2 and 271 km 2 , respectively [6,45]. Table 1 provides some key characteristics of the two dams. Mega constructions in Amazon biome with high potential to environmental and social impacts require previous and in-depth contingency plans and socio-environmental impact studies [10,43], but recommendations are rarely enforced.  [6] 2012 2011 Note: The information was from https://ejatlas.org/print/jirau-and-santo-antonio-dams-on-madeira-river-brazil; https://www.hydropower.org/case-studies/brazil-jirau (acquired on 26 June 2020).

Collection of Annual Land Use and Land Cover Data
MapBiomas provides annual LULC maps for entire Brazil from 1985 to the present (MapBiomas Project, Collection Version 3.0 of the Annual Land Use Land Cover Maps of Brazil, accessed on July-2019 through the link: https://mapbiomas.org). The MapBiomas united the experts from different fields (ecology, remote sensing, geographic information science, computing science, etc.) and relied on the Google Earth Engine platform and automated classifiers to generate Brazil's annual LULC. In this study, the annual LULC data were used from Collection 3 products covering the period 1985-2017 (published in August 2018, Figure 2a-c). The detailed generation process of LULC data is described on the website and summarized briefly here. Firstly, Landsat images were collected with low cloudy coverage and high spectral contrast among LULC categories from a defined period to generate annual mosaics. Secondly, various forms of variables from Landsat bands (105 metrics, including each of the 7 spectral bands, as well as for the calculated spectral fractions and indices) were created to run the random forest classifier. The classification was based on training samples acquired from temporally stable LULC categories, which were generated in the previous products of Collection 2.3-reference maps and manual sampling. Then, the accuracy assessment analysis was conducted in the current map Collection (i.e., Collection v.3.0) [47]. Table 2 provides the definition of the LULC classification system and area occupied by each category in three years (1985, 2011, and 2017). The detailed generation process of LULC data is described on the website and summarized briefly here. Firstly, Landsat images were collected with low cloudy coverage and high spectral contrast among LULC categories from a defined period to generate annual mosaics. Secondly, various forms of variables from Landsat bands (105 metrics, including each of the 7 spectral bands, as well as for the calculated spectral fractions and indices) were created to run the random forest classifier. The classification was based on training samples acquired from temporally stable LULC categories, which were generated in the previous products of Collection 2.3-reference maps and manual sampling. Then, the accuracy assessment analysis was conducted in the current map Collection (i.e., Collection v.3.0) [47]. Table 2 provides the definition of the LULC classification system and area occupied by each category in three years (1985, 2011, and 2017).   Land cover with exposed soils or naturally exposed rocks without soil cover 48

Examining the Impacts of Dam Construction on Land Use and Land Cover Change
Cumulative sum (CUSUM) was utilized to reveal the water area net change from 1985 to 2017. The newly flooded area was defined as the area that had never been water before the dam was completed and was covered by water at least once after the dam was completed. Difference components and intensity analysis were adopted to further analyze LULC and processes of water area change in two steps: separated LULC and water change (the difference between two adjacent years) into quantity, exchange and shift based on difference components method [39], and computed the component's intensity of LULC and water change following intensity analysis theory [32,40,48]. The gain and loss intensities and transition intensity were calculated for a water change. In addition to the difference components and intensity analysis approaches for upstream, midstream, and downstream, buffer analysis was also used to examine the impacts of dam construction on water area change according to different distances from the dam sites.

