Temporary dependency of parameter sensitivity for different flood types

Floods represent one of the most severe natural disasters in the world. Floods have caused enormous loss of economies and casualties, which threaten the development of human society (Yoo et al. 2015; Garrote et al. 2016; Winsemius et al. 2016; Molinari et al. 2019). China is a flood-prone country, especially in the southern part. Thus, flood prevention awareness has continued to rise on the political agenda over the decades, accompanying a drive to ‘improve’ flood prediction and forecasts (Cloke & Pappenberger 2009; Yucel et al. 2015; Massari et al. 2018; Belabid et al. 2019). Accurate flood simulations are a solid foundation for flood forecasts (Kay et al. 2015; Cattoen et al. 2016; Koriche & Rientjes 2016; Unduche et al. 2018; Romali & Yusop 2021).

The concentration times in small- and medium-sized basins are commonly less than 24 h (Reynolds et al. 2017). Flood forecasts are thus required to provide simulations at sub-daily temporal resolutions (<24 h) (Ficchi et al. 2019; Hegdahl et al. 2019; Winter et al. 2019). However, in many regions, time series of observation data are only available at daily or coarser temporal resolutions, especially in developing countries. Parameter transferability across temporal resolutions is an important alternative for this situation. How to estimate the possibility of parameter transferability? What effect do temporal resolutions have on hydrological model simulations? Rafieeinasab et al. (2015) assessed the sensitivity of runoff simulation in urban catchments to spatiotemporal resolutions of precipitation input and found that a spatiotemporal resolution of 500 m and 15 min or higher is a good choice for runoff prediction. Sikorska & Seibert (2018) investigated appropriate temporal resolutions of precipitation data for runoff simulations, and the results revealed that hourly temporal resolution may not be always required. Ficchì et al. (2016) analyzed the impact of temporal resolutions of inputs on hydrological model performances in 240 test catchments, and the results demonstrated that three classes of model performance behavior were found at finer temporal resolutions, including significant improvement, insensitivity and degradation of performance. It can be concluded that inputs of finer temporal resolutions may not always obtain better model performance. As is known to us, model parameters have a significant effect on runoff simulation. Then, what roles do parameters play exactly in hydrological model simulations at different temporal resolutions? Hence, it is very important and worthwhile to elucidate the effect of temporal resolutions on model parameters and investigate the parameter transferability across different temporal resolutions to achieve accurate flood simulations.

The dependency of parameters on temporal resolutions has been studied in many literatures (Ostrowski et al. 2010; Kavetski et al. 2011; Bastola & Murphy 2013; Yang et al. 2016; Jie et al. 2018). Ostrowski et al. (2010) carried out the analysis of parameters depending on time steps with physically defined soil moisture and found out that functional relationships between parameter values and time steps are linear, nonlinear or remain constant. Jie et al. (2018) identified sensitive parameters at different temporal resolutions in the Xinanjiang model via the Morris method and found that parameters controlling the water balance and runoff routing are temporally dependent. Kavetski et al. (2011) reported that parameters relevant to slow response processes are relatively constant over a range of temporal resolutions, while those dominating fast response processes are temporally dependent. As Ficchì et al. (2016) concluded, the most important variables for model performance evolution across temporal resolutions are flood duration, precipitation duration and rainfall-runoff lag time. Therefore, it is valuable to explore the difference in model response to parameter perturbations across temporal resolutions among different flood types (classified by precipitation duration and other flood indices), which is rarely investigated in the literature. This study contributes to obtain accurate flood simulations at finer temporal resolutions for different flood types, which can provide useful supports for flood forecasting and flood risk management.

The common approach for investigating the model response to parameter perturbations across temporal resolutions is sensitivity analysis (SA). In addition to determining the dominant parameters, SA also contributes to gaining a deeper understanding of model processes (Guse et al. 2016; Gupta & Razavi 2018). The typical process of SA is using classical statistical performance metrics that quantify the closeness of the model simulation response to observed data, which does not consider the underlying processes shaping hydrograph (Guse et al. 2016). A more profound assessment of how dominant parameters affect hydrological processes promotes understanding of models greatly. Model diagnostic analysis is a good choice to accomplish this aim, which is capable of estimating the variation of model processes through mining more information contained in the hydrograph (Gupta et al. 2008; Yilmaz et al. 2008). Thus, diagnostic signature analyses are beneficial to quantify the effect of dominant parameters on hydrological processes. For the runoff generation process, discharge magnitudes can be partitioned through different segments of flow duration curve (FDC) (Yilmaz et al. 2008; Pfannerstill et al. 2014). FDCs can be used to evaluate changes of different discharge segments response to parameter perturbations. The relevant processes are distinct for FDC segments (Yilmaz et al. 2008; Cheng et al. 2012; Haas et al. 2016; Kamamia et al. 2019). During the high-flow segment, discharge is mainly controlled by a fast response process, such as surface runoff, while discharge in the intermediate flow segment is governed by the hydrograph relaxing phase. Slow response processes, including base flow and evapotranspiration, are more related to discharge in the low-flow segment. Thus, the application of FDC is actually valuable to analyze the sensitivity of hydrological processes to dominant parameters as each FDC segment captures different hydrological processes (Sawicz et al. 2011; Huang et al. 2017; Kamamia et al. 2019).

