A systematic review of natural flood management modelling: Approaches, limitations, and potential solutions

The Pitt Review of the 2007 summer floods in the UK, published in 2008, commended the potential of natural flood management (NFM) for reducing flood risk. NFM is a nature‐based approach that has since gained substantial interest from both practitioners and academics. The review further highlighted the need for catchment‐based flood management (CBFM) to enhance resilience to flooding and climate change by incorporating NFM and wider nature‐based solutions into hard flood protection systems. Such integrated approaches are considered to be more sustainable and adaptable than the traditional hard‐engineered measures. More recently, the European Commission's European Green Deal also highlighted the need for greater use of nature‐based solutions including NFM for managing flood risk. Whilst there have been many attempts to quantify the effects of NFM through hydraulic and hydrological modelling, there is still no systematic review conducted for these modelling works. This review aims to summarise the current NFM modelling approaches, as well as discussing their key limitations related to data, model methods, and real‐world applications. This paper then goes further to highlight potential solutions to some of these challenges and provides guidance to assist modellers to improve future modelling and data collection process.


| INTRODUCTION
Globally, river flooding is the most wide-spread natural hazard that has caused the greatest loss of life and damages to economic systems to date (CRED, 2020;Jongman et al., 2015). From 2020 to 2030, the total number of people impacted by riverine flooding is expected to double worldwide from 65 million to 132 million due to climate change and socio-economic development in floodplains (Ward et al., 2020). Conventional hard engineered solutions are widely implemented as the main approach to manage flood risk in cities and catchments. However, these flood protection infrastructure systems require large capital investment and maintenance costs, and may potentially cause ecological and water quality degradation (Bunn & Arthington, 2002;Mueller et al., 2011;Ward et al., 2017). In recent years, a paradigm shift in flood risk management (FRM) has emerged to integrate hard engineering methods with 'soft' approaches including nature-based solutions. This approach to flood risk reduction has gained wide-spread interest in both the academic community (Dadson et al., 2017;Iacob et al., 2017), governmental institutions, and industry (Burgess-Gamble et al., 2017;SEPA, 2016). In particular, since the Pitt Review addressing the 2007 summer floods (Pitt, 2008), the UK has taken major steps towards natural flood management (NFM), referring to measures that reduce flood risk through working with natural processes (WWNP) (Connelly et al., 2020). NFM employs techniques ranging from leaky barriers and a range of runoff attenuation features (RAF's) to coastal and soil management to protect, restore and emulate natural regulating functions of rivers (Burgess-Gamble et al., 2017). These NFM techniques impact roughness and infiltration, as well as altering flow rates to reduce flood risk. In addition to reducing flood risk, NFM has been linked to a range of other benefits, such as improving biodiversity, habitat creation (Cook et al., 2016), enhancing water quality (Barber & Quinn, 2012;Howe & White, 2003), public health and well-being (de Bell et al., 2017;Maas et al., 2006;Postnote, 2016), and more widely contributing to climate change mitigation, such as carbon sequestration (Wingfield et al., 2019). The Pitt Review highlighted the UK's continued reliance on building traditional large flood defences for future flood protection, despite trends in contemporary climate change mitigation approaches which prefer more 'soft' nature-based solutions. Soft approaches are considered to be more sustainable and should complement traditional hard flood defences (Pitt, 2008).
Despite a growing evidence base around NFM, there are still major concerns around the effectiveness of the approach at large-scale, the longevity of NFM sites, and the relation of geographical setting and geology for NFM interventions (Dadson et al., 2017;Iacob et al., 2014;Murgatroyd & Dadson, 2019;Short et al., 2019;Wilkinson et al., 2019). Stakeholders are still often reluctant to use NFM for large-scale FRM strategies, and as a result, an ad-hoc approach towards NFM siting has been taken without considerations for the wider implications at the catchment scale. Modelling, varying from 1D hydraulic modelling to 2D fully distributed modelling, has become an indispensable tool for quantifying the impacts of NFM interventions. But most of existing modelling studies have been focused on a sub-catchment scale. Whilst some researchers have attempted to highlight the potential of various NFM modelling frameworks (Hankin et al., 2019), as of yet, there has been no extensive systematic review of the existing modelling studies to identify the potential pitfalls of these approaches and lay out the potential directions for future development.
Most NFM research involves 1D modelling, which simulates linear domains (e.g., river channels and pipes) decomposed into multiple cross-sections or nodes (Betsholtz & Nordlöf, 2017), to solving the 1D St Venant equations or one of its simplified forms for water level and velocity or discharge in the flow system under consideration. These 1D models can only represent flow in a single direction and are typically used for modelling large-scale channel or pipe networks or flow systems in small-scale sites.
In contrast, 2D models typically simulate the spread of water flow over a 2D horizontal domain represented using a DEM or raster grid (Costabile et al., 2012). Compared to 1D modelling, 2D modelling has largely been conducted for both site and catchment scales, with an aim to replicate complex overland flow and flooding processes interacting with irregular terrain features. This approach allows for better representation of inundation process over the floodplain. However, 2D models require a large amount of data to accurately represent domain topography and hydrological conditions, and are also computationally much more expensive, hindering their application in large-scale simulations (Simons et al., 2014). Thus, 1D models are particularly useful where 2D spatial data are not at a sufficient resolution to capture accurately the channel geometries and can be more easily implemented to represent in-channel flow dynamics and hydraulic structures such as weirs. While 1D and 2D models are most often applied independently, modellers have taken to integrating these types of models through a coupling process to gain the advantages of both approaches.
1D-2D coupled modelling is typically used to capture overland flows, flooding process, and in-channel flow dynamics and their interactions. There are two main types of 1D-2D coupling approaches: loose and tight. Loose coupling involves running the two models separately, with the 1D model used to generate boundary conditions for the 2D model (McMillan & Brasington, 2007;Yu & Lane, 2011). This approach has been known to encounter issues in maintaining the continuity of flow. Tight coupling is implemented to enforce mass or momentum conservation, or both, by estimating the flow between the 1D and 2D domains using lateral weir equations based on the water level differences predicted by the two models (Morales-Hern andez et al., 2016). 1D-2D coupled modelling has been effectively used to reduce the long run time and the large amount of data that a pure 2D model requires. Further model integration has also been achieved by cascading different modelling approaches, which involves coupling semi-distributed or distributed hydrological models for catchment-wide hydrology with 1D or 2D hydraulic models for channel flow dynamics (Hankin et al., 2019;Rodríguez-Rinc on et al., 2015). This allows for a more realistic representation of water flows around in-channel features, such as leaky dams or barriers (Ferguson & Fenner, 2020c).
This review is focused on these current approaches and challenges in modelling NFM features at multiple scales, aiming to (1) assess their effectiveness in modelling NFM features; (2) identify their major limitations; and (3) propose a potential approach to modelling NFM features at a larger scale.

