Causal effects of dams and land cover changes on flood changes in mainland China

Quantifying the effects of human activities on floods is challenging because the knowledge and observations toward the effects are limited. Many previous methods fail to isolate different effects and reduce the uncertainty caused by small samples. We use panel regressions to derive the sensitivity of annual maximum discharges (Q) to the changing values 10 of three human factors: urban areas, cropland areas, and reservoir indexes for large and middle dams. We also test whether the effects increase or decrease with increasing initial values of human factors. This method is applied in 2739 hydrological stations in China. Results show that a 1% increase of urban areas causes around a 2.9% increase of Q. Cropland areas have no significant effect on Q. Reservoir index has a decreasing effect: a 1 unit increase of reservoir index causes a decrease in Q from 23.1% to 5.4% for catchments with initial reservoir indexes from 0 to 5. From 1992 to 2017, increasing urban areas 15 cause more than 10% increases in Q in 10.4% of the 2739 catchments, most of which are located in the North Plain of China. From 1960 to 2017, increasing reservoir indexes cause a more than 10% decrease of Q in 53.4% of 777 catchments with at least one dam. Among 1074 catchments with limited impacts from urban areas and reservoir indexes, 210 (19.6%) catchments have more than 10% unexplained decreasing rates in Q per decade during 1960-2017, and 62.4% of the 210 catchments are located in the middle and down streams of the Yellow River Basin and the upper streams of the Haihe River 20 Basin. This study extends the panel regression method in hydrology and sheds light on the attribution of flood changes on a national scale.


