Why do we have so many different hydrological models? A review based on the case of Switzerland

Hydrology plays a central role in applied and fundamental environmental sciences, but it is well known to suffer from an overwhelming diversity of models, particularly to simulate streamflow. We discuss here in detail how such diversity did arise based on the example of Switzerland. The case study's relevance stems from the fact that Switzerland, despite being a small country, shows a variety of hydro‐climatological regimes, of water resources management challenges, and of hydrological research institutes that led to a model diversification that stands exemplary for the diversification that arose also at larger scales. Our analysis, based on literature review, personal inquiry, and an author survey, summarizes the main driving forces behind model diversification. We anticipate that this review not only helps researchers from other fields but in particular also the international hydrology community to understand why we have so many different streamflow models.


| INTRODUCTION
Since the advent of hydrological modeling, the number of models keeps increasing at a fast pace. It has become common to talk about the "plethora of hydrological models" (Clark et al., 2011). Some authors support the idea that there are too many hydrological models, which might lead to a waste of time and effort, and that the hydrological community should gather on a Community Hydrological Model (Weiler & Beven, 2015).
While any newcomer to hydrological modeling will easily find some guidance on navigating the sheer diversity of hydrological models, on understanding their concepts and limitations (Beven & Young, 2013;Kauffeldt et al., 2016;Solomatine & Wagener, 2011), and on their historical emergence (Beven, 2020a(Beven, , 2020b, the question of how this diversity has emerged receives much less attention. Existing historical analyses of model diversity (Peel & McMahon, 2020) generally focus on the technical evolution of model types. However, according to our personal experience, much of the knowledge about why many similar models have emerged is transferred informally.
One of the key drivers for the pronounced model diversity in hydrology is undoubtedly the wide range of model applications (Weiler & Beven, 2015) that all require appropriate modeling; this concept can be defined following Rosbjerg and Madsen (2005) as "the development or selection of a model with a degree of sophistication that reflects the actual needs for modeling results". Indeed, there is not a single valid model fitting every purpose (Hamalainen, 2015). Two well-accepted characteristics that models should exhibit are parsimony and adequacy to the problem at hand, that is, a model should not be more complex than necessary and should be fit-for-purpose (Beven & Young, 2013). In other words, the hydrological model diversification is strongly driven by the modeling context and by what is now often called uniqueness of place (Beven, 2000). However, the hydrologic literature also offers other explanations, ranging from legacy reasons for model selection (Addor & Melsen, 2019) to a lack of agreement on concepts for process representations and to the simple wish to try to do better with yet another model parameterization (Weiler & Beven, 2015).
We attempt here an analysis of what might explain the emergence of multiple hydrological models at the scale of Switzerland, a country that is small enough to do an exhaustive analysis of published work. Despite Switzerland's small area (41,285 km 2 ), we make the assumption that the range of hydro-climatological regimes (Aschwanden & Weingartner, 1985) and of water resources management challenges (FOEN, 2021;Schneider & Homewood, 2013) are diverse enough to shed light on some of the dominant drivers of model diversity that are also at work at larger scales. Most importantly, Switzerland has a federal system, with universities and research institutes depending on different national and cantonal funding sources. Accordingly, the country shows a relatively high density of independent and competing hydrological research groups, which can be assumed to replicate at a small scale the model development dynamics that exist at a larger scale.
Digital hydrological modeling in Switzerland began in the 1970s, shortly after the first hydrological computer models emerged internationally (Peel & McMahon, 2020). During these early times, there was a strong focus on the simulation of snowmelt runoff (Braun & Lang, 1986;Martinec, 1970) and glacier-melt runoff (Braun & Aellen, 1990). Snowmelt modeling probably started with the work of Hoeck (1952), and modeling of snow runoff kept a strong influence on hydrological model development in the following decades (see Supporting Information for additional details), along with a focus on understanding the role of forests in the water cycle (Forster, 1989;Keller & Forster, 1991). In addition to modeling studies in experimental catchments (Iorgulescu & Jordan, 1994), first model-based climate change (Bultot et al., 1992) and land-use change (Jordan et al., 1990) impact studies appeared. Quantitative real-time forecasts for water resources management (Lugiez et al., 1969) and hydropower production (Jensen & Lang, 1973) started being based on hydrologic models rather than statistical approaches. The work of Hottelet et al. (1993) is an early example of model adaptation to a specific catchment; they added three parameters to the Hydrologiska Byråns Vattenbalansavdelning model (HBV)-Eidgenössische Technische Hochschule Zürich (ETH) (Swiss Federal Institute of Technology in Zürich) to account for aspect-dependent snowmelt and karst runoff for the Thur catchment.
Most notably, model diversity already began to puzzle the Swiss research community in the late 1970s; Naef (1977) presented a first model intercomparison study, comparing complex and simple models, and Naef (1981) notably asked: "But, given that the results are good, why do new models continue to be published?" This work aims to disentangle the motivations and reasons behind the choices that led to the current co-existence of a wide range of models. Our analysis relies on a literature review of hydrological modeling of Swiss catchments as well as on a survey conducted with 50 scientific actors (authors of published journal articles) in this domain to assess the more informal and subjective drivers of model selection. We focus here on hydrological models (Box 1) that simulate hydrological processes, including surface and subsurface flow and the resulting streamflow at the catchment scale. We exclude models that simulate the water balance without providing streamflow at the catchment outlet. The hydrological models developed in Switzerland (Table 1 and Appendix) range from rainfall-runoff (Box 1) models (PREVAH, GSM-SOCONT, RS, SEHR-ECHO, WaSiM) to models with emphasis on snow processes (Alpine3D), glacier-hydrology models (GERM), and water temperature models (StreamFlow). Some models have their origin outside Switzerland but are now actively being developed in Switzerland (HBV-light, TOPKAPI-ETH, SUPERFLEX) or were applied to Swiss case studies (CemaNeige-GR6J, LARSIM, VIC, SWAT, mHM).