| METHODOLOGY
This section introduces the methodological approach taken for this systematic literature review, including: the search terms, selection, outline of search strategy (including search engines used), the screening and eligibility process (which led to the final number of research items included in the review), and how search outcomes were coded.

| Search terms for NFM modelling
A range of nature-based solutions are used to address flood risk and many of them are highlighted within the WWNP's directory (Burgess-Gamble et al., 2017), as well as the EU Natural Water Retention measures platform (NWRM, 2021). The site features involved in NFM tend to overlap with projects conducted for other purposes, such as ecological restoration and conservation. Furthermore, many countries have different synonyms for the same or similar features or sites. To capture as many studies as possible, a range of search terms for referring to sites and features were considered. For individual sites, terms such as 'Nature Based Solutions', 'Catchment Based Flood Management', 'Natural Flood Management', 'Low Impact Development', 'Soft Engineering', 'Green Infrastructure', 'Sustainable Drainage', 'Working with Natural Processes' and 'Runoff Attenuation' (Benedict & McMahon, 2002;Burgess-Gamble et al., 2017;Dadson et al., 2017;Dietz, 2007;European Commission, 2015) were considered. Additionally, some models may only consider one or two NFM features and thus may not use terms for sites. To ensure that studies focusing on these various NFM related features are included in the review, specific search terms were also included such as 'Leaky barriers', 'Offline storage areas', and 'Rain gardens'.
Search terms also included model type because a focus of the review was to discuss how different modelling approaches simulate NFM features, and compare the advantages and disadvantages of hydrological, hydraulic, and hydrodynamic models. The model-related search terms included 1D, 2D and 1D-2D coupled models, as well as more general terms like 'Flood*', 'Simul*' and 'Model*'. The use of the star function (*) allows for any documents that contain any part of the starred word to be picked up in the search, for example, 'Simul*' will detect both 'Simulated' and 'Simulation'. This was done in an attempt to also include those papers that may not denote the model as 1D or 2D and capture the international papers that may use slight variations on wording. All the search terms used for different search engines can be found in Table 1.