Introduction
River flooding is one of the most severe disasters in the world. China experiences tremendous damages from floods in the past decades with expanding urban areas, a booming economy, and increasing populations (Du et al., 2019;Kundzewicz et 25 al., 2019). Sharply changing flood characteristics make flood risk management more difficult. According to a national investigation of flood peak changes in China conducted by Yang et al. (2019), abrupt changes due to human activities are the predominant mode of flood changes. Understanding how floods change in a changing environment helps flood risk management in the future. Therefore, a quantitative attribution of flood changes is urgent on a national scale for policy decisions. 30 To detect flood changes and pinpoint the underlying reasons, scientists need to answer the following questions: 1. whether a factor affects floods? 2. If the effect presents, how strong is the effect? The drivers of flood changes can be classified into three categories: atmospheric factors, catchment factors, and river factors (Merz et al., 2012;Blöschl et al., 2015).
Atmospheric factors refer to the meteorological forcing of water fluxes such as natural climate variability and anthropogenic climate change; catchment factors refer to the alternating physiographic characteristics of catchments, such as land cover 35 changes; river factors refer to hydraulic infrastructures that change river morphology and flood routing, such as dams and channelization (Merz et al., 2012;Blöschl et al., 2015). Catchment and river factors are mainly attributed to human activities, which attract increasing attention in hydrological systems in the era of "socio-hydrology" (Di Baldassarre et al., 2019;Müller and Levy, 2019). However, quantifying human impacts on floods is challenging for the following reasons. Firstly, due to the highly unpredictable human behaviors, we have limited knowledge to reproduce the process of how human 40 activities affect floods (Pande and Sivapalan, 2017). For example, the expansion of cropland and urban areas not only casts deterministic effects on floods through changing soil physics and surface roughness but also brings uncertain effects through irrigations and water diversions. We consider these effects "uncertain" because they are related to unknown human decisions.
Secondly, the observations of human activities are limited (Pande and Sivapalan, 2017). In the example above, many regions lack long-term and large-scale data of soil physics, roughness, irrigations, and water diversions, which are highly dependent 45 on a high-cost network of in-site measurements.
Previous studies have used three methods to quantify human impacts on floods. A first method is a model-based approach.
This method regards the impacts of human activities as either the difference between actual observations and the model simulations of floods (Viglione et al., 2016;Lu et al., 2018) or flood changes with time-varying model parameters (Peña et al., 2016;Umer et al., 2019). However, this method suffers from limited model accuracy. The second method is a paired-50 catchment experiment. This method either compares the floods before and after human impacts in one catchment or compares floods in two groups of catchments with and without human impacts (Prosdocimi et al., 2015;Hodgkins et al., 2019). However, the comparisons above cannot rigorously isolate multiple impacts on floods since we cannot actually control everything except one targeted human factor (Runge et al., 2019). The third method is empirical variable dependence, i.e., using regressions or non-stationary probability distributions to link human factors to flood characteristics (FitzHugh et 55 al., 2011;Prosdocimi et al., 2015;Bertola et al., 2019;De Niel and Willems, 2019). The third method is cost-efficient for large-scale studies, but it has two problems. Firstly, to derive the causal effects of human factors, all confounders -which correlates with human factors and floods at the same time --should be explicitly accounted for in the empirical relationships. However, defining numerous variables to represent confounders may be an endless task. For example, climatic confounders are ambiguous because floods are caused by different climatic factors (e.g., long rainfall, short rainfall, 60 snowmelt, and rain on snow) in different regions (Stein et al., 2020;Yang et al., 2020a;Merz et al., 2020). Therefore, in a large-sample study, we do not have a unified regression form to control all possible variables for all catchments. Secondly, empirical methods require sufficient data for robust statistical inference, while flood samples are rare for each catchment. Panel regression (Steinschneider et al., 2013;Wooldridge, 2016) solves the problems of the empirical method in two ways.
Firstly, panel regression adds virtual variables to the regression to represent a fixed individual or regional effect 65 (Steinschneider et al., 2013). In such a way, the regression can account for the effects of ambiguous confounders that are constant in time or region. Secondly, panel regression pools all samples into one model and trades space for time (Steinschneider et al., 2013). Therefore, the regression result is more reliable even with short flood records for each catchment. Although panel regression is a tool in economics, it has been introduced in hydrology to estimate the effects of forests on floods (Ferreira and Ghimire, 2012), urbanization on runoff coefficients (Steinschneider et al., 2013), dams on 70 streamflow (McManamay et al., 2014), rainfall on streamflow (Bassiouni et al., 2016), deforestation on streamflow (Levy et al., 2018), urbanization on floods (Blum et al., 2020), and rain/snow fraction on floods (Davenport et al., 2020). However, these studies only focused on one factor at one time. Considering more human factors can provide a more comprehensive picture of the human impacts on floods. In addition, only Blum et al. (2020) and Davenport et al. (2020) tested the nonlinear effects of factors. Previous studies rarely examined whether the effects increase or decrease with increasing initial values of 75 human factors.
In this study, supported by an unprecedentedly large dataset of Chinese floods from 2739 streamflow gauge stations, we quantify the national average sensitivities of annual flood peaks to changing urban areas, cropland areas, and reservoir indexes for large and middle dams using panel regression. We also design a workflow to test whether effects increase or decrease for catchments with increasing initial values of the targeted factor. The causal effects of factors distinguish the 80 flood changes explained and unexplained by the three human factors in recent decades. This study is organized as follows.
Section 2 introduces methods. Section 3 describes the data. Section 4 presents the results. Section 5 discusses the methods and the insights gained by this study. Section 6 gives conclusions.

Causal map of flooding 85
Causal maps depict the dependency relationship between variables, and they help discover confounders and focus on the causal effects of different factors when fitting a regression model (Pearl and Mackenzie, 2020). A causal effect is defined as the sensitivity of floods to a factor when all possible confounding variables are controlled. Similar to Blum et al. (2020), we draw a causal map of flooding in Fig. 1. This study estimates the causal effects of the changes in dams, urban areas, and cropland areas on floods, as the three dashed lines show in the figure. Variables lying above the dashed lines are unknown or 90 unobserved mediators. Urban areas and cropland areas are interrelated because they may change into each other during the process of land cover change. We consider two major confounders. The first confounder is the regional time-varying confounder, which means that it changes over time but keeps spatially constant in a region. Climate is such a variable since it varies temporally and has spatial homogeneity. The trends of climate may correlate with human activities. For example, decreasing annual precipitation exacerbates water shortage and may therefore promote the reservoir constructions or the 95 implementation of the Grain for Green Project. The second confounder is the individual time-invariant confounder, which means that it keeps constant in time but varies by different catchments. This confounder is mainly represented by the characteristics of catchment landscapes, e.g., topography, soil types, geology. They may also correlate with human activities.
For example, urban areas are likely built on flat and plain catchments.