Difference Components
The quantity component of each category is the absolute difference between the increase of gross area from other categories and the decrease of the gross area to other categories. The exchange component represents the area exchanged between two categories but without a change in the area of both (e.g., category i changes to category j in some locations, while category j changes to category i in other locations during the same period). The shift component is the difference that belongs to neither the quantity component nor the exchange component. Equations (1)-(4) are change component calculations of LULCC at category level: Equation (1) calculates the total difference by summing the area changed to category j from other categories (gross gain) and the area changed to other categories from category j (gross loss).
Equation (2) expresses the quantity component by calculating the absolute difference between gross gain and gross loss for category j (q j ).
Water 2020, 12, 1921 7 of 20 Equation (3) indicates the exchange component by calculating the area that changed to other categories from j and with the same area changed to j from other categories for category j (i j).
Equation (4) gives the shift component by subtracting the areas of quantity and exchange components from the total difference area (d j ) for category j.
where lowercases i and j represent LULC categories, and J is the total number of LULC categories, which were 6 in this study. C ij is the area changed from category i to category j. In this study, Equations (1)-(4) were used to calculate change components for all 6 categories, but this study only focused on water. The q j reflects the net changed area of category j, but the exchange (e j ) and shift (s j ) components do not change the area of category j. Equation (5) was used to compute overall difference size (D) and overall change component sizes (Q (overall quantity change), E (overall exchange change), S (overall shift change)) by summing all categories. The results of Equation (5) were used to evaluate the impacts of the dam constructions on change components of all categories (i.e., LULC).
where X equals D, Q, E, and S; x equals d, q, e, and s, respectively.

Intensity Analysis
The intensity analysis method analyzes LULC at the interval, category, and transition levels. These three levels explore very fine detail in the patterns of LULC [36,48]. The interval level can be used to examine the change dynamics of all LULC categories as a whole. Category level compares observed gain and loss intensities of each category with the uniform change intensity based on LULCC in every time interval. Transition level analysis shows the intensity of category i changing to category j compared with the uniform transition intensity. Due to the disproportionate distribution of area within the 6 categories (Table 2), water gain size from other categories was calculated to realize the water-related LULC at transition level, but not compared with uniform transition intensity.
Equation (6) shows the annual change intensity U t for a time interval [Y t , Y t+1 ]. It can be compared with the uniform annual change U of the study period [Y 1 , Y T ]. If U < U, then the change of this time interval is slow. If U > U, then the change is fast. The time interval refers to a year to year change [Y t , Y t+1 ], and when the interval is more than two years in a temporal extent of our statement Equation (7) was used to compute the three overall change component intensities (Q' (overall quantity change intensity), E' (overall exchange change intensity), and S' (overall shift change intensity)), with d equals D and x equals Q, E, and S, respectively, at the interval level. Equation (7) was also used to compute the three change component intensities (q', e', and s') for each category j (here only focused on water) by taking the size of the component (x equals q, e, and s) divided by the total changed area (d). This study presented LULCC dynamics with a focus on the analysis of water dynamics at category and transition levels.
The loss intensity is the ratio of the lost area to the initial area for category i in each interval (Equation (8)).
Equation (9) calculates the gain intensity of category j as the ratio of gain area to the area of category j at the final time in each interval. If l' (g') < U, then l' (g') is considered dormant, meaning this category has experienced lower intensive loss (gain) than the mean level across the spatial extent. If l' (g') >U, then l' (g') is considered active, indicating that the loss (gain) of the category j is more intensive than the mean level across the spatial extent.
Equation (10) gives the transition intensity by computing the ratio of the area changed from category i to j by the initial area of category i.

Buffer Analysis
Buffer zone analysis was adopted to identify the areas sensitive to the dams by checking the water area changes according to the distances from each dam. Buffer zones surrounding the river were created at 12 km on two sides. The buffer zones were divided into three parts-upstream, midstream, and downstream-according to the locations of the dams (Figure 2d). To examine the distance effects on water area change, we considered the two dams as the centers to set buffer zones at 10 km intervals. In the upstream region of the Jirau dam, there were 8 buffer zones. The midstream region was an upstream area of Santo Antônio dam and downstream of the Jirau dam, and the buffer zone was most overlap with each other when setting buffer zones based on each dam. Therefore, the downstream of the Jirau dam used the buffer zone of the Santo Antônio dam. There were 8 buffers for upstream and midstream and 14 buffer zones for downstream, respectively.  (Figure 3e). The proportions of quantity, exchange, and shift did not significantly change during the study period (Figure 3f). The change intensities were lower in the upstream than downstream before 2011 but higher after 2011. This indicated that the construction of the dams had significantly impacted the LULCC process in the upstream.