| DRIVERS FOR MODEL DIVERSITY
The hydrology of Switzerland is shaped by the interplay of its two mountain ranges (Alps and Jura) with the Central Plateau and some of the largest lakes in Europe (Michel et al., 2020). Accordingly, the characteristics of this landscape and its dominant hydrological processes have a prominent place in model development and selection. This modeling context, characteristic of any alpine country, is generally acknowledged in modeling papers through the choice of a model that is "specifically designed to capture hydrological processes that are important for catchments in complex terrain" (Anghileri et al., 2019, referring to the PREVAH model). Multiple similar statements can be found in the literature, for the aforementioned PREVAH model (Brunner, Liechti, & Zappa, 2019;Köplin et al., 2010;Verbunt et al., 2007;Zappa & Kan, 2007) as well as, for example, for WaSiM (Jasper et al., 2002;Jasper & Kaufmann, 2003;Thornton et al., 2021) or for HBV-light .
The high, topography-induced spatial variability in alpine environments is a driver for increased spatial resolution in hydrological models, compared to regions with more gentle topography . Indeed, most models used and developed in Switzerland are semi-distributed or fully distributed. Very few applications rely on lumped models (e.g., Keller et al., 2019;Müller-Thomy & Sikorska-Senoner, 2019).
While natural processes and characteristics of the landscape have an impact on model selection and development, the context of the model application also comes into play, for example, in terms of constraints on the computational efficiency for long-term climate change impact studies or for operational forecasting. We attempt here a clustering of modeling studies according to either the representation of key processes in the model (Section 2.1) or the influence of the application context (Section 2.2) in various Swiss catchments ( Figure 1).
The information sources considered in this analysis are as far as possible peer-reviewed articles with applications to hydrology. The articles were retrieved based on searches by authors (hydrologists in Switzerland) and keywords. While we tried to search all applications as exhaustively as possible, biases in the search and citing network effects are possible. Where necessary, conference proceedings, PhD theses, research, and government reports are also included and complemented with information obtained upon personal inquiry. A few Swiss models are exclusively used or developed in engineering companies, and these are not included here. Furthermore, our analysis focuses on catchment-scale modeling and excludes studies that focus on hydrogeological modeling (Carlier et al., 2019) and those with a focus on urban hydrology (Peleg et al., 2017) or urban hydrogeology (Schirmer et al., 2013). All articles are not directly referenced in this paper, but a complete table is available in the Supporting Information.

| Cryosphere
Contributions from snow and glaciers are key components of the alpine environment as they play a significant role in hydrological regimes through a temporal redistribution of water (Barnett et al., 2005). These processes thus need to be BOX 1 What do we mean by hydrological model?
A hydrological model is an input-output model that simulates the evolution of water storage, water fluxes, and potentially associated chemical and physical properties at the Earth's surface and subsurface, based on the water balance equation. The term "rainfall-runoff model" is often used for hydrologic models that simulate streamflow at a catchment outlet based on input time series of rainfall. The term "rainfall-runoff" stems from the early times when such models simulated how much water of a rainfall event ran off to the stream (rather than being stored in the catchment), that is, "runoff" designated the part of rainfall that appears as streamflow (WMO, 1992b). Nowadays, rainfall-runoff models are continuous simulation tools that simulate all components of streamflow (including baseflow), and the term "runoff" now designates the lateral (as opposed to vertical) movement of water (at the surface or in the subsurface) towards a river (WMO, 2012). Modern rainfall-runoff models further transform simulated hillslope-scale runoff to catchment-scale streamflow; some of them include in-stream routing. Such models can be generalized to precipitation-runoff models in the presence of snowfall. The term "water balance model" is sometimes used as a synonym for rainfall-runoff models (Boughton, 2004). The correcter term "rainfall-streamflow" model appeared rather early (Young & Minchin, 1991) but is to date (November 22, 2021) only used in 17 WebOfScience publications. Streamflow is in many papers called interchangeably "discharge" and sometimes even "runoff," which is a legacy effect.
accurately represented in the models, which can be achieved with different levels of complexity. Most hydrological models used in Switzerland rely on a simple temperature-index-based snow routine (see, e.g., Jenicek et al., 2018). On the other end, Alpine3D, which is built on the snowpack and ground surface model SNOWPACK , has undoubtedly the most complex representation of snow processes among Swiss models. It allows analyses that require a fully distributed simulation of mass and energy balance, such as the analysis of the influence of snow distribution and water transport within the snowpack on the catchment-scale hydrologic response (Brauchli et al., 2017), or the influence of snow processes on soil moisture and streamflow .
Despite the relatively long history of development and application of snow hydrological models, detailed comparisons between simple snow routines and Alpine3D have not been conclusive so far in Switzerland, particularly in the context of climate change applications (Kobierska et al., 2011;Shakoor et al., 2018). It is noteworthy, however, that applications to extended periods remain challenging due to the high computational demand of Alpine3D, with only few attempts to complete long model runs (Michel et al., 2021). The only other hydrological model used in Switzerland that solves the energy balance to simulate snow accumulation and melt is one of the most recent versions of WaSiM; it solves the full energy balance and includes lateral redistribution of snow, but has seen a single application in Switzerland so far (Thornton et al., 2021).
This quasi-dominance of a single snow model that resolves the energy balance at the Swiss level is comparable to the situation in other Alpine countries, for example, in Italy (Geotop; Endrizzi et al., 2014), in France (Crocus; Vionnet  Strasser et al., 2004). This can certainly be explained by the substantial meteorological data requirements and the considerable effort required to develop and implement such models. This is nevertheless surprising given the importance of snow hydrological processes and related water resources for alpine countries (Beniston et al., 2018). One possible explanation is the strong focus on the impact of glacier retreat on alpine hydrology, which is at the heart of many climate impact studies in Switzerland (Addor et al., 2014;Etter et al., 2017;Finger et al., 2015;Horton et al., 2006;Junghans et al., 2011;Schaefli, Hingray, & Musy, 2007), and in alpine regions in general . In fact, most models developed in Switzerland have a glacier module to quantify the contribution of glaciers to the hydrological response of a catchment (see, e.g., Finger et al., 2011;Uhlmann et al., 2013a;Verbunt et al., 2003;Zappa & Kan, 2007).
However, in the past, all hydrological models used in Switzerland relied on simplified glacier retreat routines, even in the context of climate change impact simulations (e.g., Horton et al., 2006). This gave rise to the development of the glacio-hydrological GERM model (Farinotti et al., 2012;Finger et al., 2013;Huss et al., 2008;Junghans et al., 2011); this model combines a simple snow-hydrological routine to a new glacier retreat parameterization scheme, which is now widely applied internationally by the glacio-hydrological modeling community (Huss et al., 2010, called Δh-parametrization). The awareness of the limitations stemming from poor representations of glacier dynamics in hydrological models is currently rising and further drives hydrological model development. An example is the implementation of the aforementioned Δh-parameterization in HBV-light (Seibert, Vis, Kohn, et al., 2018).