| Search strategy
The search strategy used here is similar to the approach outlined by the systematic map to evaluate the effectiveness of UK-relevant NFM measures (Connelly et al., 2020). A systematic search of the databases including Web of Science, Scopus, and Proquest Dissertations was completed between March 2021 and June 2021 for academic literature. Google Scholar's first 300 search results were also included to cover both academic and grey literature, until the results were no longer appropriate (Haddaway et al., 2015). Academic literature here refers to papers or reports published to academic journals or proceedings, while grey literature refers to literature published by government, public and private agencies in a non-commercial form such as government reports and statements.
In addition, the publications of following organisations were specifically searched for further grey literature: Jeremy Benn Associates (JBA) consulting, Arup Group Limited (ARUP), Environment Agency (EA), Scottish Environment Protection Agency (SEPA), Catchment Based Approach (CBA), River Restoration Centre (RRC), Natural Resources Wales (NRW), The Rivers Trust (RT), Natural Environment Research Council (NERC), U.S. Environmental Protection Agency (U.S. EPA) and the European Commission (EC).
The search terms used varied for each search engine or database due to the different common search terms identified for individual search engines. For all grey literature, the search was completed manually due to some organisation sites having a large volume of irrelevant results. However, all search results up until the cut-off were included for later screening. The search string used for each search engine or site is provided in Table 1 and  the table shown in the bottom right of Figure 1 shows a breakdown of how many items were returned from each search.
During all the searches, no restrictions on language, country, or document type were imposed but only literature from 2007 onwards were considered. The year 2007 was selected to ensure that only the most up to date NBS or NFM works following the publication of the Pitt Review are considered.
T A B L E 1 The full list of search terms used for each search engine or site, as well as any specific sections of the site that were searched.

| Screening process
The screening process, as illustrated in Figure 1, began by removing duplicates found between the grey literature and academic literature. The titles and abstracts were then screened manually to find modelling relevant papers that focused on NFM. Initially, ecological models were retained as some papers referred to their runoff attenuation capabilities. Grey and academic literature were then screened for the following criteria. Blogs, news, magazines, publicity models and material, conference abstracts, and other non-modelling relevant information were excluded and only papers pertaining to modelling were considered (Higgins & Green, 2011). Articles returned from any search that could not be accessed were excluded. Any study that pertained to NBS or NFM modelling that was non-fluvial or pluvial (e.g., coastal) was removed, unless the model was hydroecologically related and contained either hydrological or hydraulic modelling components that were calibrated. Studies that did not adhere to best practice were excluded, such as those that did not involve appropriate model calibration or validation, using either primary data as a baseline or alternative statistical methods. Any papers that were identified by manual screening as not relevant to NFM modelling, including those that focused primarily on green infrastructural features (e.g., green roofs or rain barrels), were removed as these are more restricted to urban pluvial flood risk management rather than wider catchments. Finally, any grey literature published before 2007 was manually removed as the academic search could limit the papers to 2007 or newer while some grey searches could not.
Finally, all papers were checked for any final duplicates in either the grey or academic literature. Duplicates were removed and any papers removed manually were included in this count. The full screening process is summarised in Figure 1, including the number of papers removed at each stage.

| Search outcome coding
Outcomes from the searches were then analysed and coded by: source (academic, grey, thesis), year (2007, 2008 etc.), model name (SWMM, HEC-HMC etc.), model type (1D, 2D, 1D-2D), setting (urban, rural, catchment etc.), location (North America, Europe etc.), scale (<1, F I G U R E 1 Diagram illustrating the screening process, based on the PRISMA diagram for systematic reviews (Page et al., 2021), undertaken with a count of the number papers retained or removed at each stage. The table in the bottom right shows the number of items returned following each site searched. *The full exclusion list of types of documents is noted within Section 2 of this paper. <10, <100, and <500 km 2 ), NFM features modelled (RAF's, leaky barriers etc.) and limitations (lack of data, coarse resolution etc.). This allowed for a detailed breakdown and assessment of each paper highlighting their main uses and limitations as well as what sort of features are modelled by each model and the setting in which each model has been adopted. A list of all papers and their coding can be found in Appendix 1.