Given the above, the aim of this study is therefore fourfold: (1) to classify flood events into several flood types based on precipitation duration and other flood indices; (2) to identify sensitive parameters of the hydrological model at various temporal resolutions for flood types; (3) to explore the temporary dependency of parameter sensitivity for flood types; and (4) to investigate the effect of highly sensitive parameters on each FDC segment.

The Jinhua River basin is located in the Midwest of Zhejiang Province, East China (Figure 1). The basin above the Jinhua hydrological station is considered in this study and its catchment area is 5,996 km² (Pan et al. 2018). The study area is subject to Asian subtropical monsoon, which is characterized by abundant precipitation and high temperature in summer and dry and cold winters. The annual average temperature is 17 °C, and the elevation ranges from 29 to 1,296 m. The main soil types are clay loam, sandy loam and loam, and the land-use types of the study area are primarily croplands, mixed forest and grasslands (Pan et al. 2017). The annual mean precipitation is 1,404.9 mm, based on 50 years of precipitation data (1962–2011). Moreover, precipitation is strongly summer-dominant, occurring mostly from May to September. Because of the uneven temporal distribution of precipitation, the Jinhua River basin experiences many floods, including typhoon-induced and plum rain-induced.

The current event selected is discarded if one of the following criteria is confirmed: (1) if the duration of rising (declining) limb is shorter than 3 h, (2) if the event time period overlaps with a previously identified event time period and (3) if the cumulated precipitation amount before the discharge peak is lower than 2.5 mm. More criteria are omitted here and interested readers can refer to Ficchì et al. (2016) and Lobligeois et al. (2014). In total, 36 flood events are selected in the study area.

Many researchers have classified flood types focused either on catchment condition, flood damage, time period or meteorological condition (Viglione et al. 2010; Yoo et al. 2015; Turkington et al. 2016; Liu et al. 2019). The drawbacks of these methods are that they only focused on physiographic characteristics of catchments. In this study, we adopt the classification approach of flood type based on the flood process. This method focuses on the catchment state, catchment dynamics and atmospheric inputs, and no prior assumptions on which the process type is more likely to occur are made (Merz et al. 2006). This method classified floods into five flood process-oriented types, including flash floods (FFs), short-rain floods, long-rain floods, rain-on-snow floods and snowmelt floods (Merz & Bloschl 2003; Merz & Bloeschl 2009). According to the section ‘Study area’, snow is very rare in the study area, then flood types concerning snow are not considered. The detailed information of the other three flood types is shown as follows:

• 1.

  FFs are driven by short and highly intensive rainfalls and usually last less than 2 days, occurring mostly in the typhoon season. This event is often a local event, and thus major floods of this type only occur in a small catchment or a subcatchment.

• 2.

  Short rainfall floods (SRFs) caused by rainfall with short duration and a high intensity saturating parts of the catchment. Such floods mainly occur in the plum rain period. Depending on the rainfall patterns, this event can be of a regional or a larger local scale than FF.

• 3.

  Long rainfall floods (LRFs) induced by long lasting rainfall of several days or possibly weeks, including low intensity rainfall which gradually fills the storage capacity. Such floods mainly occur in the plum rain period. This type of flood event is usually of a regional range or an entire catchment.

A fuzzy decision tree method developed by Sikorska et al. (2015) is adopted here to classify flood events into three flood types, i.e., FF, SRF and LRF. As shown in Figure 2, the decision tree is built up of branches, nodes (queries), and ends up in leaves. Each leaf stands for one of the three flood types. The evaluation of tree is conducted by the examination of sequential queries corresponding to flood indices (shown in Table 1), called attributes here. A query stands for a test on an attribute and each branch stands for an outcome of the test.