Design of panel regression
Panel regression is a statistical technique for panel data (Steinschneider et al., 2013;Wooldridge, 2016). Panel data are 105 observations on several subjects in different periods. Panel regression controls the constant effects of each subject or each period to mitigate regression bias due to omitted variables. Panel regressions in this study are extended from the equation in Blum et al. (2020) and are presented in Eq. (1) and (2) as follows. (1) (2) 110 , is the annual flood peak of catchment in year (m 3 • s −1 ).
, is the urban percentage of catchment in year (%).
, is the cropland percentage of catchment in year (%).

catchment in year ;
( ) is the degree of regulation of reservoir in catchment , which is the ratio between the storage capacity and total annual flow of the reservoir; ( ) is the upstream area of reservoir ; is the area of catchment . is a region dummy which equals 1 or 0. is a year dummy which equals 1 or 0.
(1) and (2) are the time-invariant constant 115 effects of catchment in Eq. (1) and (2) respectively. , (1) and , (2) are the time-varying constant effects of region in year in Eq. (1) and (2) respectively. , (1) and , (2) are the model residuals in Eq. (1) and (2) respectively. The response functions 1 (•), 2 (•), and 3 (•) represent various response types of to different factors. We assigned , , and into two regressions because their data sources had different temporal lengths. Although may correlate with and , the effects of dams and land cover on floods can be derived independently since we have controlled their common drivers (Pearl 120 and Mackenzie, 2020) in each equation, i.e., the regional time-varying term and the individual time-invariant term.
A region consists of a group of spatially coherent catchments with a similar climate. Unlike Blum et al. (2020) who used predefined physiographic regions, we delineated regions by using the partitioning around medoids (PAM) algorithm (Reynolds et al., 2006) based on the distance matrix of all catchments defined as follow: where ( , ) is the distance of Köppen-Geiger class (Beck et al., 2018) ratios between catchment and ; ( ) is the area percentage of Köppen-Geiger class in catchment ; ( , ) is the standardized distance between the geometric centers of catchment and ; ℎ.
( , ) is the spherical distance on the earth between the geometric centers of 130 catchment and .
The effect of a factor on , i.e., the percentage change in given a fixed change in , is expressed as:  The mathematical assumptions of the panel regressions in this study are as follows: 1. No other important time-varying sub-145 regional variables that significantly affect both human factors and floods; 2. No interactions between human factors and regional or individual characteristics that produce significant spatially heterogeneous effects. The regressions and statistical tests were performed in R (R Core Team, 2019) using packages lfe (Gaure, 2019), lmtest (Hothorn et al., 2020), and sandwich .

Flood change quantification 150
After obtaining the coefficients in the response function (•), we can derive actual flood changes attributed to factor changes in a long period. The regressions in this study calculate a common percentage change in all flood peaks rather than heterogeneous changes of different flood events, according to Eq. (6). Therefore, the percentage change in annual maximum discharges for catchment from year 1 to 2 can be written as: To examine the changes in flood peaks unexplained by changing urban areas, cropland areas, and dams, we selected catchments with less than 10% changes in flood peaks due to those factors respectively. Specifically, for a factor , we selected catchments with �exp� � , 2 � − 0� − 1� < 10% where 2 was the most recent year of the data. For the annual maximum discharges of each selected catchment, we used Theil-Sen slope estimator (Theil, 1992) to derive the trend and Mann-Kendall test (Mann, 1945) to derive the statistical significance. 160 3 Data

