Detectability of variation in river flood from satellite images

Floods are major natural disasters that have considerable consequences worldwide. As the frequency and magnitude of flooding are expected to be affected by ongoing climate change, understanding their past changes is important for developing adequate adaptation measures. However, the limited spatiotemporal coverage of flood gauges hinders detection of changes in flooding, particularly in poorly gauged regions. Here, we propose a method using surface water data of river floodplain inundation as a proxy of the magnitude and frequency of flooding. Surface water data − Aqua Monitor which represented the probability linear trend changes in land and water surface area based on 30-m Landsat images between 1984–2000 and 2000–2013 was used in this study. The changes in water surface area over the floodplain obtained from Aqua Monitor showed high correspondence with historical trends observed or simulated annual maximum daily discharge, indicating the potential to detect changes in frequency and magnitude of flood from satellite data. In regions where changes could be measured with sufficient satellite images, 29% showed an increase in water surface area in the flood plain, 41% showed a decrease, and 30% showed small or no changes.


INTRODUCTION
Flooding is a common major natural hazard in many areas of the world, and directly affects people's lives and livelihoods. Global warming will increase the frequency of floods, especially in Asia and Africa where rapid growth of the population and assets are expected (Hirabayashi et al., 2013). However, changes in flooding in these regions are poorly understood due to limited numbers of in situ observations.
A satellite remote sensing-based approach has been widely applied for assessing recent changes in water surface area, including flood changes. For example, Najibi and Devineni (2018) showed an increase in the frequency and duration of floods worldwide using a global flood database based on water surface area images derived from remote Correspondence to: Haireti Alifu, Department of Civil Engineering, Shibaura Institute of Technology, 3-7-5 Toyosu, Koto-ku, Tokyo 135-8548, Japan. E-mail: haireti.alifu.f4@sic.shibaura-it.ac.jp sensing satellites. Ji et al. (2018) generated the 500-m Resolution Daily Global Surface Water Change Database (DGSWCD500m), derived from more than 1.9 million frames of the MODIS surface reflectance product (MOD09GA) from 2001 to 2016 which have high potential applicability to the determination of flood inundation. Donchyts et al. (2016) developed a Deltares Aqua Monitor, which shows a global map of water and land surface area changes over 1984-2013 at 30 m spatial resolution using the multi-petabyte archive of millions of Landsat series images. It contained probability linear trend changes in land and water surface area, and this information around river floodplain reflects the river flood dynamic. Hence, assessing the utility of high-resolution water surface data for revealing changes in river flooding patterns is possible and important, especially in data-poor regions.
The aims of this study are to: (1) evaluate the ability of the satellite-based water surface area change in a floodplain to detect changes in magnitude and frequency of flood during 1984-2013 by comparing those of in situ observations or simulated river discharge, and (2) discuss the observed changes in satellite-derived water surface area and peak discharge data at global scale.

DATA AND METHODS
The high-resolution Multi-Error-Removed Improved Terrain (MERIT) global digital elevation model (DEM) , stream gauge station data from the Global Runoff Data Center (GRDC), global daily discharge reanalysis, and Aqua Monitor (Donchyts et al., 2016), are the main data used for this study. Other auxiliary data, such as the FLOod PROtection Standards, FLOPROS (Scussolini et al., 2016), dam index derived from GRanD (Lehner et al., 2011), percent tree cover data (Kobayashi et al., 2016), DGSWCD500m (Ji et al., 2018), and Emergency Events Database, EM-DAT (Guha-Sapr et al., 2016) are also used for this study. Detailed information for data used in this study is available in the supplementary materials (SM) (Table SI-SIII).
In order to detect the flood variation, it is necessary to perform several preprocessing steps, such as upscaling, masking out pixels that are out of the river floodplain, measuring pixel changes in the sub-basins scale and manmade large surface area changes, and excluding locations where there were too few satellite observations. Detailed descriptions of the datasets and the method are provided in the following sections.