Impacts of Dam Construction on LULCC
LULCC process in the upstream.
To examine the impacts of dam constructions on each LULC category in the upstream, midstream, and downstream, the statistics of gross loss, gross gain, and net changed area were calculated for pre-dam (1985-2011) and post-dam (2011-2017) periods, respectively (Tables S1-S3, Figure S1). Farming increased at the cost of forest reduction in the pre-dam period, but the pattern became more complicated in the post-dam period. The gross loss and gross gain of areas of each LULC category increased significantly in upstream (Table S1). Some categories (e.g., gross losses of the forest, other non-forest natural formation (ONFNF), and other non-vegetated areas (ONVA); gross gains of ONFNF, ONVA, and water) accounted for the largest proportions to the overall study area in the post-dam period (Table S2). The flooded area and ONFNF were the major drivers of deforestation (Table S3, Figure S1). The average annual increased area of farming changed from 17.95 km 2 to 4.43 km 2 . In the midstream, the gross loss area of farming increased significantly, and the gross gain was similar between the two periods. A net reduction of the farming area was observed, which was mainly caused by large amounts of the flooded area, and returned to the forest in the post-dam period in midstream. A large area of farming land was inundated in 2011-2012, but no significant replacement of flooded farming land was observed ( Figure S1). Forest showed an adverse change trend compared with that of farming, indicating the deforestation area decreased during 2011-2017 in midstream. Some farming land was recovered after a few years of dam completion in upstream and downstream ( Figure S1b,d).  To examine the impacts of dam constructions on each LULC category in the upstream, midstream, and downstream, the statistics of gross loss, gross gain, and net changed area were calculated for pre-dam (1985-2011) and post-dam (2011-2017) periods, respectively (Tables S1-S3, Figure S1). Farming increased at the cost of forest reduction in the pre-dam period, but the pattern became more complicated in the post-dam period. The gross loss and gross gain of areas of each LULC category increased significantly in upstream (Table S1). Some categories (e.g., gross losses of the forest, other non-forest natural formation (ONFNF), and other non-vegetated areas (ONVA); gross gains of ONFNF, ONVA, and water) accounted for the largest proportions to the overall study area in the post-dam period (Table S2). The flooded area and ONFNF were the major drivers of deforestation (Table S3, Figure S1). The average annual increased area of farming changed from 17.95 km 2 to 4.43 km 2 . In the midstream, the gross loss area of farming increased significantly, and the gross gain was similar between the two periods. A net reduction of the farming area was observed, which was mainly caused by large amounts of the flooded area, and returned to the forest in the post-dam period in midstream. A large area of farming land was inundated in 2011-2012, but no significant replacement of flooded farming land was observed ( Figure S1). Forest showed an adverse change trend compared with that of farming, indicating the deforestation area decreased during 2011-2017 in midstream. Some farming land was recovered after a few years of dam completion in upstream and downstream (Figure S1b,d).

Water Surface Area Change and Newly Flooded Area after Dam Construction
Water surface area in upstream, midstream, and downstream regions increased and then decreased after the dam was completed in 2012 ( Figure 4)

Dynamics of Water Change Components
In the upstream region, the average percentage of quantity change was 30.6% during the period of 1985-2011 (before the dam was built) and increased to 60.6% during the period of 2011-2017 (post-dam construction) (Figure 6a)

Dynamics of Water Gross Gain and Gross Loss
There were 12 intervals (for a total of 26 time intervals) with water gross gain greater than gross loss before 2011, and the water area net increased by 2.4 km 2 during this period in the upstream region ( Figure 7a). The water area had a net increase of 89.2 km 2 during 2011-2017 as most intervals (within this period) showed larger gross gain than gross loss. Water gross gain and gross loss were 18 times at dormant state (less than uniform category intensity) before 2011, but the distribution was different (Figure 7b). It was observed that water gross gains were continuously active from 2011, while gross losses at a dormant state in 2011-2014 but active from 2014 onwards.
There were 13 intervals (for a total of 25 time intervals) showing gross gain less than gross loss before 2010, and the water area net reduced by 2.9 km 2 in midstream region (Figure 7c). There were four intervals (total of seven time intervals) with a net gain, and the water area net increased by 123.