| Sediment production and transport
A particular topic that deserves a more detailed discussion is sediment production and transport. Despite the importance of sediments in the alpine water cycle (Hegg et al., 2006), work in this field in Switzerland essentially focused on process observations (Rickenmann et al., 2012). Attempts were made to model sediment transport with the streamflow F I G U R E 1 Map of Switzerland with its major drainage divides (green), the extent of the "hydrological Switzerland" (orange), and some catchments (brown) that are referenced in the text (Federal Office of Topography swisstopo; Bühlmann & Schwanbeck, 2018) model PREVAH (Raymond Pralong et al., 2015), but such studies fall short of accounting for sediment source dynamics and connectivity. The modeling of sediment sources, as well as transport capacity, requires complex models that yield reliable spatial patterns of hydrological processes.
A single model has been developed for this purpose in Switzerland, TOPKAPI-ETH, which has been modified to account for river bed erosion and deposition processes using a sub-grid modeling scheme . This modification allows the simulation of erosion and deposition patterns, sediment transport rates, and evolution of the channel slope. Battista, Schlunegger, et al. (2020) later developed a new soil erosion and sediment transport module for TOPKAPI-ETH to investigate the effects of precipitation and surface erodibility and their spatial variability on sediment fluxes, and to analyze the role of localized sediment sources . Sediment management, for example, in the context of hydropower production (Gabbud & Lane, 2016;Raymond Pralong et al., 2015), is gaining importance in Alpine countries and might drive the development of new modules to simulate the interplay of hydrological and geomorphological processes at the catchment scale.

| Ecohydrology
Ecohydrology studies the feedbacks between ecosystems and the water cycle, at stream reach or catchment scale (Tague et al., 2020). To date, catchment-scale studies with ecohydrological models accounting for feedback with vegetation are scarce in Switzerland. An early example is a work of Zierl and Bugmann (2005), who used the model RHESSys to study climate and land-use change impacts on alpine streamflow in Switzerland. Later, Fatichi et al. (2012aFatichi et al. ( , 2012b developed an ecohydrological model, called Tethys-Chloris (T&C), to simulate vegetation-hydrology interactions at large scales. It has been applied to catchments in Switzerland to study soil moisture spatiotemporal dynamics (Fatichi, Katul, et al., 2015), as well as to assess the vulnerability of Alpine ecosystems to climate change (Mastrotheodoros et al., 2019).
Another approach consists of coupling ecosystem models to distributed hydrological models, such as TOPKAPI-ETH for the analysis of the interactions between rivers, aquifers, and ecosystems (Foglia et al., 2009;Pappas et al., 2015). In the context of simpler streamflow models, model coupling for ecohydrological purposes is limited to one-way coupling, as in the example of PREVAH that was coupled to a model for bedload transport and channel morphodynamics to assess climate change impacts on brown trouts (Junker et al., 2015).
Most hydrological studies that analyze the interplay of vegetation and hydrology focus on the effect of land-use change by analyzing different vegetation scenarios without modeling the actual feedback, in Switzerland, as well as internationally (Dwarakish & Ganasri, 2015). We further discuss such land-use change studies in Section 2.2.3. Overall, catchment-scale ecohydrological studies remain rare in Switzerland, a situation that might change due to increasing awareness of human pressure on natural ecosystems and potential water use conflicts (Milano et al., 2016).

| Anthropogenic streamflow modifications and water quality
Anthropogenic impacts on water flow in the stream network are growing with expanding urbanization and related hydraulic infrastructure, with efforts to produce more hydropower (Schaefli et al., 2019) or with growing irrigation infrastructure. Related streamflow perturbations should not be ignored in hydrological modeling, but can be challenging to incorporate in a model due namely to the absence of detailed data on anthropogenic water use (FOEN, 2021).
In Switzerland, large hydraulic infrastructure is dominantly related to hydropower operations, including large dams and accumulation reservoirs, and water diversions (Schaefli et al., 2019). Thus, specific modules have been implemented in several hydrological models, which in some cases simply led to model extensions, for example, for WaSiM (Verbunt et al., 2005) or TOPKAPI-ETH . Such model extensions were namely essential to analyze anthropogenic versus climate change effects on natural streamflow regimes  or to anticipate hydropower operations in a climate change context (Anghileri et al., 2018;. One model developed in Switzerland has been specifically designed to simulate hydropower operations: RS MINERVE García Hern andez et al., 2020). The development of a specific model was considered necessary to simulate complex hydropower operations in the Valais region (upper Rhone catchment), which shows a particularly high density of hydropower infrastructure.
Compared to hydropower, the main other water user, agriculture, has not received much attention in terms of hydrologic model development and streamflow simulation at catchment scale. A key reason might be that in Switzerland, limited water availability and potential droughts have only recently become an issue (FOEN, 2021). One example that we could find is the study by Fuhrer and Jasper (2012) that has shown that WaSiM is suitable to study the demand and supply of water for agriculture, including irrigation. In this area, models that are widely used internationally have seen little application in Switzerland. One such example is SWAT and SWAT+ (Bieger et al., 2017), whichgiven their user-friendliness (Abbaspour et al., 2007)-might see more applications in the future despite not being specifically designed for alpine environments (Andrianaki et al., 2019;Rahman et al., 2014). The need to model the feedback of agricultural water use and the catchment-scale hydrological response could lead to further hydrological model diversification; corresponding knowledge is currently missing in Switzerland (FOEN, 2021).
Future developments or applications of hydrological models specifically designed to reflect agricultural water use in an alpine context might in particular also be driven by the need to simulate corresponding water quality dynamics. While there are studies related to water quality at very small scales (area of <2 km 2 ; see e.g., Ammann et al., 2020;Frey et al., 2011;Gassmann et al., 2013), where intensive field studies are manageable and local-scale transport can be more easily resolved, only few models exist at a catchment-scale of several square kilometers (Wittmer et al., 2013). Water quality studies at meso-catchment scales are even more rare to date in Switzerland. One example is the work of Abbaspour et al. (2007) who tested SWAT for modeling water quality and nutrient loads in the Thur catchment (1700 km 2 ): they concluded that "in watersheds similar to Thur-with good data quality and availability and relatively small model uncertainty-it is feasible to use SWAT". Even the basic water quality variable, water temperature, has only received little attention in catchment-scale hydrology in Switzerland. It led to the development of an Alpine3D extension, Streamflow Michel et al., 2021).