The fuzzy decision tree method is a modification of the classical crisp tree (Solomatine 2003; Sauquet & Catalogne 2011). This method assumes that there is some unavoidable vagueness in the classification, as different processes may simultaneously contribute to flood formation. Hence, one event can be governed by mixed flood types. This method uses a quantitative description of the three types by assigning a degree of membership to each flood type. To allocate the degree of membership, the tree attributes are represented as soft thresholds in a similar way to the fuzzy concept as shown in Figure 3 (Pradhan 2013). These are defined as a threshold value (⁠⁠) with a certain range assigned (⁠⁠). The certain range is equal to 20% of the threshold ⁠. Afterwards, for each flood index (shown in Table 1) at each attribute test, a corresponding degree of acceptance that an event belongs to a certain flood type is calculated in the following way. If the input value is equal to ⁠, the degree of acceptance is assigned as 0.5. Then, if all flood indices fall into the corresponding threshold intervals, i.e., from to ⁠, the degree of acceptance is assigned between 0 and 1 assuming a linear interpolation (Figure 3). Moreover, the degree of acceptance is set as 0, if input values fall below the lower threshold ⁠. The degree of acceptance is set as 1, if input values fall above the upper threshold ⁠. Each branch is evaluated till the leaf in this way. The overall degree of acceptance for each flood type is calculated as the product of all degree values distributed at nodes along the branches and the sum of values from all leaves.

Therefore, after using the fuzzy decision tree method (Figures 2 and 3), we acquire a spectrum of all flood types with allocated degrees of their belonging to each flood type. Thus, the type assigned with the highest value is the dominant flood type (Sikorska et al. 2015).

DHSVM (distributed hydrology soil–vegetation model), a physically based hydrological model, provides a dynamic representation of hydrologic processes at the spatial scale of digital elevation model (DEM) data (the horizontal resolution of the DEM is typically 10–200 m) (Wigmosta et al. 1994; Wigmosta & Burges 1997; Wigmosta et al. 2002). Catchment is divided into computational grid cells based on DEM. Moreover, DEM is used to direct downslope water movement and extrapolate climatic data. Soil and vegetation properties are allocated to each computational grid cell. These properties vary spatially throughout catchment. At each time step, DHSVM offers simultaneous solutions to energy and water balance equations of every grid cell in catchment. Individual grid cells are hydrologically linked through surface and subsurface flow routing. In this study, the spatial resolutions are 200 m, and five temporal resolutions (1, 3, 6, 12 and 24 h) are used to investigate temporary dependency of parameter sensitivity. Version 3.1.1 of DHSVM is adopted in this study.

The DHSVM model consists of seven modules, including evapotranspiration, snowpack accumulation and snowmelt, canopy snow interception and release, unsaturated moisture movement, saturated subsurface flow, surface overland flow and channel flow. Evapotranspiration is presented by a two-layer canopy model with each layer partitioned into wet and dry areas. An individual grid cell may contain an overstory canopy and either an understory or bare soil. Modules concerning snow (snowpack accumulation and snowmelt, canopy snow interception and release) are not turned on here due to the fact that snow is very rare in the study area. Darcy's law is used to assess unsaturated moisture movement through multiple root zone soil layers (Domenico & Schwartz 1998). Saturated subsurface flow is routed by a cell-by-cell approach using either a diffusion or kinematic approximation (Wigmosta et al. 1994; Wigmosta & Lettenmaier 1999). Saturated subsurface flow is largely controlled by lateral hydraulic conductivity and an exponential decrease rate of lateral hydraulic conductivity with soil depth.

Surface overland flow can be routed using an explicit cell-by-cell approach or a unit hydrograph approach. The former method is adopted in this study. Surface flow is generated in a grid cell when meeting one of the following conditions: (1) throughfall occurs on a saturated grid cell (saturation excess runoff); (2) the input of throughfall exceeds the defined infiltration capacity (infiltration excess runoff) and (3) the water table rises above the ground surface (return runoff). The downslope movement of surface flow is based on a cell-by-cell method similar to the method used for subsurface flow. Flow in stream channels and road drainage ditches is routed using a cascade of linear channel reservoirs. Roads are not considered in this study due to the fact that detailed road map is not available. The area percentage of roads is very low in this study area. Despite that, it should be kept in mind that roads often intercept subsurface flow at road cuts and generate surface flow from compacted surface. Detailed description of DHSVM can be found in the DHSVM website (https://www.hydro.washington.edu/Lettenmaier/Models/DHSVM) and Wigmosta et al. (1994, 2002).