Global floodplain and sub-catchment data
To remove water pixels outside of the floodplain, 1 min resolution of a global floodplain area was generated from the MERIT DEM . The MERIT DEM was generated through removal of multiple errors from existing DEMs, and performs better than similar previous products. The floodplain was defined as the area within 5 km of major rivers (total drainage area >3,000 km 2 ) with a water surface height of <10 m. The MERIT DEM was also used to define 445 global subcatchments with areas >100,000 km 2 . The inundation area of each sub-catchment was then analyzed.

Stream gauge data and global daily discharge reanalysis
Data for 70 stream gauge stations were provided by the GRDC (https://www.bafg.de/GRDC/EN/Home/homepage_ node.html). These stations are selected based on the following criteria: 1) stations have larger drainage basin area (area >100,000 km 2 ) and long-term discharge data (at least 30 years of observed daily discharge data for the period 1975-2015 and data were missing for fewer than 73 days in the year, i.e. 20% of the year), 2) stations are located at or near the outlet region of the sub-catchment. Moreover, if there were several GRDC stations in the same sub-catchment, we analyzed the data from the station with the largest drainage basin ( Figure S1).
On the other hand, discharge reanalysis for the 1984-2013 period was performed using a river and inundation model, the CaMa-Flood, using daily runoff input derived from a land surface model driven by bias-corrected meteorological reanalysis (Tanoue et al., 2020, Text S1). CaMa-Flood realistically simulates floodplain inundation dynamics and improves predictions of daily scale river discharge, at the high spatiotemporal resolution, by incorporating floodplain inundation dynamics (Yamazaki et al., 2011). Evaluation of discharge reanalysis indicated that (Figure S1-S2) peak discharge has a good correlation with most of the peak discharge derived from GRDC. However, lower correlation coefficients (R) were estimated in the cold river catchment and highly regulated river catchment due to the CaMa-Flood model does not consider human river management activities of reservoirs, irrigation, or land use. Hence, the modeled discharge reflects flood changes driven by natural hydrological processes (e.g. changes in precipitation, snow, and evapotranspiration). Despite the inevitable biases, as the discharge reanalysis showed anomalous high peaks in flood years, we assume that the discharge reanalysis can be used as a proxy of the magnitude of the flood where in situ observation was unavailable.

Satellite-derived water surface area changes between 1984-2013
Aqua Monitor (Donchyts et al., 2016), performed waterland and land-water surface area change by calculating linear regression of the wetness index values derived from a large number of Landsat series images at 30 m pixel. The proportions of pixels showing an increase (land to water) or decrease (water to land) over the floodplain of each subcatchment was calculated based on existence of water pixels between the period of 1984-2000 and 2000-2013. As Landsat has a 16-day interval, it has sufficient opportunity to detect inundation area extent due to flooding in large river basins having large river extent for a month or more (Bender, 1991). We assume that the increase or decrease of water pixels within inundation area reflects changes in frequency or magnitude of flooding between the two periods.
To minimize the effects of artificial and natural river changes, such as shifting, capturing, and straightening of river channels, we first scaled the original images (30 m) up to a resolution of 1 min (~1.84 km) and then measured the increase or decrease in the number of water pixels in the river floodplain using floodplain and sub-basin masks defined by the MERIT DEM. We expected the upscaling procedure to cancel-out the increase and decrease of water pixels associated with route management and natural river shifting. The average linear trend change values in land and water surface area were then calculated over the floodplain in each sub-catchment. To exclude locations where there were too few satellite observations to determine the change in inundation extent, we selected sub-catchments with on average >170 Landsat images between 1984 and 2013. To avoid images of small inundation areas, the proportion of water pixels in the analyzed floodplains of each subcatchment was required to be >15%. Eighteen subcatchments with new reservoir construction after 2000 were removed from the analysis using reservoir and dam data of GRanD because the number of water pixels was considered to have increased due to new reservoir construction (Text S2, Figure S3). Finally, we selected and analyzed 177 subcatchments.
Generally, heavy rainfall induced floods cause an immediate response (lasting from only a few hours to a few days) in small streams, but in large rivers, flood runoff may exceed channel capacity for more than a month (Bender, 1991). Ponds are a very common and abundant habitat in natural floodplains. Large extreme flooding always created ponds/lakes or expanded the area of ponds/lakes that already exist in the natural floodplain. These ponds/lakes can keep their area for a long time (except in arid regions) as they are regularly flooded or are connected to a river. Thus, land-water and water-land change information from Aqua Monitor has the potential to represent flood trend in the case of a large river basin. Our analysis of DGSWCD500m (Text S3), which has a lower spatial resolution than the Landsat images implemented in Aqua Monitor, using a method similar to that of Alifu et al. (2019), confirmed the ability of satellite-derived daily water surface area data to detect the instantaneous magnitude of flood events, when and where large floods occurred in the past, especially when the magnitude of the past flooding was large (return period >100 years) ( Figure S4-S6 and Table SII). Importantly, it also supports the view that satellite-derived water surface area can detect flood events in regions with high flood protection (Text S3).