Dynamics of Water Gross Gain and Gross Loss
There were 12 intervals (for a total of 26 time intervals) with water gross gain greater than gross loss before 2011, and the water area net increased by 2.4 km 2 during this period in the upstream region ( Figure 7a). The water area had a net increase of 89.2 km 2 during 2011-2017 as most intervals (within this period) showed larger gross gain than gross loss. Water gross gain and gross loss were 18 times at dormant state (less than uniform category intensity) before 2011, but the distribution was different (Figure 7b). It was observed that water gross gains were continuously active from 2011, while gross losses at a dormant state in 2011-2014 but active from 2014 onwards.
There were 13 intervals (for a total of 25 time intervals) showing gross gain less than gross loss before 2010, and the water area net reduced by 2.9 km 2 in midstream region (Figure 7c). There were four intervals (total of seven time intervals) with a net gain, and the water area net increased by 123.    (Figure 8b). In the midstream region, the water gain was mainly from the forest, farming, and ONVA before 2011. Water gain was concentrated on forest and farming from 2011 to 2017 (Figure 8c). The contribution from ONVA to water significantly reduced compared with that of the upstream region. Gain intensity from ONVA was larger than other categories before 2011. The whole gain intensity significantly increased in 2011-2012, and the gain intensity of ONFNF showed an obvious increase compared with the period before the dam was constructed. The overall gain intensity decreased rapidly after 2014

Water Gain Size and Intensity from Other Categories
The three regions showed different change trajectories in water gain size and intensity from other categories. Water gain was mainly from the forest, ONFNF, and farming with a percentage of 46.6%, 21.2%, and 23.8% before 2011 (pre-dam period) in the upstream region (Figure 8a). Large gain intensities occurred on construction land and ONVA. After 2011 (post-dam period), water gain was mainly from forest and ONVA (42.3% and 26.2%), and ONVA showed the largest gain intensity in 2011-2012. The water gain from the forest mainly occurred between 2013 and 2015 (Figure 8b). In the midstream region, the water gain was mainly from the forest, farming, and ONVA before 2011. Water gain was concentrated on forest and farming from 2011 to 2017 (Figure 8c). The contribution from ONVA to water significantly reduced compared with that of the upstream region. Gain intensity from ONVA was larger than other categories before 2011. The whole gain intensity significantly increased in 2011-2012, and the gain intensity of ONFNF showed an obvious increase compared with the period before the dam was constructed. The overall gain intensity decreased rapidly after 2014 (Figure 8d). In the downstream region, there was no significant difference between 1985-2011 and 2011-2017 for water gain patterns (Figure 8e). Water gain from forest accounted for 50.15% in 1985-2017. ONVA and ONFNF showed strong gain intensity in most intervals (Figure 8f).

The Uncertainty of Classification and Its Influence
It is important to know how the classification uncertainties in the LULC data affect the analysis and the conclusions. The hypothetical error intensity at the category level for water gain and loss in each interval (i.e., year to year from 1985 to 2017) was computed to test a null hypothesis that the change intensities among all LULC categories are uniform [36,48]. A larger hypothetical error gives stronger evidence against this null hypothesis. The whole study area was taken as an example to illustrate the error analysis of LULC. Water gain commission/omission error intensities ranged from 0.71 to 86, and water loss commission/omission error intensities ranged from 0.06 to 74, both with the average 33 under the uniform change intensity hypothesis (Table S1). Thus, given these intervals with very low hypothetical error intensity, the conclusion that water change was dormant/active had weaker confidence. MapBiomas reported the overall and per-category classification accuracy for each year during the period of 1985-2017 at the Amazon biome level. The overall accuracy of the LULC classification was 94 at level 1 in each year. The producer's accuracy and user's accuracy of water ranged from 72 to 86 and 83 to 93 in 1985-2017, respectively (Table S4). Based on this accuracy assessment, there were 7 years (1985, 1991, 1995, 2001, 2006, 2008, and 2012) when the actual commission/omission error intensities were larger than hypothetical error intensity for water loss. There were 8 years (1989, 1995, 1998, 1999, 2000, 2003, 2006, and 2016) when the actual commission/omission error intensities were larger than hypothetical error intensity for water gain. These years had low confidence in the dormant/active conclusions. Thus, the conclusions about the pattern of water change, whether stable or not, across these time intervals in terms of gain