In this study, we use a 90 m DEM downloaded from the Shuttle Radar Topography Mission (SRTM) website (http://srtm.csi.cgiar.org/). The resolution of DEM is refined to 200 m due to computational burden. The soil and vegetation data are obtained from the Nanjing Institute of Soil Research of China and WESTDC Land Cover Products 2.0 (http://westdc.westgis.ac.cn), respectively. The soil classes are reclassified according to the U.S. Department of Agriculture (USDA) soil texture classification system needed in DHSVM. Soil types in the study area include clay loam (55.4%), sandy loam (16.5%), loam (15.8%), clay (7.3%), silty clay loam (4.6%) and water (0.4%) (Pan et al. 2017). The vegetation types of the study area are primarily croplands (36.7%), mixed forest (29.6%), grasslands (22.9%) and evergreen needleleaf forest (5.0%) (Pan et al. 2017).

The other data needed for DHSVM are morphology, stream network and soil depth, which are generated from Arcinfo macro language scripts (DHSVM website). Hourly climate data are available at 23 climatic stations (shown in Table 2). The climate variables include average air temperature, relative humidity, wind speed, precipitation, incoming shortwave radiation and incoming longwave radiation. The majority of these stations have a complete record of several climatic variables from 01/01/2006 to 08/31/2014. Hourly observed runoff data of 36 flood events for the period of 04/12/2006–07/01/2013 (via flood event selection from the section ‘Flood event selection’) are available at the Jinhua hydrological station (Figure 1). The limited shortage of climatic data is complemented through inverse distance interpolation from nearby stations. Runoff and climatic data at different temporal resolutions were generated by aggregating the hourly data to 3-, 6-, 12- and 24-hourly time series (Ficchì et al. 2016; Reynolds et al. 2017).

DHSVM parameters can be classified into constant, soil and vegetation parameters. The parameters concerning snow are set to model defaults, owing to the fact that snow is rare in the study area. Constant parameters usually are defined as constant values, depending on characteristics of the study area. In the original model setup, there were six soil types and nine vegetation types. In this study, we focused on soil/vegetation types whose area percentages are larger than 10%, i.e., clay loam, sandy loam, loam for soil types and croplands, mixed forest, and grasslands for vegetation types. The parameters for other soil/vegetation types are set similar to the dominant soil/vegetation types (clay loam and croplands) (Cuo et al. 2011). For each soil type, the number of parameters is the same, i.e., 13. For vegetation types, there are 10 related parameters for croplands and grasslands (only have understory canopy) and 19 parameters for mixed forest (has overstory and understory canopy). Hence, in total, 78 parameters are considered in this study (Table 3). The parameter ranges are determined via combining model defaults, published values of similar catchments, and local history record of study area.

A local sensitivity test (Cacuci 2003) using a stepwise, single-parameter perturbation approach of soil parameters and vegetation parameters is adopted in this study. This method has been successfully used to identify sensitive parameters for DHSVM simulation (Du et al. 2014). In this study, each parameter is varied by the following perturbations to produce 11 separate cases: −50, −40, −30, −20, −10, 10, 20, 30, 40, and 50%, plus the baseline value (see Table 3). The baseline value is derived from the best model parameter set by Pan et al. (2017), who proposed a two-step SA approach for analyzing model parameter sensitivity in continuous daily simulation in the Jinhua River basin.

Figure 4 shows the degree of acceptance of three flood types. As can be seen, half of these flood events are classified as a mixed flood type. This fact reveals that fuzzy decision tree considers the uncertainty in flood-type classification. Thus, this method is more reliable than crisp decision tree. The type with the highest degree of acceptance is the dominant flood type. SRF is the main flood type in the period 04/12/2006–07/01/2013, accounting for 50% of flood events. About 36% of flood events are classified as LRF and 14% of flood events are FF. As presented in Table 4, most events occurred in the period from May to October (32 events), which consist of plum rain period and typhoon period. Four events occurred in the other period. FFs are restricted to the period from May to October. The mean precipitation duration of LRF is the highest (154 h), followed by SRF (74 h) and FF (35 h).