| Model improvement and uncertainty analysis
A large body of the international hydrologic modeling literature focuses on model performance with respect to reproducing observed streamflow (Beven, 2011;Clark et al., 2011), for example, as a function of model parameterizations (Girons Lopez et al., 2020), of spatiotemporal model resolution (Brunner, Zappa, & Stähli, 2019), of precipitation input data (Müller-Thomy & Sikorska-Senoner, 2019; Sikorska et al., 2017;Sikorska & Seibert, 2016), or parameter estimation techniques Foglia et al., 2009). However, in the context of model diversity, this field of research has overall little impact because most modeling groups who work on such theoretical aspects often use their in-house models for proofs of concepts or to improve them (e.g., Hingray et al., 2010;Schaefli, Talamba, & Musy, 2007).
Overall, there are only a few parameter regionalization studies, that is, studies on the spatial transfer of model parameters, a key topic for hydrologic prediction in catchments without streamflow observations (Guo et al., 2021). On an international level, this question led to the development of specific models in conjunction with spatial parameter transfer approaches (e.g., mHM; Samaniego et al., 2010). In Switzerland, there was rather a focus on understanding the benefit of parameter regionalization combined with few streamflow observations for the calibration of existing models and to counterbalance sparse operational observation networks especially in mountain areas (Viviroli & Seibert, 2015).

| Characterization and quantification of floods and droughts
Infrastructure planning, water resources, and natural risk management heavily rely on probabilistic quantification of extremes, that is, an estimation of what could happen in terms of floods and droughts and their associated probabilities (called return periods in hydrology). Work in this field continues to be based on statistical analyses and extrapolation of observed streamflow time series (Asadi et al., 2018;Brunner, Furrer, et al., 2018), but hydrological models play an everincreasing role to complement missing or insufficient streamflow data.
Any model-based flood estimation method is computationally intensive since long model simulation runs are required at an hourly time step (see Sikorska-Senoner et al., 2020, about reducing computational requirements for extreme flood estimation by hydrological modeling). Accordingly, simple models such as PREVAH (Felder & Weingartner, 2017;Viviroli, Mittelbach, Gurtz, & Weingartner, 2009;Viviroli, Zappa, Schwanbeck, et al., 2009), HBV-light (Brunner & Sikorska-Senoner, 2019;Sikorska et al., 2017;Sikorska-Senoner et al., 2020) and RS MINERVE (Bieri & Schleiss, 2013;Zeimetz et al., 2017Zeimetz et al., , 2018 are often used for flood estimation; these models are all deemed to perform well enough for flood estimation in Swiss catchments by their respective authors and users. Furthermore, at the time of writing, the open-source DECIPHeR model is further developed and implemented for Swiss catchments to bring diversity in the type of models used for flood modeling (Kauzlaric, personal communication).
Other more complex distributed models are also used for flood modeling, but more often on an event-based scale, as in the case of a reconstruction of the 1816 Tambora eruption and its impact on floods in the upper Rhine basin with WaSiM ( Figure 1; Rössler & Brönnimann, 2018). Such a model can be highly relevant to study specific flood types that involve a detailed description of small scale processes, such as to analyze rain-on-snow flood events (Rössler et al., 2014, with WaSiM) or the interplay of rainfall temporal variability and the clustering of saturated areas (Paschalis et al., 2014, with TOPKAPI-ETH).
Work on droughts is much less abundant in Switzerland than work on floods, which is related to the fact that missing water was, in the past, not a hot topic in this country known as the water tower of Europe (Milano, Reynard, Bosshard, & Weingartner, 2015). What can be highlighted here is that the same models are in use to assess droughts and floods, potentially with specific recalibration, but without modification of the model structure. This is motivated by the fact that existing models are deemed to reproduce well all dominant processes in the Swiss environment, as, for example, explicitly stated in the work of Zappa and Kan (2007) on quantifying the hydrological impact of the 2003 heatwave with a distributed version of PREVAH. It was also later on used for additional drought analyses (Brunner, Liechti, & Zappa, 2019;Zappa et al., 2019), where a spatial mismatch between water scarcity and storage availability has been highlighted for Switzerland (Brunner, Liechti, & Zappa, 2019). Similarly, HBV-light served in several drought studies, such as for the definition of a new drought index that accounts for snow , the predictability of low flows , or the sensitivity of catchments to meteorological droughts . It was also used to assess low flow drivers in Alpine catchments (Arnoux et al., 2020). However, significant efforts to improve the representation of groundwater and the corresponding baseflow during droughts remain to be done in Switzerland, which might lead to further model development or diversification.

| Climate change and land-use change impact analysis
Climate change impact studies emerged in Switzerland in the 1990s, including a large national research program on climate change and natural hazards (SNFS, 2021). Since then, all model-based studies are mostly conducted with the models that established themselves in Switzerland, which have, however, not been specifically designed for climate change impact analysis; detailed assessments of how well these models can simulate future conditions are largely missing. Climate change impact analyses require, among others, a good representation of the processes that play a key role in the studies. The representation of these processes in the models is discussed in Section 2.1, both for present and future climate conditions. These studies also require models that are not too time-consuming due to the processing of long transient periods as well as due to the increasing ensemble size of modeling chains.
WaSiM was chosen by Middelkoop et al. (Middelkoop et al., 2001) for a climate change impact study primarily due to its good interpolation of meteorological data for mountain environments, and by Jasper et al. (2004) to assess the effect of different regional climate scenarios in the Thur and the Ticino catchments (Figure 1). WaSiM has also been applied to assess future soil water patterns Rössler et al., 2012) and future summer evapotranspiration regimes , applications where a spatially detailed representation of vertical and horizontal flow processes-involving the Richards equation-is important. It was even applied for the entire Rhine basin at a 1 km 2 resolution down to Rotterdam by Kleinn et al. (2005).
TOPKAPI-ETH, the other frequently used fully distributed model with explicit simulation of horizontal and vertical fluxes, has also been applied to several climate change impact studies, such as in the analysis of internal climate variability , or the assessment of future water resources . It was also selected by Fatichi, Rimkus, et al. (2015), who argues that "it represents a reasonable compromise between the physically meaningful representation of hydrological processes and computational time for large-scale (>1000 km 2 ), long-term (>20 years), high-resolution (<1 km 2 ) distributed simulations".
However, the most widely used models to study climate change impact on streamflow are to date the so-called reservoir-based models PREVAH (Brunner, Farinotti, et al., 2019;Köplin et al., 2012;Milano, Reynard, Köplin, & Weingartner, 2015;Speich et al., 2015, and others; see Supporting Information) and HBV-light Etter et al., 2017;Hakala et al., 2020;Jenicek et al., 2018, and others). In the western part of Switzerland, RS MINERVE and GSM-SOCONT have been used, especially for high elevation sites (Horton et al., 2006;Terrier et al., 2015;Uhlmann et al., 2013aUhlmann et al., , 2013b. The rationale for using these models is well summarized by Köplin et al. (2010) who, for PREVAH, states that the model "has been developed especially to suit conditions in mountainous environments" and that it "has proved to be a reliable and flexible tool for various scopes of application and climate conditions ranging from drought analysis over water balance modeling to flood estimation and forecasting".
One point to note is that the adequacy of the models to a different climate is generally not discussed. References to previous studies are sometimes provided, without the latter having addressed this point explicitly. While it is relatively easy to demonstrate a model's ability to reproduce floods or drought conditions, its transferability to other climate conditions is more difficult to prove directly and generally not tackled (Section 4).
Compared to climate change impact analysis, land-use change studies are rare in Switzerland. Examples are the work on the effect of forest change by Köplin et al. (2013), Schattan et al. (2013), and Speich et al. (2020), all three with PREVAH or by Alaoui et al. (2014) with WaSiM. This scarcity of detailed land-use change studies might be explained by the dominance of climate change impact studies over the last few decades.