| RESULTS
A total of 4143 documents were identified from the initial literature search. Following an initial title and abstract screening, 1082 grey and academic items of literature were manually screened and checked between April and July 2021. From this, a total of 115 pieces of literature were identified and any duplications were removed. Following this process, a final manual screening left a total of 95 papers to be included in the review. These papers were then coded as shown in Appendix 1, and percentage breakdowns of each area of coding are identified below.
The results are visualised by a range of graphs in Figure 2a-e. Of the 95 papers identified, 86% of them were from academic sources and 10% were from grey sources. The remaining 4% were theses that had been published and gave detailed information on the modelling process. In total, 97% of the papers identified were from 2013 onwards and 25% of all papers were published in the 2020-2021 period. Very few NFM modelling papers F I G U R E 2 The breakdown of items identified from the search, showing the number of papers attributed to (a) model type (1D, 2D, 1D-2D or other); (b) sources (academic, grey or thesis); (c) setting of study (urban, catchment or subcatchment, rural or physical); and (d) scale of study (<1 km, <10 km, <100 km, <500 km or other). The final graph (e) shows cumulative lines indicating the number of papers, broken down by model type, which have been produced since 2007. published before 2013 involved proper model calibration, which is one reason for the limited number of papers identified prior to this date. Furthermore, the term 'NFM' was formally defined in 2011 (Pescott & Wentworth, 2011), although NFM techniques were modelled prior to its definition, and mentioned in policy documents such as the Flood Risk Management (Scotland) Act 2009 (Spray et al., 2009). The types and versions of models used have changed over time, meaning that more recently developed models or newer versions of existing models may have more advanced capabilities when compared to the older models.
Of all the studies identified, 60% involved 1D models, 20% involved 1D-2D coupled models, 18% involved 2D models, and 2% compared the performance of 2D and 1D models. Herein, we widely classify the distributed models performing simulations over a 2D horizontal domain as 2D models, which includes hydrological, hydraulic, and hydrodynamic models. In particular, we see that 1D models have been used far more than 2D models although the application of 2D, 1D-2D models has been gradually increasing as we get closer to 2021.
Overall, the three most prolific models identified from the search were: the U.S. EPA's Storm Water Management Model (SWMM) n = 47 (1D), HEC-RAS (variants) n = 11 (1D or 2D) and TOPMODEL (variants) n = 10 (2D). SWMM was referenced throughout most of the literature and featured mainly in papers relating to low impact developments (LIDs) or sponge cities, while HEC-RAS and TOPMODEL (variants) were featured in a range of papers and settings with the main focus being small-scale catchment modelling or cascade modelling. Figure 3 presents the paper count per continent and clearly shows the global North and South divide in where modelling has been conducted (and subsequently published), with no papers coming from the whole continent of Africa. Majority of the papers included in this study are from Europe (n = 35), Asia (n = 28) or North F I G U R E 3 The count of papers on modelling NFM features from each continent with a further breakdown of where papers within each continent are from for Asia, South and Central America, and Europe. For continental breakdowns, the largest segment comes first in the key list by country and continues in that order clockwise. America (n = 22) with the UK (n = 27), China (n = 21) and the United States (n = 22) being the main countries of origin for these studies. There is no breakdown for North America, as all the studies were conducted in the United States.
All models reported in this review provided information on the site and scale of the area modelled. A majority of studies were conducted in discrete urban areas (58%), while the second largest group of studies were conducted across wider catchments or subcatchments (34%). The rest were conducted in entirely rural areas (7%), apart from one study which focused on modelling a laboratory test case using a 1D model solving the St Venant equations. Finally, 88% of the models covered a catchment area of 100 km 2 or less, and 99% of them covered a catchment area of 500 km 2 or less. Only one model considered a domain area larger than 500 km 2 , and all models that covered a catchment area of over 100 km 2 were 1D.
A wide range of NFM features have been modelled using different modelling approaches. For example, SWMM has been extensively used to model LID features while TOPMODEL has been applied mainly to model NFM, RAF and land use features. The names of different types of features may vary, depending on where they are implemented in the world. For example, what might be referred to as 'Best management practices' in some countries may be called 'Low impact developments' or 'Working with natural processes' in others. To ensure that different nature-based solutions features are captured in the search, a wide range of features of varied names are included and coded in the results (see key at bottom of Appendix 1).
In regard to limitations identified from the search, some themes clearly emerged relating to data, computation, and real-world or stakeholder implementation, which could not be coded as simply as features. Rather, we consider and discuss the limitations reported by the authors of each paper in the discussion.