As mentioned in the section ‘Model implementation’, we analyzed the sensitivities of 78 parameters in this study. Figure 5 shows SA results at 1 h temporal resolution for flood types. The parameter abbreviations shown in Figures 5–12 represent the parameters of croplands and clay loam. The parameters of other soil and vegetation types are not sensitive. Only sensitive parameters are displayed in Figure 5 (SAPC ≥ 1%). From this figure, it can be observed that 13 parameters are sensitive. The sensitivities are very different among three function objectives. For TD, 13 parameters are sensitive for FF, 12 for SRF, and 11 for LRF. For PD, 12 parameters are sensitive for FF, and 11 for SRF and LRF. For TPD, nine parameters are sensitive for FF, eight for SRF and four for LRF. It can be concluded that FF and SRF simulations are more easily affected by parameter perturbations based on the sum of SAPC from 13 sensitive parameters. Furthermore, highly sensitive parameters among three function objectives and their SAPC vary greatly for three flood types (see Table 5). As presented in Table 5, TD of FF and SRF are highly sensitive to EDR, P and f_c. TD of LRF is highly sensitive to P and f_c. Perturbations of f_c result in the largest SAPC of TD for three flood types, and the values are 72, 66 and 54% for FF, SRF and LRF, respectively. The highly sensitive parameters based on PD of FF, SRF and LRF are the same, i.e., K, EDR, P and f_c. P is the most sensitive parameter for PD, and its SAPC values for three flood types are more than 100%. TPD of FF and SRF is highly sensitive to K, EDR, P and f_c. However, TPD of LRF is only highly sensitive to P. This result reveals that TPD of FF and SRF is more sensitive to parameter perturbations. Moreover, it can be observed that soil parameters are always highly sensitive, which affects runoff generation greatly. P determines the maximum soil water content. K dominates the rate of water movement in the soil column and EDR describes the exponential decrease of K with soil depth. P, K and EDR, together, to a large extent, determine the partition of water between the soil column and stream channel, and routing time. f_c plays a dominant role in the unsaturated moisture movement module and has an impact on the amount of runoff. Overall, FF and SRF are more sensitive to parameter perturbations than that of LRF, and the numbers of sensitive and highly sensitive parameters are larger.

Figure 6 shows parameter sensitivity across different temporal resolutions, i.e., 1, 3, 6, 12 and 24 h. As SAPC caused by P is much larger than other parameters, the logarithm of SAPC is adopted for y-axis in Figure 6(a)–6(f). As most SAPC of TPD caused by vegetation parameters are 0, SAPC is used for y-axis in Figure 6(g)–6(i). The blue horizontal line stands for sensitive-dividing line. The red horizontal line stands for highly sensitive-dividing line. From this figure, it can be observed that parameter sensitivities vary greatly with temporal resolutions. It is revealed that temporary dependency occurs in parameter sensitivity (see Table 6).

There are large differences between three function objectives. For TD, all of sensitive parameters are more than 10%, which illustrates that all of them are highly dependent. For instance, TD of FF is sensitive to K at 1 h temporal resolution, while it becomes highly sensitive at 3 h temporal resolution. Moreover, the sensitivity of UCH for LRF changes from not sensitive at 1 h temporal resolution to sensitive at 3 h temporal resolution. The temporary dependency of this parameter is the highest and the is 68%. Its sensitivities of 1, 3, 6, 12 and 24 h are 0.2, 1.5, 3.0, 4.2 and 4.9%, respectively. For the most sensitive parameter, P for LRF, the sensitivities of 1, 3, 6, 12 and 24 h are 29.7, 46.3, 52.1, 46.5 and 35.0%, respectively. The temporary dependency of P for LRF is 24%. The fact reveals that highly sensitivity is not equal to highly temporary dependency. The temporary dependency of TD of SRF is the largest, followed by LRF and FF based on the average values of from all sensitive parameters. However, for PD, the temporary dependency of LRF is the largest, followed by SRF and FF. As presented in Table 6, the temporary dependency of parameter sensitivity based on TPD is much larger than TD and PD. From Figure 6(g)–6(h), it can be observed that most parameter sensitivities at 24 h temporal resolution are 0, mainly owing to temporal resolution aggregation and the short flood duration of FF and SRF. The parameter sensitivities are not similar for LRF, because the flood duration of LRF is much longer than that of FF and SRF. Based on the average value of TPD, the temporary dependency of LRF is much higher than that of SRF and FF. Taking three function objectives into consideration, the temporary dependency of parameter sensitivity from LRF is the highest, followed by SRF and FF.