| Operational forecasting
The ever-increasing need for reliable real-time streamflow forecasts leads to a continuous evolution of the underlying hydro-meteorological modeling systems. Operational forecasting started with deterministic forecasts from a single meteorological forecast applied to a single hydrological model; today, users expect full probabilistic ensemble forecasts at hourly time scales, updated every few hours and with several meteorological inputs applied to different hydrological models (Jasper & Ebel, 2016). Coupled atmospheric-hydrologic ensemble prediction systems were proven to provide better forecasts than deterministic simulations Liechti et al., 2013;Verbunt et al., 2007;Zappa et al., 2008). These might also include data assimilation schemes (Jörg-Hess, Griessinger, & Zappa, 2015;Jörg-Hess, Kempf, et al., 2015) or the assessment of hydrologic uncertainty related to meteorological forcing, model parameters, and initial conditions Jaun & Ahrens, 2009;Zappa et al., 2011).
Such modern forecasting systems require hydrological models that provide forecasts at many locations in a stream network, and that includes the effect of hydraulic infrastructures (e.g., of hydropower water intakes and accumulation lakes, Section 2.1.4). Since the early times of flood forecasting in Switzerland, HBV and PREVAH were used in governmental offices (Jasper & Ebel, 2016) as well as in research institutes because of their relative simplicity and low computational costs (Addor et al., 2011;Antonetti et al., 2019;Murphy et al., 2019;Verbunt et al., 2006). Along with HBV, PREVAH, and LARSIM, WaSiM is today part of the Swiss operational ensemble forecasting system (Jasper & Ebel, 2016), which uses the FEWS platform (Flood Early Warning System; Werner et al., 2013) to provide forecasts for the cantonal authorities and the public (Swiss Federal Office for the Environment, 2019). A key advantage of the WaSiM model is that it can explicitly account for lake regulations and hydropower operations (J. Schulla, personal communication, October 23, 2020). WaSiM has also been used for research studies on improving operational flood forecasting in mountainous areas (Ahrens, 2003;Ahrens et al., 2003;Jasper et al., 2002;Jasper & Kaufmann, 2003).
In parallel to the models mentioned above, RS MINERVE is being used as a specific flood forecasting tool for the upper Rhone river catchment, a large catchment (5220 km 2 , Figure 1) strongly influenced by glacier melt and hydropower production (García Hern andez, García Hern andez, Horton, et al., 2009;Jordan et al., 2010). The model has been primarily developed for this application. Furthermore, its interpolation of the meteorological forcing and its partitioning between rain and snowfall have been enhanced to improve the forecasts in this catchment (Tobin et al., 2011(Tobin et al., , 2012. Applications or studies of sub-seasonal to seasonal (lead times up to 4-6 weeks) streamflow forecasts are relatively scarce in Switzerland, but few applications using PREVAH exist (Anghileri et al., 2019; Monhart et al., 2019). PREVAH also plays a prominent role in operational drought forecasting (Bogner, Liechti, Bernhard, et al., 2018;Fundel et al., 2013;Jörg-Hess, Kempf, et al., 2015) and within the operational Swiss drought information platform (Stähli et al., 2013).

| Large scale modeling
To complete the picture, we address here the application of some international hydrological models implemented for Europe or large European river basins such as the Rhine and thus covering at least a significant part of the hydrological domain of Switzerland (Figure 1). However, we restrict ourselves to those models whose code is publicly available and/or whose results are published and/or directly available for Swiss basins. Kauffeldt et al. (2016) presented a technical review of large-scale hydrological models with regard to their suitability for the European Flood Awareness System (EFAS). Specific criteria must be met for a model to be suitable for continental-scale forecasting, such as a representation of all major processes in the domain, flexibility in resolution and spatial discretization, a possibility for data assimilation, and so forth . Among the models evaluated in the study, three have been specifically deployed for Europe: LISFLOOD, HYPE, and mHM (Appendix). While LISFLOOD and HYPE (or E-HYPE for the version covering the pan-European continent) are already running operationally (Appendix) at the European scale, mHM has only recently been applied for the development and evaluation of a pan-European multi-model seasonal hydrological forecasting system (Wanders et al., 2019).
Several other models have been applied specifically for the Rhine basin, mainly focusing on forecasting discharge or climate change impact applications. Examples include the so-called wflow_hbv model (van Osnabrugge, 2020;van Osnabrugge et al., 2017;van Osnabrugge et al., 2019) for hourly/daily streamflow forecasting of the Rhine, allowing lake level data assimilation. Another example is the LARSIM model, which was implemented at a 1 km 2 resolution in combination with HBV-light to assess the origin of streamflow components . The major regulated and unregulated lakes were included in LARSIM, as well as four of the most influent "clustered" hydropower reservoirs present on the upper Aare, upper Reuss, upper Rhine, and in the Ill river catchment (Figure 1).
In general, the skill of most large-scale models is found to be inferior near the main Alpine ridge compared to mountainous or lowland areas. The high Alpine catchments have been identified early as posing a major challenge to large-scale hydrological modeling (Kleinn et al., 2005). Besides larger errors in the meteorological variables (precipitation in particular) and the important effect of water management practices, the smaller the catchment area and the greater the elevation ranges, the more detailed the model structure and the spatial resolution need to be to achieve good model performances . While most likely being canceled out downstream (Kleinn et al., 2005), these problems remain yet to be addressed in large-scale modeling and certainly partly explain why specific Swiss-scale models continue to be extremely popular.