Trends in the annual maximum daily discharge
The trends in flooding were analyzed for 177 global subcatchments using gauge data ( Figure S7a), the daily dis-Y. HIRABAYASHI ET AL. charge reanalysis ( Figure S7b), and Aqua Monitor ( Figure  S7c). The percentage changes per year in the observed annual maximum daily discharge (GRDC) were calculated by fitting values to a linear trend in the latest 30 years between 1975 and 2015, with the available observed period varying for each station. The same calculation was performed for discharge reanalysis using data from 1984-2013.
For the global analysis of trends in flooding, the discharge analysis at the outlet pixel which has the largest accumulated flow within the sub-catchment, defined by a river routing network in CaMa-Flood, was calculated to represent the discharge in that sub-catchment.

Evaluation of flood change
To determine whether the satellite-derived surface area change reflected historical trends in river flooding, the data were compared with the observed annual maximum daily discharge data for 22 sub-catchments where tributaries had little effect (Table I and Figure S8). By tracking changes in the proportion of water pixels using Aqua Monitor, we were able to detect cases showing the same direction of change in observed annual maximum daily discharge. Of 22 basins, 19 showed the same direction of change. Among these, 12 showed increased flooding, and 7 showed a decrease.
However, in three basins (Missouri, Burdekin, and Snake), although the observed annual maximum daily discharge showed increasing trends, the inundation areas, as analyzed using Aqua Monitor, exhibited decreasing trends. One potential reason for the discrepancy is the effect of the water surface area of reservoirs. Because we defined floodplains according to the distance from the main river, water pixel changes may reflect changes in water surface area due to reservoirs located along the main channel of the river. To know the effects of reservoir/dam to river discharge, we calculated the dam index ( Figure S3) as the total volume of reservoirs in the upper basin obtained from the Global Reservoir and Dam Database (GRanD) (Lehner et al., 2011) divided by the average daily discharge for the period 1958-2013. Table I and Figure S3 indicated the Missouri River is highly controlled, as shown by the high dam index. There were several flood events during the analysis period, which caused increasing trends in both the observed and discharge reanalysis annual maximum discharge. However, at the same time, the Missouri river basin experienced a prolonged drought between 2000 and 2007, which caused a significant decrease in reservoir storage (Gao et al., 2012) that could have been responsible for the decreasing trend in Table I. Summary of past changes in flood in 22 selected basins for which observations are available from the Global Runoff Data Centre (GRDC). Numbers in the table correspond to the location map in Figure S8a. *, p < 0.10; ** p < 0.05 These floods may have caused the increasing trends in the observed and discharge reanalysis annual maximum daily discharge. Another potential reason for these trends is water management; for example, there are five streamflow gauges in the same sub-catchment of the Snake River, and our analysis showed an increasing trend (+ 1.12% yr -1 ) based on the observation data of the station situated within the largest drainage basin area, whereas the remaining four gauges at upstream locations showed decreasing trends (ranging from -0.20 to -2.26% yr -1 ). This may have been due to intensive irrigation activities. The increasing trend of the gauge located farthest downstream was less affected by the irrigation activities due to return flows from surface water irrigation (Hoekema and Sridhar, 2011). The satellite-derived trend in the water surface area showed similar direction of trend change with the observed trend (Table I and Figure  S8) and indicated that exclusion of reservoirs from the floodplain mask and consideration of irrigation water usage would increase the likelihood of detecting changes in flooding trends.