Figure 7 shows percentage changes of PD to highly sensitive parameter (see Table 5) perturbations at temporal resolutions for flood types. Figure 8 presents PD variations caused by highly sensitive parameter perturbations at five temporal resolutions for three flood types. According to Table 6, these parameters are highly temporally dependent. As shown in Figure 7, absolute percentage changes of PD resulting from parameter perturbations are very different at five temporal resolutions. The larger parameter perturbations are, the bigger the differences among five temporal resolutions exist. For instance, when P from FF increases by 10% (50%), absolute percentage changes of PD are 4.9% (40.0%), 6.0% (36.6%), 4.3% (22.5%), 2.4% (11.1%) and 2.0% (6.0%) at 1, 3, 6, 12 and 24 h temporal resolutions, respectively. In general, highly sensitive parameter perturbations at finer temporal resolutions have much stronger effects on PD, except f_c. This fact is also confirmed by Figure 8(a)–8(i). The lines of PD at finer temporal resolutions are much steeper. For f_c, all absolute percentage changes of PD at five temporal resolutions are in a small range. The larger changes occur at coarser temporal resolutions, which is due to the fact that f_c dominates unsaturated moisture movement. Hence, variations of f_c at finer temporal resolutions do not cause considerable changes in PD, but in TD.

Figure 9 presents the FDC of simulated discharge response to highly sensitive parameter perturbations for representative events of FF (No. 20110619), SRF (No. 20101013) and LRF (No. 20060517) at 1 h temporal resolution. All the degrees of acceptance of these representative events to corresponding flood types are 1.0, and they are the typical event of three flood types. FDC is widely adopted to illustrate the capability of catchment to produce different magnitudes of runoff. To achieve this, the curve is partitioned into three segments, and the effects of perturbing highly sensitive parameters on three segments are investigated hereafter. The three segments include high-flow segment (0–0.2 flow exceedance probabilities, i.e., Q0–Q20), intermediate flow segment (0.2–0.7 flow exceedance probabilities, i.e., Q20–Q70), and low-flow segment (0.7–1.0 flow exceedance probabilities, i.e., Q70–Q100). In order to compare the effects of highly sensitive parameters on three segments clearly, Figures 10–12 display the percentage changes of low flow (Q95), intermediate flow (Q50), and high flow (Q5).

As shown in Figure 9(a)–9(c), perturbations of lateral saturated hydraulic conductivity (K) cause small variations of low flow (Q95), intermediate flow (Q50) and high flow (Q5) for three flood types. As shown in the first row of Figures 10–12, it is worth noting that low flow decreases with an increasing parameter value; however, high flow increases with an increasing parameter value. This is due to the fact that K controls the rate of water movement in the soil column. For the exponential decrease rate of K with soil depth (EDR), the effects of perturbations on low flow and high flow are larger than intermediate flow (Figure 9(d)–9(f)), which is owing to EDR describing the exponential decrease of lateral conductivity with soil depth. When EDR increases, the amount of subsurface runoff will decrease, which leads to decrease of high flow and increase of low flow (see the second row of Figures 10–12). Furthermore, both parameters (K and EDR) have much bigger effects on LRF than SRF and FF.

The sensitivities of three quantile flows to porosity (P) are very different for three flood types (Figure 9(g)–9(i)). This is because of the difference in attributes of flood types. Low flows of three flood types increase with increasing porosity, because the amount of infiltration increases with increasing porosity, leading to the increase of base flow (see the third row of Figures 10–12). High flow of three flood types decreases when porosity increases. This is mainly owing to the fact that the decrease of surface flow with increasing porosity. Porosity perturbations resulted in much larger changes in LRF than FF and SRF. For field capacity (f_c), its perturbation has a small effect on high flow, since that f_c plays a dominant role in the unsaturated moisture movement (Figure 9(j)–9(l)). Hence, considerable changes of low flow occur with perturbations of field capacity (see the last row of Figures 10–12). From these figures, it can be observed that intermediate flow of LRF changes are stronger than FF and SRF. This is similar to low flow. Overall, highly sensitive parameters affect low flow and high flow more than intermediate flow for three flood types. Highly sensitive parameter perturbations result in stronger changes in three FDC segments of LRF than FF and SRF.

Exploring the temporary dependency of parameter sensitivity for flood types provides a valuable reference for parameter transferability across temporal resolutions in flood simulations. This study investigated the dependency of parameter sensitivity on temporal resolutions for flood types. The results showed that parameter sensitivity at different temporal resolutions for three flood types showed, with considerable variation to be highly dependent on the temporal resolution of data. Moreover, the temporary dependency of parameter sensitivity from LRF is the highest, followed by SRF and FF. The generality of our results is so far restricted to one distributed hydrological model and one catchment. For catchments at alpine areas, their flood types include snowmelt flood and rain-on-snow flood types (Merz & Bloeschl 2009; Nied et al. 2014; Sikorska et al. 2015), which were not considered in this study. Although the results in this study are beneficial to understand the parameter transferability for rainfall-dominated flood types, a systematic investigation of parameter transferability for snow-related and ice-related flood types can be followed.