| THE MOTIVATIONS BEHIND THE MODEL CHOICE
In total, we reviewed 157 peer-reviewed journal articles on hydrological modeling in Swiss catchments (Table S1). Excluding the large-scale applications (Section 2.2.5), a Swiss hydrological model (category D or E in Table 1) is selected in 93% of cases, leaving little room for international models. PREVAH takes the lion's share with about 30% of the applications, followed by HBV-light (16.5%) and WaSiM (14.6%). The most used international model is SWAT with a small 4% usage (7 cases), mainly related to research led from outside Switzerland.
As depicted in Section 2, some models are specialized for certain processes, such as Alpine3D for snow and GERM for glaciers, and are thus proportionally more used in these contexts (Figure 2). TOPKAPI-ETH tends to be more used for applications that require a spatially distributed output of horizontal and vertical fluxes, for example, in view of coupling to another model. RS specifically targets flood modeling and hydropower operations, as it was designed for operational flood mitigation with hydropower plants. The three most used models, that is, PREVAH, HBV-light, and WaSiM, are general models and are applied to different topics, such as climate change impact studies, floods, droughts, cryosphere-related processes, and operational forecasting (Figure 2).
In the analyzed articles, about 25% of the authors specifically address the model's adequacy with the context or the landscape. However, this does not mean that adequacy has been formally tested or that it actually drove the choice of the model, but rather that the model is argued as suitable to the intended application. About 53% of the articles do not mention the adequacy of the model to the context. The rest provide some description of the model characteristics that might be interpreted as arguments for suitability to the case study. Furthermore, there are only few examples where the authors explicitly discuss their perceptual model (Beven & Chappell, 2021), that is, their perception of how nature works, which is generally left to papers dedicated to model development.
Other factors than adequacy can drive the model choice: some of a practical nature and others of a more subjective nature. Addor and Melsen (2019) argue that the choice of a model is driven by legacy rather than adequacy, where by the legacy they understand: "practicality, convenience, experience, and habit". This can include, on the one hand, the experience available in the modeling group and, on the other hand, code and data availability. Some of these practical aspects are not reflected in the literature: Peer-reviewed articles involve a reshaping of the narration (Babel et al., 2019;Pontille, 2007;Rinck, 2010), that is, the chronology of decisions is often modified to fit a standard paper structure and the drivers behind the decisions are adapted retrospectively. The practical aspects that drove model selection are thus difficult to assess objectively without additional information. For that reason, we conducted a survey inviting all first authors of the analyzed papers to participate, as well as other actors in the research community actively using hydrological modeling approaches in Switzerland. About 100 persons were contacted, and 50 took part in the survey. The webbased survey could be completed anonymously.
The survey started with four general questions on the experience of the researcher in the domain of hydrological modeling and model development (Figure 3). The objective of this first part was to provide a context to the answers given in the rest of the survey. The survey then included multiple-choice questions (Figure 4)  F I G U R E 2 Hydrological models applied to different contexts in Switzerland. The importance of the link is proportional to the number of scientific articles. The importance of some models can be inflated by the fact that an article can address multiple contexts, such as floods and climate change. Models with too few use cases (less than three) are not included for the sake of clarity. A3D stands for Alpine3D and G-SCNT for GSM-SOCONT on the model development experience of the researcher (Figure 3d). The stratification based on experience was most relevant compared to the others.
We first look at the effect of the host institutions of the first author. About half (52.5%) of the researchers stayed at the same institute when they published-as first author-most of their articles on hydrological modeling (Figure 3b). In such a case, the model choice is likely strongly biased in favor of the model developed and used at the corresponding research institute (Addor & Melsen, 2019). Indeed, in 66% of analyzed articles, the first author is affiliated with the F I G U R E 4 Results for the survey questions about model selection. The histograms have been stratified based on the involvement of the researcher in model development ( Figure 3d). As the question on the involvement offered multiple selections, the stratification was performed according to the hierarchy of Figure 3d, and researchers were thus assigned to the highest selected class institute where the chosen model is being developed. When asked if the model developed at the institute was chosen without further comparison, the researcher's opinions differed (Figure 4). While 36.2% agreed with that statement (the percentage of agreement is defined here as the sum of the frequency of the two positive categories, i.e., "relevant" and "highly relevant"), 44.7% did not agree, including both nondevelopers and researchers mostly involved in model development. Researchers with less development experience were also more influenced by their superiors when it came to the model choice than researchers with more experience. 72% of all researchers considered know-how at the institute as a relevant driver for previous work (67.4% for the present situation), significantly more than the personal knowledge of a model (which, however, seems much more important for the present situation). The strong link between the institutes and some models (see also Figure 5) brings the advantage that the model at hand is thus well known, that the code is available, and that specialists for potential modifications are present. All of these factors are a positive aspect, as in-house experience (or habits) can contribute to efficiency and expertise (Babel et al., 2019). Care has to be taken that habits do not come at the expense of adequacy. The risk of habits lies in the automatism of our decisions (Babel et al., 2019), and in the Hammer and Nail syndrome ("if all you have is a hammer, everything looks like a nail"), that is, a researcher will seek to solve every problem with the known model (Hamalainen, 2015). When the chosen model is not developed at the institute where the research is conducted, collaborations are often established with the model developer or lead researcher. Consequently, the model developer (or team leader) is co-authoring the paper in 72% of the published articles. Survey results also show that collaboration was a driving factor for 50% of the researchers (Figure 4). Another reason driving the hydrological model choice is the reuse of a model already set up for a catchment of interest, sometimes without the need to recalibrate it. About 20% of the articles explicitly state that the applied model comes from another study. This number should likely be higher, as some authors publish multiple articles targeting different topics using a single model setup without explicitly mentioning it. The survey revealed that 51% of the researchers agreed that model setups had been reused (Figure 4). Access to the model source code has been qualified as a relevant driving factor by 55.1% of the researchers for previous work, unsurprisingly with a clear contribution from model developers (Figure 4). This number even increases to 81.6% for the current situation, which reveals a strong desire to be able to edit the model code to adapt it when F I G U R E 5 Links between hydrological models and the Swiss institutes that use them. Same conventions as Figure 2. Models or institutes with too few use cases are not included necessary. However, despite access to the source code, researchers do not necessarily modify it. This might also partly be a psychological factor, that is, people feel more secure to invest time and energy in a model that they know they can eventually adapt, rather than being stuck in a situation where they cannot modify the model code. Although access to the model code is important to many authors, it does not mean that everyone sees the necessity of public availability (as open-source) of the model code. Indeed, 61.2% of the researchers considered the public availability of the model code as important to them, which is high; but the number is considerably smaller than the 81.6% of researchers who would just like to have (at least personal) access to the code. Some models have been open-source for a long time, such as VIC or FUSE (Clark et al., 2008), but either were used very sporadically for an individual catchment-generally by foreign institutions-or have not yet been used in Switzerland.
Adequacy to specific processes or landscapes was considered as important in model selection by 57.1% of the researchers for past studies and by 65.3% for the current situation (Figure 4). These numbers seem to nicely fit the comparatively lower numbers regarding technical constraints and data availability (36.7%), or project-related constraints (26.5%). Interestingly, these numbers are of the same order of magnitude than for the ease of use (30.6%), even if ease of use seems to become more important over the course of the career (with 55.1% for today's choices).
The widespread usage of a model (34.7%) is of intermediate importance for today's choices. Similarly, the presence of a strong model developer community is also deemed of intermediate importance. Jointly, these two results might reflect that model diversity (rather than convergence towards a widely used model with a strong community) is not perceived as being problematic in the hydrological model user and developer community.
External factors also influence the model choice. While education seems to play only a marginal role, sociological aspects, such as recommendations from colleagues or peers, have a nonnegligible impact (38.8%), mainly among researchers not involved in model development. Babel et al. (2019) showed, based on interviews with model developers, that the choices made during model conception and development can be indirectly influenced by various actors. Some transferred knowledge becomes self-evident for the developer, who embraces external influences. This process of incorporation of external influences is also likely to impact the model choice.
Finally, emotions are not stated as a key factor, mainly for past work. We would have expected a stronger signal of personal attachment of the developers to their model. The survey terminology was maybe not ideal here, leading to biases in the answers; an emotional relationship to a model might be seen as something negative. Other terms, such as "familiarity" or "confidence" might have led to different results. This dimension is highly subjective, and we dare to suggest that this subconscious driver is more important than stated on the survey answers, mainly for developers invested in a specific model. Indeed, Chamberlin (1890) shows that people tend to develop a parental affection for their theories, which can be transposed to models, and which hinders recognition of the benefits of other approaches. Moreover, emotions were shown to play a significant role in decision making and can be seen as "invisible actors in the system of problem solving" (Hamalainen, 2015).