Streamflow data
Annual maximum instantaneous discharge data in 2739 streamflow gauge stations were obtained from the Ministry of Water Resources in China (http://www.mwr.gov.cn/english/). Figure 3 shows the outlet locations of all stations. The catchment areas are from 1 km 2 to 1,705,383 km 2 with a median of 1,660km 2 . Catchment boundaries were extracted using  Hydro hydrography data (Yamazaki et al., 2019). Differences between extracted catchment areas and reported areas were less than 20% for all catchments. We only used data from 1960 to 2019 because less than 1,000 stations had available data before 1960. Notice that a few stations in the northeast lie outside mainland China. They were not excluded from this study because all other data were globally available. The 1-km resolution data of Köppen-Geiger climate classes were obtained from Beck et al. (2018). 170

Land cover and dams data
Land cover maps were obtained from the CCI-LC product produced by the European Space Agency ( dams. For simplicity, we use "dams" to represent large and middle dams in the rest of the paper. We obtained the locations, upstream areas, storage capacities, and total annual flows of each dam from the database. The reservoir index can be calculated with the information above.

Catchment selection for regression setup
We selected catchments with sufficient numbers of annual maximum discharges ( ) to fit Eq. (1) and (2), as shown in Table  195 1. For Eq. (1), the CCI-LC data were available in 1992-2018, and thus, we selected 1644 catchments with at least 20 years of data in this period. For Eq. (2), the discharge data were limited before 1960 and the GRanD data were available until 2017, and thus, we selected 1744 catchments with at least 30 years of data in 1960-2017. The number of catchment groups in Section 2.2 had no optimal value. In light of the number of selected catchments in the regression models, we set to be 10, 20, 30, 40, 50, and 60 to test the robustness of the models. We premised that groups were too small when ≥ 70, since the 200 average number of catchments in a group was smaller than 25 for Eq. (1) and (2). ( � , � = , and > 0), which means a fixed percentage increase of brings a fixed percentage increase of no 210 matter how large the initial is. exhibits no effect. exhibits a negative and decreasing effect ( � , � = , 1/2 and < 0), which means a fixed increase of brings a lower percentage decrease of with a larger value of initial . The details about the regression coefficients (e.g., values, standard deviations, and p values) in the effect testing process (Fig. 2) can be seen in Table S1 and Table S2. The maps of catchment groups for all values can also be seen in Fig. S1 and Fig. S2.  are relatively consistent when the number of catchment groups ≥ 30. We regard the average value ∆ = 2.9% for ∈ 220 {30, 40, 50, 60} as the sensitivity of to . has no significant effect on , therefore we do not calculate the corresponding sensitivity. Figure 5 shows the percentage change in caused by 1 unit increase of according to Eq. (6).      (8.8% of 777) have more than 25% decreases of attributed to increasing . Spatially, the impacts of dams on floods are larger in northern basins (the Huaihe River Basin, the Haihe River Basin, the Yellow River Basin, and the Songhua and Liaohe River Basin) than that in southern basins (the Yangtze River Basin, the Southeast River Basin, the Southwest River Basin, and the Pearl River Basin). In the northern basins, increasing leads to more than 10% and 25% decreases of in 66.1% and 20.0% catchments, respectively. By comparison, in southern basins, increasing leads to more than 10% and 25% 255 decreases of in only 45.4% and 2.0% catchments, respectively. Basin. Such regional coherence of similar trends cannot be found in other regions. https://doi.org/10.5194/hess-2020-609 Preprint. Discussion started: 26 November 2020 c Author(s) 2020. CC BY 4.0 License.