| DISCUSSION
A full list of papers included in this review can be found in Appendix 1. This section explores three characteristics of the studies reviewed in detail. The first section considers model type, scale of use, and features included in modelling internationally. The second section identifies the key limitations of modelling approaches. The final section of the discussion will conclude by presenting some potential solutions to the problems highlighted by the systematic review, as well as some of the future approaches that the field would benefit from adopting.

| Current approaches to modelling NFM
This review revealed that NFM modelling approaches vary in relation to type of model employed, the scale of modelling, and the features included in modelling. Table 2 breaks each of the following modelling scenarios down into brief segments with a summary of the major conclusions. The characteristics of models by type are expounded below in detail.

| Channel and piped network modelling: 1D modelling
The 1D modelling of NFM features has mainly been conducted for two separate scenarios. The first scenario is for modelling small-scale urban piped networks, which is typically performed at a scale of <10 km 2 , in mainly urban settings or sponge city scenarios using software such as SWMM (Platz et al., 2020;Steis Thorsby et al., 2020), or L-THIA-LID 2.1 Liu et al., 2015). Features that are modelled in these scenarios typically pertain to LID features or green infrastructure, such as rain barrels, rain gardens, pervious pavements, and green roofs, as these are common ways of integrating NFM into urban settings. 1D models used here typically consist of a nodeto-node network, or a sewer network, and use predesigned LID features implemented at points in the network to investigate the effect on water levels, such as determining whether manholes are likely to overflow.
The second scenario is for modelling large-scale channel networks that are greater than 100 km 2 in size. Modelling of these scenarios should focus on a mix of inline channel modifications and features, as well as differing levels of LID implementation (Li et al., 2015). The 1D modelling approach is normally used to represent these systems due to a wide range of restrictions including computational, and data limits for 2D models. Largescale channel network models are normally implemented at a coarse resolution and incapable of representing complex flow dynamics. As a result, the complex small-scale interactions of NFM features can be left out of the modelling process. These large-scale models are useful for indicating the potential effect of NFM solutions at larger catchment scales, as well as the response of the catchment at defined points in relation to water depth and stream power (Singh et al., 2018), rather than for overland flows. Randall et al. (2019) stated that including high-resolution land use data as a component of largescale modelling is important for reducing uncertainties, consistent with the observations of others (Krebs et al., 2013). Further, studies suggest that additional T A B L E 2 A table showing the breakdown of different modelling scenarios highlighting; each of the models used in relation to this type of scenario, the type of model (2D or 1D), and references for these approaches. Note: A short summary of the major conclusions from each scenario has been added to highlight key take-home points. calibration is needed due to the uncertainties of modelling results at such a large scale (Gülbaz et al., 2017). A number of studies (e.g., Singh et al., 2018;Sisson, 2017) have highlighted a need for 2D modelling at the large scale to provide further insights into inundation process intervened by different types of NFM measures, although this would require high spatial resolution datasets and be restricted by computational power.