Comparison of discharge reanalysis with observed discharge and satellite-derived water surface area changes indicated that human water management activities may have had an effect. In the Yukon, Peace, Ohio, Yellowstone, Colorado, and Severnaya river basins, the annual maximum daily discharge of discharge reanalysis showed the opposite trend to those of observed discharge and satellite-derived water surface changes. Several sub-catchments showed increasing trends in discharge reanalysis but decreasing trends in observed or satellite-based inundation area. For example, the discharge reanalysis showed increases in flooding in two Colorado basins, and in the Peace and Severnaya River basins, whereas observed and satellitederived flood areas showed decreasing trends. Conversely, discharge reanalysis indicated that the likelihood of flooding decreased in the Ohio, Yukon, and Yellowstone River basins, but observed discharge and satellite-derived data indicated that the occurrence of flooding increased.
The discrepancies between the discharge reanalysis and observed trends suggested that human activities leading to a reduction in peak floodwater and reservoir operations affected river discharge, as the model did not consider river management activities. We assumed that when the dam index was larger, river discharge was more affected by reservoir operations in the upstream area. The results showed substantial reservoirs along the Ohio, Yellowstone, and Colorado Rivers (Table I), indicating that reservoir operations or other river management activities affect flood trends. Other potential sources of discrepancy include uncertainties in modeled hydrological processes and climate inputs. The current modeling framework cannot replicate ice jam flooding, which is the main cause of floods in the Peace, Yukon, and Severnaya basins (Beltaos, 2018;Brabets et al., 2000;Peters and Prowse, 2001;Vuglinsky, 2002). Although discharge reanalysis has been used as a good indicator to show changes in river flooding where in situ observation is unavailable (e.g. Alfieri et al., 2020), additional information derived from satellite images compensate for its disadvantage in particular for basins with human water management.
We also evaluated the effects of local flood protection measures on the ability to capture flooding trends using the global database of FLOPROS (Scussolini et al., 2016, Text S4, Table I and Figure S9). The result in Table I shows that the flooding trend derived using Aqua Monitor is in the same direction as the actual trend, although these catchments vary in flood protection level. This illustrated that despite levees and other flood-protection measures, Aqua Monitor can detect high volumes of water, changes in river width between levees, and changes in natural inundation area (including flood retention ponds).
Despite the discrepancy between the modeled and actual results for some basins, due mainly to the effects of human activities, 19 of 22 events for which observation data were available showed concordance in the direction of change between the satellite-and gauge observation-derived trends. Thus, satellite-based water surface area changes in a floodplain have a potential to detect changes of extension of flood water (particularly important in poorly gauged regions) within the observed time in Aqua Monitor, which may relate to both frequency and magnitude of flooding. This illustrated that despite levees and other floodprotection measures, or some potential impacts of land use change along rivers, Aqua Monitor can detect high volumes of water, changes in river width between levees, and changes in natural inundation area (including flood retention ponds).

Water surface area change analysis at the global scale
In regions where changes could be measured with sufficient satellite images from Aqua Monitor, 29% showed an increase in water surface area in the flood plain, 41% showed a decrease, and 30% showed small or no changes. Additional trend data from Aqua Monitor provided a wider view of past water surface area changes in data-poor regions (Figure 1b). For example, increasing trends were found across the upper Amazon, and in several basins in East Asia. There were decreasing trends in southeastern Australia, eastern Brazil, Argentina, and the western part of Southeast Asia. These results are consistent with a study based on observed peak discharges in the period 1971-2010 (Gudmundsson et al., 2019). Our results showed similar upward and downward trends in the United States (Archfield et al., 2016) and Western Europe (Blöschl et al., 2019).
Considering the relatively high concordance between the observed annual maximum daily discharge and satellitederived flood area change, the results increase our understanding of trends in water surface area, which may shed light on past flooding events, especially in South America and Southeast Asia where in situ observations are limited. We expect that the continuing accumulation of satellite images will enhance our understanding of trends in flooding over a larger area in the future.