The Uncertainty of Classification and Its Influence
It is important to know how the classification uncertainties in the LULC data affect the analysis and the conclusions. The hypothetical error intensity at the category level for water gain and loss in each interval (i.e., year to year from 1985 to 2017) was computed to test a null hypothesis that the change intensities among all LULC categories are uniform [36,48]. A larger hypothetical error gives stronger evidence against this null hypothesis. The whole study area was taken as an example to illustrate the error analysis of LULC. Water gain commission/omission error intensities ranged from 0.71% to 86%, and water loss commission/omission error intensities ranged from 0.06% to 74%, both with the average 33% under the uniform change intensity hypothesis (Table S1). Thus, given these intervals with very low hypothetical error intensity, the conclusion that water change was dormant/active had weaker confidence. MapBiomas reported the overall and per-category classification accuracy for each year during the period of 1985-2017 at the Amazon biome level. The overall accuracy of the LULC classification was 94% at level 1 in each year. The producer's accuracy and user's accuracy of water ranged from 72% to 86% and 83% to 93% in 1985-2017, respectively (Table S4). Based on this accuracy assessment, there were 7 years (1985, 1991, 1995, 2001, 2006, 2008, and 2012) when the actual commission/omission error intensities were larger than hypothetical error intensity for water loss. There were 8 years (1989, 1995, 1998, 1999, 2000, 2003, 2006, and 2016) when the actual commission/omission error intensities were larger than hypothetical error intensity for water gain. These years had low confidence in the dormant/active conclusions. Thus, the conclusions about the pattern of water change, whether stable or not, across these time intervals in terms of gain and loss intensities should be taken with caution. These years did not influence the conclusion about the dam constructions on water resource change. For upstream, midstream, and downstream regions, low confidence also most occurred before 2011, similar to the whole study area, and only water loss showed low confidence in 2012 after the completion of the dams (Table S1).

Planned Versus Observed Impacts of Dam Construction on Water Area
The Jirau and Santo Antônio dams in the Madeira River have been extremely controversial [5]. Cella-Ribeiro et al. [9] concluded the two dams were less harmful to fish assemblage compared with vertical-axis turbine systems based on dams, and given the results were suggested for other hydroelectricity projects in the Amazon. Latrubesse et al. [2] found that the Madeira River was seriously affected by the two dams and represented the most vulnerable area in the entire Amazon basin. Thoroughly understanding how dam construction affects the spatial-temporal process of LULC would provide key knowledge for future hydropower development and for ecosystem health assessment. The original environmental impact assessment (EIA) predicted that the water area of the Santo Antônio and Jirau dams' reservoirs (corresponding to our upstream and midstream regions) would cover 529 km 2 . Our results showed that the lowest water area was 470.2 km 2 in 2012, and the largest water area was 626.3 km 2 in 2014 (abnormally high precipitation year) over the period of post-dam construction. It seems that EIA provided a relatively reasonable assessment of how the water area would change as an effect of the dams. The newly flooded area of 424 km 2 in the upstream and downstream was larger than EIA estimated (i.e., 308 km 2 ). The proportion of permanent water coverage throughout 2012-2017 was relatively low, and one-third of the area was flooded only once after the dam was completed.