| Site scale modelling: 1D and 2D modelling
Anything modelling conducted at under 1 km 2 was considered site-scale. Within site-scale modelling there has been a mix of 1D and 2D modelling studies reported in the literature. In most cases, models used at this scale could not simulate extreme rainfall events. This is likely due to the adoption of the diffusive wave approximation of the shallow water equations rather than the full hydrodynamic solutions (Krivtsov et al., 2020), or a lack of data on extreme rainfall events at this scale (Lin et al., 2021). Another observation is that some models do not consider certain processes such as evapotranspiration. Evapotranspiration may be particularly important for these smallscale models (Avellaneda et al., 2017;Xu et al., 2019), yet often these data are not available at a fine enough scale.
In most instances where 1D modelling has been used for small-scale sites, the model has either used pre-designed parameters for LID modules, or assumed the feature size. This approach is often used when sites were not validated with field measurements. Lin et al. (2021) suggest that a potential reason for the lack of onsite validation is that monitoring plans tend to not be initially designed into sites, and tend to be implemented post-development, which makes gathering site observation data challenging and time consuming. The SWMM model, which includes pre-designed modules, was used extensively for these 1D modelling scenarios at the small scale. Where 2D modelling of small-scale sites has been conducted, a range of different models have been used including multiple studies using 1D-2D coupled models (Burns et al., 2015;Yin et al., 2020), and some purely 2D models designed for urban environments (Krivtsov et al., 2020;Tan et al., 2019). These models mainly focused on LID features with the exception of Fry and Maxwell (2018) who examined land use change, highlighting why the use of 2D models is particularly important for representing complex systems compared to lumped or 1D models. In all studies of this nature, the researchers identified difficulties around either limited data at very small scales, or a lack of data on the routing of complex overland flows in urban environments.
Further, the importance of high-resolution elevation datasets is recognised for capturing site-scale processes (Yin et al., 2020), and improving flow routing in urban environments. 4.1.3 | Catchment scale modelling: 1D-2D coupled, 2D and cascade models At the small catchment or subcatchment scale (10 km 2 < x < 100 km 2 ) a large amount of 2D or 1D-2D modelling has been performed, particularly within rural UK landscapes. Works of modelling catchments in North America (Doubleday et al., 2013;Juan et al., 2017) and China (Li et al., 2018) have also been conducted, but it is not as abundant as in the UK. This is likely due to the relatively good availability of high-resolution elevation data in the UK, as well as the fact that NFM sites tend to be in rural landscapes rather than urban settings.
In the UK, models such as TOPMODEL (variants), ISIS, TUFLOW, HEC-RAS and Infoworks ICM have been used extensively to simulate catchment scale flooding through 1D-2D coupled, 2D, and cascade modelling. Modelling work in 2017 and prior was largely achieved through1D-2D coupled models with a 2D hydrological or hydraulic model capturing the hillslope run-off to provide input as a lumped channel discharge unit into a 1D channel network model . This approach was similarly adopted by practitioners using an ISIS-TUFLOW coupled model (JBA, 2015; Wilkinson & Jackson-Blake, 2017) for modelling small catchments. However, the limitations of using coupled models are recognised when modelling more complex flows and interactions in 1D models.
Following the development of these 1D-2D coupled models, improvements to distributed 2D hydrological models have allowed coupled models to not only capture catchment scale interactions, but also more accurately simulate the dynamics of drainage or river networks (Hankin et al., 2019). This development, as well as the wide availability of higher resolution datasets, in the UK, have allowed for a fully distributed modelling approach capturing catchment scale interactions to become possible. However, there is still a major problem around the coarseness of DEMs at larger scales, as well as the lack of data for calibration. Calibration is required to ensure 2D models sufficiently represent the complex flows around RAFs and leaky barriers. Very high-resolution models of these features are required to ensure accuracy.
Finally, a 2D cascade model integrated into a 1D urban sewer network model has recently been used, in a UK context, to demonstrate how changes to land use or instream features at the catchment scale can impact urban drainage systems and sewer overflows. Ferguson andFenner (2020a, 2020b) developed a model that routes the outputs from TOPMODEL through channel flows in 2D HEC-RAS to improve the numerical representation of in-channel features when compared to purely 1D models. Such cascade models are particularly useful for investigating the integration of catchment interventions into urban systems, and capturing complex overland flows and their interactions with instream features. There are still limitations with this approach as it is challenging to sufficiently represent these NFM features at coarse resolutions, and the approach relies on simple assumptions about system characteristics, such as roughness values and the way NFM features perform during a rainfall or flood event (Ferguson, 2020). Table 3 summarises the characteristics of studies involved in the research including, type of model employed, and the NFM features being modelled. In addition, the interventions involved in each study are listed and classified, following the classification framework proposed by Moore and Rutherfurd (2017). However, this analysis does not offer insight into how modellers can parameterise or conceptualise NFM features. One challenge is that there are many methods for parameterising and conceptualising individual NFM features, using many different models. For example, leaky barriers have been represented simply using the floodexcess volume method in 1D models (Bokhove et al., 2018). However, this ignores losses through the barriers, and assumes the features are empty prior to flood events. Another method is through changing the Manning coefficient (n) of the channel, but again this is inappropriate due to its impact on both high and low flows and little effect on steep slopes (Leakey et al., 2020). A more physically based approach would be to change channel geometries (Thomas & Nisbet, 2012), or to use weir equations to estimate the flow rate passing through . Both approaches can be validated using field data, and are likely to provide a better representation of how barriers function in comparison to other parameterisation and conceptualisation methods. Following these more specific morphological approaches, each NFM feature may be modelled more specifically depending on location, season, and the detailed design of interventions. Figure 4 synthesises which type of models could be used by practitioners for specific NFM interventions throughout a hypothetical catchment.