The classification of 36 observed events into three flood types illustrates that diverse flood-generating mechanisms exist in the Jinhua River Basin. The classification is based on multiple flood indices, including precipitation duration, occurrence timing, precipitation intensity, and catchment wetness. Similar to many natural phenomena, the boundaries among different flood processes are not sharp (Merz & Bloschl 2003). One flood event may be a combination of various flood processes. For instance, event No. 20060614 is governed by FF and SRF. Moreover, event No. 20080616 is a combination of SRF and LRF. The fuzzy decision tree method considers the unavoidable vagueness in classification through a quantitative description of various types by assigning a degree of membership to each flood type. In this study, only one catchment in the humid region and a limited number of flood events were classified into FF, SRF and LRF. For catchments at alpine areas, their flood types include snowmelt flood and rain-on-snow flood types (Merz & Bloeschl 2009; Nied et al. 2014; Sikorska et al. 2015), which were not considered in this study. Further research is necessary to investigate the applicability of the fuzzy decision tree method in flood classification in other regions.

The comparison of SAPC from PD showed that 12 parameters are sensitive for FF, and 11 parameters are sensitive for SRF and LRF. The highly sensitive parameters for three flood types are the same, i.e., K, EDR, P and f_c. Beside the SAPC of PD, the other two function objectives also showed similar results. The high sensitivity of K and EDR is due to the fact that K and EDR control the rate of water movement in the soil column. The PD usually consists of lateral surface flow, lateral subsurface flow and base flow (Wigmosta & Lettenmaier 1999; Du et al. 2014). From Table 2, it can be observed that there are obvious differences in the SAPC of PD for three flood types. For instance, the SAPC of parameter K from SRF is more than twice as much as FF. This is due to the fact that shallow lateral subsurface flow accounts for a larger proportion of SRF than FF, which can also explain the difference between sensitivity of EDR from SRF and FF. The high sensitivity of P is due to the fact that P determines the maximum soil water content. Soil with high porosity can store more water, which has a direct effect on lateral surface flow. f_c is used to describe available water for subsurface soil column. Because the lateral subsurface flow is not dominant in PD, f_c is less sensitive than P for PD. However, f_c is more sensitive than P for TD, which reveals the composition of TD. Therefore, flow component research is much demanded to get further recognition to difference in the parameter sensitivity among various flood types. We can obtain the proportion of lateral surface flow, lateral subsurface flow and base flow, through runoff component research. Therefore, the sensitivity of various flow components of different flood types to model parameters can be achieved. Subsequently, we can get much clearer understanding of the effect of parameter perturbations on flood simulations.

A local sensitivity test was applied in this study, which means that only one parameter was perturbed at each time. For example, field capacity remained constant, when porosity was perturbed. This basically does not happen in the real world. A global sensitivity test, such as Sobol's method, is more reasonable. However, global sensitivity tests have a high demand on model runs and computational capacity (Du et al. 2014). In another study (Pan et al. 2017), a two-step global SA was used to investigate parameter single sensitivity and interactions between different parameters, and parallel computing was applied to improve computational efficiency. Although a local sensitivity test can be enough to achieve the goals of this study, a global sensitivity test is expected to further explore the sensitivity of parameters in the hydrological model in the future study.

Parameter sensitivity at different temporal resolutions for three flood types is shown to vary considerably (Table 6). Four parameters such as K, EDR, P and f_c are highly sensitive and highly temporal dependent, which control the dominant runoff mechanisms of the study area, fast flow processes (i.e., surface and near-surface flow). This confirms the finding from a previous study by Jie et al. (2018) that sensitivities of runoff-related parameters of the Xinanjiang model are temporally dependent in a tributary of the Minjiang River, China. The response time of hydrological processes in a certain watershed, such as runoff formation processes, is usually stable. Therefore, it is appropriate that the temporal resolution of data matches the response time. Parameters relating to slow flow processes remain constant over a range of temporal resolutions, while parameters relating to fast flow processes converged toward stable estimates as the temporal resolution approached the timescale these processes present (Kavetski et al. 2011). As the runoff data averaged at coarser temporal resolutions, a loss of information on the dynamics of fast flow processes occurs (Ostrowski et al. 2010). Hence, parameters describing fast flow processes are temporally dependent, which illustrates the temporal dependency of parameter sensitivity as shown in Table 2. As presented in Figure 6, it can be observed that parameter sensitivity is not becoming relatively constant, when temporal resolutions range from 24 to 1 h. This fact reveals that higher temporal resolution is desired for further research. Meanwhile, other studies have explored the impact of temporal resolutions on model simulations and parameter transferability across temporal resolutions (Ren et al. 2016; Reynolds et al. 2017; Chouaib et al. 2018). We can obtain temporary dependency of parameter sensitivity for different flood types through this study. Subsequently, highly sensitive and highly time-dependent parameters need to pay more attention in the parameter transfer process among different temporal resolutions. Hence, this study is beneficial to improve efficiency of parameter transfer process and not waste efforts to insensitive and independent parameters.