Water Change Intensity and Component Among Different Regions
Compared with traditional methods that subjectively divided study periods into two or several segments based on the dam construction date [7], intensity analysis quantified the land change area and intensity and checked the stationarity state at the interval, category, and transition levels every year. For example, upstream and midstream regions had fast dynamics in 2011-2017 and 2010-2017 intervals, and the downstream kept consistently slow in 2006-2015 at an interval level for LULC. Thus, the dam constructions immediately changed the LULC patterns of upstream and midstream regions. Both upstream and downstream regions reached the highest change intensity at the 2015-2016 interval, while midstream peaked at 2011-2012 in terms of LULC. The dam constructions had a more direct and intense impact on midstream, while more gradual influences on the upstream region and no direct impact on downstream. Intensity analysis provided much information about each category changes from category level analysis. This is very helpful for studies that concern about a particular category of LULCC. Water gain intensity was continuously active during 2011-2017 in the upstream, meaning that water area due to dam constructions experienced more intensive gain than the uniform overall change for a long period. However, it should be noted that the loss intensity became active with an increasing trend during 2014-2017, meaning the water area experienced more intensive loss than overall change after a few years of dam operation. The midstream showed a different pattern with the water gain intensity only activated in 2011-2014. The gain intensity was dormant, and the loss intensity was active in 2014-2016. This dynamic brought a significant decrease in the water area and high proportions of quantity change. This trend was also observed by Cochrane et al. [4]. One of the important reasons might be the two dams were designed as run-of-the-river hydropower stations [9], and they did not need to preserve large reservoirs. However, in our study, the water area was based on the LULC data from the low water period (August to November) because the selected Landsat images were from June 1 to October 31 each year. The impacts caused by the dams might be substantially underestimated [4].
The water gain size and intensity also provided more information about the LULC category's transition from transition level analysis. Forests were the main category replaced by water (area changed) before and after the construction of the dams in the three regions. Intensity analysis attracted more attention to ONVA and ONFNF after the dam constructions in the upstream and midstream, respectively. Thus, we need to pay more attention not only to these newly flooded forests but also to non-forest and bare lands. With the prolonged and frequent submersion of forests and farmlands, these areas might change to ONVA and be sensitive to the hydrological changes related to the dams. The change in their ecological functions deserves more attention.
Change components analysis enhanced the interpretation of the difference between a category and overall LULC in terms of area and intensity. Moreover, the overall component intensity was conducive to the comparisons among different regions and time intervals. The dams had different impacts on LULC change components and water change components. The exchange was the major change component of LULC during 1985-2017 without significant changes after the construction (2012). According to the change components of water, the proportions of quantity change significantly increased after the dams were completed (2012) and became the major change components in upstream and midstream. For the water category in the Amazon region, the main omission error occurred in the forest (ranged from 0.11 to 0.25), and the main commission error occurred in ONFNF (0.04-0.11) and forest (0.02-0.08). Misclassification in the LULC data can contribute to exchange. Some of the forests on the floodplain of the Madeira River grow in the water for several months each year, i.e., they are called flooded forests [5]. This ecosystem dynamic inevitably challenges the LULC classification of these targeted regions, a challenge considerably augmented, given the high inter-annual variability due to the amount of precipitation and water level. The largest exchange size in water before dam constructions occurred in different years for upstream, midstream, and downstream regions, meaning the classification error might play a limited role in the exchange. Although the proportion of exchange in change components decreased after the construction of dams, the average exchange area of water changed from 2.54 km 2 to 25.3 km 2 after the construction of the dams and the area reached 57.1 km 2 in 2016-2017 in the upstream region. The construction of dams significantly increased the exchange area in the upstream and midstream regions and led to a more fragmented watershed in the river system [4]. This is all the more surprising, given that the intent of building two dams, rather than one, was said to be to avoid flooding in Bolivia. Why Jirau is not releasing more of this water, producing more energy, and reducing the flooded area needs an explanation from those who manage the energy sector.