Strengths and limitations of panel regressions 275
We use panel regressions to derive the causal effects of urban areas, cropland areas, and dams on annual maximum discharges across mainland China. In this study, the panel regressions exhibit the following strengths. 1. We obtain a nationally generalizable sensitivity of floods to each human factor. This sensitivity helps understand the overall added risks of specific human activity on floods on a national scale. In addition, with quantitative sensitivity, scientists are able to select catchments with limited impacts of dams and land cover changes for studying the effects of climate change, e.g., Blöschl et 280 al. (2019). 2. Compared with previous studies using panel regressions in hydrology (Ferreira and Ghimire, 2012;Steinschneider et al., 2013;McManamay et al., 2014;Bassiouni et al., 2016;Levy et al., 2018;Blum et al., 2020;Davenport et al., 2020), we take a further step by considering multiple types of human impacts simultaneously and distinguishing their increasing or decreasing effects. Blum et al. (2020) and Davenport et al. (2020) considered non-linear forms of response functions for the targeted factors, but they did not distinguish increasing and decreasing effects. These improvements provide 285 a more comprehensive understanding of human impacts on floods.
The limitations are as follows. 1. The assumptions in the regressions are difficult to test. As stated in Section 2, we assume no more important time-varying sub-regional confounders and no interaction terms between human factors and regional or individual characteristics. Testing these assumptions requires detailed information about catchment characteristics such as topography and geology. Moreover, adding too many variables into the regressions will decrease model interpretability.

290
The method cannot distinguish the heterogeneous effects of human factors on different floods. As stated in Section 2.3, the method derives a common percentage change in all flood peaks given changing human factors, which means no changes in coefficients of variation. However, practically, the variability of floods may change by human activities. For example, reservoirs tend to regulate extreme floods but omit small floods. 3. This study does not comprehensively assess the effects of total human impacts on floods. We omit many other human factors due to the lack of data. For example, the data about water 295 diversion, irrigation, channelization, and afforestation on a national scale are currently not available to the public.

Consistency with knowledge and other large-sample studies
We detect a stable positive effect of urban areas on floods. In theory, expanding urban areas magnify floods in two major ways. Firstly, natural soil grounds are replaced by impervious surfaces, which lead to more rainfall water appearing on the surface rather than infiltrating into the soil (Villarini and Slater, 2017). Second, urban areas have smooth surfaces, where 300 floods propagate faster and become more flashy (Mogollón et al., 2016). This study finds a 2.9% increase in annual maximum discharges given a 1% increase in urban areas. This finding accords with the result from a national investigation in the US (Blum et al., 2020), which reported a 3.3% increase in annual maximum discharges based on panel regressions.
The cropland areas impose no significant impacts on floods according to our results. Theoretically, expanding cropland areas affect floods in many ways. For example, during agricultural practices, soil depths may decrease due to erosion while increase due to soil compaction (Rogger et al., 2017). Cropland may also bring artificial drainages that lower groundwater tables (Rogger et al., 2017). Some effects may be offset by others, which masks the relationship between cropland areas and floods. Similar to our result, Bertola et al. (2019) found that agricultural land-use intensification rarely caused flood changes in 95 catchments of Austria using covariate-based non-stationary flood probability distributions. To our knowledge, largesample studies are limited on the relationship between cropland and floods. Therefore, more detailed in-site investigations 310 are required to uncover the causal chain from cropland changes to flood changes.
This study suggests that dams have a negative decreasing effect on floods. Generally, dams buffer water during floods and thus decrease flood peaks. More dams may not necessarily decrease floods at a constant rate because existing dams with sufficient storage capacities are already capable to control floods. This effect was confirmed by Wang et al. (2017), who used detailed conceptual models of reservoir regulation and found that the mean annual floods had a slowing decrease with 315 increasing reservoirs. In a large-sample study on 4859 catchments in the US (FitzHugh et al., 2011), median annual 1-day maximum flows were estimated to decrease by more than 20% when the storage ratios, i.e., the total storage capacity of upstream dams divided by average annual runoff, were larger than 1. If a dam with storage capacity equaling the annual runoff is established at the outlet of the catchment without any dam before, both the reservoir index defined in this study and the storage ratio defined by FitzHugh et al. (2011) increase from 0 to 1. In this special case, the annual maximum discharges 320 decrease by 23.1% in this study, comparable to the 20% decrease from FitzHugh et al. (2011). It is noteworthy that this study only focuses on the effects of human factors on annual maximum discharges. Generally, the effects are larger for less frequent floods. Zhao et al. (2020) investigated floods in 1403 catchments in the US and found a decrease of 100-year floods by more than 60% in 47% of catchments with a dam upstream.