The perturbations of highly sensitive parameters cause different variations on three flow segments for FF, SRF and LRF. For low flow, the perturbations of K and EDR have a greater influence on FF than SRF and LRF. This is because subsurface flow accounts for a larger proportion in low flow of FF. The differences among effects of the perturbations of K and EDR on intermediate flow and high flow of three flood types are also owing to the different composition of runoff. Similarly, perturbation of P means that the larger variations in high flow of LRF than SRF and FF are also due to the bigger proportion of subsurface flow in high flow. Flow components research is beneficial to get the proportional composition of base flow, subsurface flow and surface flow in three flow segments (low flow, intermediate flow and high flow). For instance, Poulain et al. (2018) partitioned runoff into diffuse and quick runoff components through the master recession curve. Saraiva Okello et al. (2018) adopted chemical hydrograph separation to quantify runoff into base flow and quick flow in monthly and annual scales. Based on the runoff components result, we can conduct the sensitivity of different runoff components in three flow segments to model parameters. Therefore, a clear understanding of the impact of model parameters on runoff components in low flow, intermediate flow and high flow can be obtained. This clearer understanding will contribute to recognition of the difference among parameter sensitivity in various quantile flows of FF, SRF and LRF.

In this study, 36 flood events in the Jinhua River basin, East China, were firstly classified into three flood types using the fuzzy decision tree method. The sensitivity of DHSVM flood simulation for flood types to parameter perturbations was identified through a local sensitivity approach. The temporary dependency of parameter sensitivity for flood types was finally investigated. The key findings of this study are summarized below:

• 1.

  The classification results show that the dominant flood type in the analysis period is SRF (its percentage is 50%), followed by LRF (36%) and FF (14%). 89% of flood events occurred in the period from May to October, which consists of both plum rain period and typhoon period.

• 2.

  The SA results indicate that runoff simulations in FF and SRF are more sensitive to parameter perturbations than that of LRF, and the numbers of sensitive and highly sensitive parameters are larger. The sensitive parameters mainly consist of five soil parameters (K, EDR, P, f_c, and w_p) and eight vegetation parameters (ULAI, UALB, UCH, Rsmax, URsmin, MT, VPD and RZD). The highly sensitive parameters mainly include four soil parameters, i.e., K, EDR, P, and f_c.

• 3.

  The results of parameter sensitivity across various temporal resolutions reveal that sensitive parameters are highly dependent on temporal resolutions. The larger parameter perturbations are, the higher temporary dependencies are. The temporary dependency of parameter sensitivity from LRF is the highest, followed by SRF and FF.

• 4.

  The results of diagnostic signature analysis elucidate that the effects of parameter perturbations on FDC of different flood types are various. Highly sensitive parameters affect low flow and high flow larger than intermediate flow. Highly sensitive parameter perturbations result in stronger changes in three FDC segments of LRF than FF and SRF.

• 5.

  This study provides a valuable reference for selecting parameters in the subsequent parameter transfer process in catchments with hydrological similarity. Moreover, this study contributes to further reduce uncertainty in parameter estimation by considering parameter temporary dependency.

We acknowledge the financial support from the National Nature Science Foundation of China (51909233), the National Key Research and Development Plan ‘Inter-governmental Cooperation in International Scientific and Technological Innovation’ (2016YFE0122100), and the Key Project of Zhejiang Natural Science Foundation (LZ20E090001). National Climate Center of China Meteorological Administration, Zhejiang Provincial Metrological Administration, and Zhejiang Provincial Hydrology Bureau are greatly acknowledged for providing meteorological and hydrological data used in this study. The valuable comments and suggestions from editor, and two anonymous reviewers are greatly appreciated.

Data cannot be made publicly available; readers should contact the corresponding author for details.