CONCLUSIONS
We investigated the utility of a satellite-derived water surface area dataset for detecting past floods on a global scale. The analysis demonstrated the ability of these data to provide information on flood area changes in data-poor basins, such as those in Asia and South America. The surface water area change derived from Aqua Monitor successfully detects direction of changes in flood in most analyzed basins. In addition to the discharge reanalysis, our results showed that satellite-derived surface water change can be utilized to detect occurrence of flooding or past changes in frequency or magnitude of flooding where in situ observation is unavailable. In regions where changes could be measured with sufficient satellite images, Aqua Monitor indicated that 29% of the basins experienced an increase in water surface area in the flood plain, 41% experienced a decrease, and 30% experienced small or no change (Figure 1).
Furthermore, comparison of the discharge reanalysis and observed discharge indicated that water management activities potentially affect flooding, particularly in highly regulated river basins. This is indeed the disadvantage of discharge reanalysis which has been often used to analyze river water change at global scale. As the observations and satellite-derived water surface area change showed relatively high concordance in many basins, additional information derived from satellite images would enhance our understanding of flood change where in situ observation is unavailable. Moreover, flood trend data obtained from different sources, as well as models incorporating the effects of water management activities, could clarify the mechanism underlying changes in flood trends.

ACKNOWLEDGMENTS
The flooding trend data for the sub-catchments used to draw the figures are available for research purposes upon request. This research was supported by the Environment Research and Technology Development Fund Figure 1. Detected trends in historical floods over the past few decades. a) Trends in the observed and simulated annual maximum daily discharge (% yr -1 ). Trends were plotted only when they were in the same direction as that seen in the satellite-derived inundation area data. b) Trends in inundation area derived from satellite data. Sub-catchments (gray and white colored) without significant trends (p < 0.05), few satellite images (i.e. <170 for both 1984-2000 or 2000-2013), or a small proportion of water pixels in the inundation area (< 15%) were excluded

SUPPLEMENTS
Text S1. Global daily discharge reanalysis Text S2. Dam index derived from Global Reservoir and Dam Database (GRanD) Text S3. Ability of the satellite-derived water surface area to detect historical floods Text S4. Catchment based global database of FLOod PROtection Standards (FLOPROS) Figure S1. Correlation of annual maximum daily discharge  derived from GRDC and discharge reanalysis at the 70 GRDC stations located at outlet region of catchment Figure S2. Scatter plot of annual maximum daily discharge  derived from GRDC (m 3 s −1 ) and discharge reanalysis (m 3 s −1 ) based on the 136 GRDC stations Figure S3. a) Dam index and b) total storage capacity of dam/reservoirs constructed after 2000 Figure S4. Global percentage tree-cover of each catchment Figure S5. Scatterplot of the median normalized inundation anomaly ((Ey -M)/SD × 100) for selected flood events (see Table SII) and their return periods. Ey is the total inundation pixels of each sub-basin (ATCIE) from the extreme flood of the year, SD and M is the standard deviation and mean of the ATCIE for the whole period Figure S6. Location of GRDC stations in the subcatchments. Cyan lines are the sub-catchment boundaries and red dots are GRDC stations Figure S7. Changes in the a) observed (units: % yr -1 ) and b) discharge reanalysis (middle panel) discharges (units: % yr -1 ), and c) water surface area (units: 10 -3 % yr -1 ), as derived from satellite data Figure S8. Changes in the a) observed (units: % yr -1 ) and b) discharge reanalysis (units: % yr -1 ), and c) in water surface area derived from satellite data (units: 10 -3 % yr -1 ) for 22 basins Figure S9. FLOPROS flood protection standards (Scussolini et al., 2016). The average return period (years) of the sub-catchment is shown