| ON MODEL STRUCTURAL UNCERTAINTY AND MULTI-MODEL APPROACHES
The World Meteorological Organization (WMO) used to carry out comparisons of hydrological models and publish the results in Operational hydrology reports (WMO, 1986;WMO, 1992a). In this context, different international models were compared over various catchments, including the Dischma basin in Switzerland. The Snowmelt Runoff Model (SRM, Martinec, 1975, see Supporting Information) has also been assessed in these early model comparisons (WMO, 1986, WMO, 1992a. Since then, few model intercomparison studies have been conducted in Switzerland. The existing ones are studies on what we can gain from model complexity in terms of model performance. For example, Gurtz et al. (2003) compared PREVAH and WaSiM for two mountainous catchments, showing that, despite their significant complexity difference, both models show a similar performance in terms of streamflow simulation. This result was also reproduced in the study by Orth et al. (2015) comparing HBV-light, PREVAH, and an even simpler model, but with performance varying depending on the considered hydrological variable or hydrological conditions. They highlighted that that there is no single model that is best suited to all hydrological conditions.
Similarly, few studies are relying on a multi-model approach to predict future hydrologic behavior. The use of multiple hydrological models aims at accounting for the uncertainty associated with the models' structure, their numerical implementation, and their processes representation. The idea is to consider a diversity of models that represent the uncertainty at play (Babel et al., 2019). Such multi-model approaches are in particular extremely relevant for probabilistic flood forecasting  or for long term climate change impact predictions (Andrianaki et al., 2019;Kobierska et al., 2011Kobierska et al., , 2013, where different model formulations can lead to different model sensitivities with respect to the climate forcing and thus to different results. In the context of climate change impact analysis, multi-models are in fact of key importance to understand the dominance of different uncertainty sources, that is, those related to the climate change scenarios and those related to the hydrological modeling itself. Bosshard et al. (2013), for example, showed that none of the uncertainty sources they assessed is negligible and that the respective contributions to the uncertainty vary over the year with a larger contribution from the hydrological model in winter and spring. Similarly, Addor et al. (2014) assessed the impact of climate change on hydrological regimes comparing models of different complexity (HBV-light, PREVAH, and WaSiM) for six catchments in Switzerland. The hydrological models contributed significantly to the overall climate change impact uncertainty in the (partially) glacierized catchments. This finding was reproduced in the recent re-evaluation of climate change impacts on Swiss hydrology (Hydro-CH2018, FOEN, 2021 with a similar hydrologic model ensemble (HBV-light and two different versions of PREVAH). It showed that the largest relative differences between the models occur especially in glaciated or high-elevation catchments in the low flow seasons (winter and spring) or in summer in terms of absolute differences in simulated streamflow.
It is noteworthy that in all these studies, it was assumed that the parameters calibrated for the present conditions also apply in the future. As shown by Melsen and Guse (2021) for 605 catchments in the United States with the models SAC (Sacramento Soil Moisture Accounting model; Burnash et al., 1973), VIC, and HBV, this is an important and specific uncertainty factor in climate change impact modeling chains, but how to address this question remains one of the key challenges for hydrological modeling, and has not seen any specific developments in Switzerland so far.
Overall, multi-model studies are still rare in Switzerland. This is in strong contradiction with our survey, which revealed that 80% of the researchers considered the impact of the model structure as an important source of uncertainty, and 72% supported that multi-models should be considered (Figure 4). However, only about 20% claimed to do it regularly, while 53% admitted to (almost) never do it. The first reason given for not doing multi-model approaches is the lack of resources (time or/and money) for such exercise (supported by 76.6% of the participants). The second reason is that it requires too much effort to set up another model (61.7% of the participants); the third one is that other sources of uncertainty are considered more significant (40.4%), and the fourth is related to the lack of know-how of another hydrological model (31.9%). In addition, other reasons were provided, such as the carbon footprint of simulations and thus the need to do as few simulations as possible with improved models or the difficulty to interpret the results. Other participants suggested working with a single model but testing multiple structures and parameter values or working with model variants (such as in the work of Girons Lopez et al. (2020), who tested numerous snow routines with different degree of complexity).
Researchers are used to work with a single hydrological model (70% of the survey participants; Figure 3) and rarely learn to use another model. As discussed above, the use of another hydrological model requires a substantial effort. An eased assessment of multiple structures is the objective of the so-called "modular modeling frameworks" (used only 4% of the time by the survey participants), strongly encouraged by Clark et al. (2011). Modular modeling frameworks allow hypothesis testing by generating models using different modules and variants of numerical implementations (Follette et al., 2021). Nowadays, several of these frameworks exist, such as SUPERFLEX, FUSE (Clark et al., 2008), PERSiST (Futter et al., 2014), ECHSE (Kneis, 2015), MARRMoT , Raven (Craig et al., 2020), and SUMMA (Clark et al., 2015). Such modular frameworks can favor the generation and testing of new hypotheses about dominant hydrological processes and how to encode them in models. An example in Switzerland is given by Dal Molin, Schirmer, et al. (2020), who discusses, based on the SUPERFLEX framework, how to flexibly adapt the model structure to integrate new hypotheses about dominant hydrological processes.
Modular frameworks have theoretically the potential to counter-act model diversification, since they assemble different model structures in a single framework. The current experience and the survey show, however, those modular frameworks are not easily adopted by the modeling community.