Distance Effects of Dam Construction on Water Area Change
Generally, the influence of the construction of the dams on the observed variations in water area decreases with the increase in distances from the dam sites and gradually stabilizes at a certain distance-a trend observed in previous studies [17]. The threshold strongly depends on the characteristics of topography and reservoirs systems and size. Jirau and Santo Antônio dams were the first to adopt the horizontal bulb turbines and run-of-the-river technology in the Amazon region, therefore operating without large reservoirs. Previous studies have mostly focused on reservoir-based dams [7,11,17], while the influences of horizontal bulb turbines have been less evaluated. According to our study, the 30 km and 60 km buffer zones in the upstream were more sensitive to the construction of the dam. This was because these two buffer zones included the confluences of São Lourenço River and Mutum-Paraná River with the Madeira River, where a large area of their mouths flooded after the completion of the dams ( Figure 5). In midstream, the water area decreased from 10 km to 30 km and increased from 30 km to 60 km and decreased from 60 to 80 km in 2011-2012, but this trend disappeared in the following years. For downstream, only the 80 km buffer zone showed an obvious change. This change was from the nearby seasonal lake and not from the Madeira River. The two dams did not cause noteworthy water area change in the downstream region. This is as it was meant to be. What is less clear why are the two dams holding more water in their reservoirs and increasing flooded area when they are meant to let the river flow as naturally as possible and not hold water in storage. This may be due to the selection of horizontal bulb turbines, which could be considered in future hydropower constructions in the Amazon region.

Conclusions
In this study, intensity analysis and difference components were used to characterize the LULC during the period of 1985 to 2017, and the impacts on water area dynamics caused by the two run-of-the-river dam constructions along the Madeira River in the Brazilian Amazon were examined. Our results showed that LULC change intensities of upstream and midstream regions became consistently higher in the year before the completion of Jirau and Santo Antônio dams, respectively, but the dam constructions did not enhance LULC change intensities downstream in the first few years after the constructions. The LULC change components of upstream and downstream were not significantly impacted, while midstream showed a high proportion of exchange in the two years before the dams were built.
The construction of Santo Antônio and Jirau led to a large increase in newly flooded areas significantly greater than predicted by EIA in 2005. However, only 18.5% of the newly flooded areas became permanent water throughout 2012-2017. The water gross gain intensity was different in three regions. In the upstream, it was observed an active steady-state of water gross gain intensity after the completion of the dams, while in the midstream, it only lasted for three years. The response of water change in downstream was not obvious, but the largest gross gain and loss both occurred after the completion of the dams. The dam constructions changed the major difference component from exchange to quantity in both upstream and midstream regions. The confluences with tributaries determined the susceptibility of water area change in different distances, especially in upstream. Forest was the largest flooded category in the three regions in the post-dam period, while the highest gain intensity all occurred in ONVA after dams were completed. ONVA and farmland were the second largest flooded categories in upstream and midstream, respectively. Overall, the intensity analysis and difference component provided more information about the water area change for assessing the impacts of two run-of-the-river dams in the Madeira River than previously used methods. This study provided a comprehensive framework to assess the impacts of dam construction on water area change based on time series of LULC data.
Supplementary Materials: The following figure (S1) and tables (S1, S2, S3, and S4) are available online at http://www.mdpi.com/2073-4441/12/7/1921/s1. Figure S1: Annual gross loss and gain areas of forest and farming from 1985 to 2017 in upstream, midstream, and downstream. (Note: left bar and right bar in each year interval are gross losses to other categories and gross gains from other categories, respectively. The gross losses and gains of ONFNF, construction land, and ONVA were not provided here because of their small area. Water was analyzed in Section 4.2 in more detail. ONFNF: other non-forest natural formation, ONVA: other non-vegetated areas), Table S1: Average annual gross gain and loss areas of each category in pre-dam and post-dam periods in upstream, midstream, and downstream (km 2 ), Table S2: Average relative annual gain and loss areas to the study area for each category in pre-dam and post-dam periods (%), Table S3: Average annual net change of each category in upstream, midstream, and downstream before and after the construction of the dam (km 2 ), Table S4: Commission and omission error intensities for upstream, midstream, downstream, and study areas for each interval from 1985 to 2017 (%).