Insights toward a national investigation of flood changes 325
This study takes the first step to explain flood changes quantitatively on a national scale in China. In this study, urbanization and dam constructions significantly change annual maximum discharges in the middle and down streams of the Yellow River Basin and the Haihe River Basin, where step changes were detected by Yang et al. (2019). As a major human residence with a high density of population, the North Plain of China experiences fast urbanization in recent years (Du et al., 2018), which brings larger flood risks on lives and properties. In addition, the degree of dam regulation is larger in northern 330 China because the annual runoff is much smaller than that in wet southern China. In this study, after removing the catchments with nonnegligible impacts of urbanization and dams, the unexplained decreasing trends occur in the middle and down streams of the Yellow River Basin and the upper streams of the Haihe River Basin, where decreasing trends were also derived by Yang et al. (2019). Yang et al. (2019) interpreted these trends as the results of soil conservation practices (Bai et al., 2016) and decreasing extreme rainfall Wu et al., 2016). Besides, other reasons include decreasing soil 335 moisture (Cheng et al., 2015;Yang et al., 2020a) and the impacts of cascade small soil-retaining dams (Yang et al., 2020b).
It indicates that the impact factors of floods are complex in this region and further studies are required.

Conclusions
We conducted a data-based analysis on the causal effects of human impacts on floods using a panel regression on a national scale, based on annual maximum discharges (Q) from 2739 stations in China, CCI-LC land cover data, and GRanD dam data. 340 Furthermore, we derived nationally generalizable information about the sensitivity of Q to human factors, namely the changes in urban areas, cropland areas, and reservoir indexes for large and middle dams, and then determined the explained and unexplained changes by the human factors based on the sensitivity. The major findings are as follows.
Urban areas have a positive and stable effect on floods, i.e., a 1% increase in urban areas causes a 2.9% increase in annual maximum discharges. Cropland areas have no significant effect on Q. Reservoir index has a negative and decreasing effect 345 on Q, i.e., the decrease of Q caused by a 1 unit increase of reservoir indexes ranges from 23.1% to 5.4% corresponding to initial reservoir indexes from 0 to 5.
From 1992 to 2017, more than 10% of increases in Q were caused by increasing urban areas in 10.4% of the 2739 catchments. These catchments are mainly located in the North Plain of China, especially in the Huaihe River Basin and the middle and down streams of the Haihe River Basin. From 1960 to 2017, among the 777 catchments with at least one dam, 350 53.4% have more than 10% decreases in Q caused by increasing reservoir indexes. Spatially, the impacts of dams on floods are larger in northern basins, including the Huaihe River Basin, the Haihe River Basin, the Yellow River Basin, and the Songhua and Liaohe River Basin, where 66.1% catchments have more than 10% decreases of Q attributed to increasing reservoir indexes. Among 1074 catchments with less than 10% changes in Q caused by urban areas or dams, 210 (19.6%) catchments have a more than 10% decrease per decade during 1960-2017, and 62.4% of the 210 catchments are located in 355 the middle and down streams of the Yellow River Basin and the upper streams of the Haihe River Basin.
This study extends the panel regression method to quantify the effects of multiple human factors on floods, which helps understand the causes of flood changes on a national scale in China. Future studies may collect more data to consider more human factors and quantify the effects on different return periods of floods.

Data availability 360
Annual maximum discharge data are obtained from the Water Resources Information Center of the Ministry of Water Resources in China (http://www.mwr.gov.cn/english/). CCI-LC data are obtained from the ESA Climate Change Initiative -Land Cover project 2017 (http://maps.elie.ucl.ac.be/CCI/viewer/download.php). GRanD data are obtained from Global Dam Watch (http://globaldamwatch.org).