| DISCUSSION AND CONCLUSION
The objective of this work was to disentangle the motivations and reasons behind the choices that led to the current co-existence of a wide range of streamflow models, even at the scale of a small country. A key result of the literature review is the fact that published work crucially lacks adequately addressing model adequacy for the study context or the landscape, which not only impedes objective insights into decision processes involved in model choice, but is also in contradiction to the web-based survey where model adequacy to the analyzed landscape was of foremost importance for model developers and users.
This model adequacy to the landscape might at least partly explain the lack of adoption of international or largescale models. Researchers active in Switzerland are, in fact, very keen on using local models, either developed in Switzerland (93% of the case studies) or even at their own research institute (66% of the articles analyzed). This strong link between models and institutes provides the benefit of expertise and efficiency, but at the risk of context inadequacy for the selected models and of automatism in decisions (Babel et al., 2019). This model-institute link is likely to be one of the main causes of the existence of so many hydrological models, since each research group develops its own tools.
It is important to underline here that hydrological model diversity is interesting to represent the uncertainty related to the hydrological model structure or implementation (Babel et al., 2019) and that there will never be a single model valid for all applications (Hamalainen, 2015). Thus, model diversity is indeed desirable, as long as it is used to assess model variability. However, in Switzerland, multi-model applications to harness this diversity are largely missing. A first barrier to this approach is the challenge to master multiple models. In this regard, modular modeling frameworks could help to move in this direction, but our review shows that they are not easily adopted.
This lack of multi-model studies is in contradiction with the importance that is attributed to such studies in our author survey. This finding is certainly transferable to the larger international modeling community. Many of us advocate the development of methods that facilitate robust model development (Zheng et al., 2018), but most of us do not apply these principles ourselves. This contradiction between what we think that we should do (multi-model analysis) and what we actually do (use our local model) might be partially related to research funding, in Switzerland and elsewhere: Hydrological studies tend to receive small-scale funding (limited in time and resources) from local or regional authorities or private actors (such as hydropower companies), which does not push towards the use of widely accepted international models or towards the development of model intercomparison studies, but which can favor the development of local models. Short projects and associated efficiency constraints also push towards the reuse of a certain model setup rather than towards the optimal or most robust modeling solution for the research questions at hand.
While we see an impressive model diversity at the single or few catchments scale, it is noteworthy that our review points towards an important national-scale shortcoming: some catchments have been "over-modeled" (e.g., Thur, Dischma)-albeit with little model intercomparison (Section 4)-while there is a clear lack of large-scale or national studies. In particular, the few available studies often do not cross the Swiss national border, even though the "hydrological Switzerland" extends to its neighboring countries (Figure 1). The absence of such larger-scale studies might be explained not only by the lack of resources, but also by shortcomings and challenges more widely encountered in hydrological modeling over larger domains. These include differing quality and scales of input data and streamflow observations and large heterogeneity in hydrological behavior (possibly requiring more than one specialized model). Yet, this heterogeneity may in fact provide us with the opportunity to improve our understanding of differences in model adequacy and "benchmarking" model performance Newman et al., 2017;Seibert, Vis, Lewis, & van Meerveld, 2018), and to draw most needed conclusions on the robustness of generalizations and on estimation uncertainty (Gupta et al., 2014;McMillan et al., 2016).
Two reasons might, in our view, favor a growing adoption of international models in Switzerland or at similar regional scales and thereby slow down the rhythm of local model diversification: The first is to account for processes that most local models have not specialized in, such as, in the Swiss case, agricultural water use or ecosystem interactions. The second reason is the growing adoption of version control systems that allow collaboration on open-source code with unprecedented ease. Indeed, as the survey revealed, most researchers desire access to the source code, for potential model adaptation to specific needs. Open-source models thus offer an interesting alternative for research projects that go beyond the capabilities of in-house models. This might ultimately contribute to model improvements and lead to an international community-driven dynamic that is beneficial for all, model developers as well as model users. The future will reveal how this might affect model diversity at international scale.
To conclude, we would like to emphasize that despite some local specificities related to the alpine context, the key findings of our study are readily transferable to other contexts or scales: a foremost driver of hydrological model diversity is the use of local models, to benefit from local know-how as well as to maximize model adequacy to the analyzed landscape. Modeling contexts and landscapes might change, but our scientific, practical, and financial incentives or constraints are pulling us in the same direction. The following take-home messages, derived from this review, might improve the navigation of the hydrological modeling landscape in the future and be beneficial for future model development: • Modeling studies should come with a succinct statement about the model choice even if the choice was purely heuristic and, for example, linked to the research institute. This might in particular also contribute to the reproducibility of all modeling decisions and thus of the work. • Multi-model studies rely on model diversity; they are at the heart of hydrological model development and uncertainty assessment but are particularly challenging to implement. Additional efforts are needed, both from the developer community as well as from the funding side to ease future implementations. • High hydrological model diversity might come in pairs with a lack of diversity of studied catchments; model development might benefit from an expansion and coordination of test cases in the future. • Current catchment-scale hydrological model diversity arose with a strong focus on water quantity and climate change impact studies. Other areas will certainly catch up, which will most probably further increase model diversity. Modular approaches and collaboration on open-source codes could become key tools to avoid an explosion of the number of models in use.
Finally, we would like to emphasize that the evolution of hydrological models and associated modeling practices is strongly linked to current developments in the field of digital scholarship and its evaluation by stakeholders such as funding agencies and career committees (David et al., 2016). How these developments are impacting model diversity remains to be assessed in future work.

ACKNOWLEDGMENTS
We would like to thank various Swiss and international colleagues for their help on the history of hydrological modeling in Switzerland (see Supporting Information) and Karsten Jasper for the insight on hydrological modeling at the Swiss Federal Office for the Environment (FOEN). We are also grateful to the two reviewers who helped improve the paper, as well as to the 50 participants in the survey who provided valuable information for this analysis.

CONFLICT OF INTEREST
The authors have declared no conflicts of interest for this article.

DATA AVAILABILITY STATEMENT
The details of the analysis of the application studies have been published as: Horton, Pascal;Schaefli, Bettina;Kauzlaric, Martina (2021), "