| Limitations of current models
Three major themes emerged from reviewing the literature for limitations to modelling NFM: data-related challenges, computational challenges, and the real-world applicability aspects of the models.

| Data limitations: Resolutions, lack of data for model setup and calibration
A majority of papers identified 'lack of data' as a serious limitation to the modelling process. Specifically, three issues were identified: the lack of model set-up data (rainfall and land use), the lack of calibration data (water level or flow data), and the lack of high-resolution DTM or DEM data. Some papers (n = 16) did not necessarily state limitations with data used in the model, but instead identified a need for improved data quality, which is still a potential limitation to modelling.
One of the largest limitations noted in the literature was the difficulty in finding data to enable the set-up of all the features of the model. Set-up data (n = 35) limitations varied between studies and included limited rainfall gauging data (Ferguson & Fenner, 2020a), land use data (Miguez et al., 2009), and soil moisture data (Korgaonkar et al., 2018). In regard to land use data, while there is a growing availability of open-source spatial data since Miguez et al. (2009), some of these data still have quite coarse resolutions, although this will likely become finer with technological and remote sensing advancements.
Furthermore, a potential reason for the lack of data at site scale is the absence of monitoring pre-and post-site development. In small catchments or site level scenarios there may not be any rain gauges close to the site, or land use maps may be outdated or too coarse to capture smallscale changes. To overcome these limitations, some studies incorporated publicly available datasets into models to validate land use (Towsif Khan et al., 2020), while some others used rainfall estimates from nearby basins (Ferguson & Fenner, 2020a). If appropriate, using rainfall from nearby basins could be an effective method of estimating model inputs, but it is not an ideal scenario and increases model uncertainty.
Adding to this, a lack of calibration data (n = 21) was also noted as being a significant limitation. Lack of calibration data often stems from an absence of site-specific data, such as the outflow of a certain NFM feature (Lin et al., 2021), or lack of data at the catchment scale, such as the absence of rain and flow gauging stations in smaller or more rural catchments (Ferguson & Fenner, 2020c). In most cases, while there was some form of data for calibration, this tended to be limited and only for a small number of low-intensity events (Gülbaz & Kazezyilmaz-Alhan, 2014;Schubert et al., 2017). There may also be issues with gathering calibration data in low and mid-income countries as there is less funding available for monitoring and gauging. Given that this lack of Note: Some articles appear multiple times in the table as they model multiple NFM interventions within the study. Hydrological models are again included into 2D, 1D and 1D-2D coupled scenarios as appropriate.
gauging and calibration data increases the uncertainty for modelling (Norbury et al., 2018), some models use more statistical methods or alternative modelling results as their basis for model calibration.
Another issue around data is the lack of high-resolution DTM or DEMs (n = 13). This was apparent in many different settings of both rural and urban areas, and even occurs in data-rich regions, like the UK. However, England does have very high-resolution terrain data, 1 m coverage in most areas, with a view to expand this to the whole of the UK region (EA, 2022). The issues with resolution have a knock-on effect across different scales. While finer resolutions are better for improving the accuracy of models, they tend to be harder to obtain at larger scales. Furthermore, while finer resolution data do improve accuracy, they also require a large amount of computational power in order to run models in a reasonable time. This leads to trade-offs between reasonable model run time and coarser resolutions at scale. Highresolution DEMs or DTMs are particularly important for capturing small-scale processes, as well as NFM features themselves. Some features may be less than 1 m wide, such as in-channel leaky barriers or small RAFs, and therefore a coarse resolution may miss these features (Yin et al., 2020). Coarse data resolutions make it difficult to accurately represent the location of features at the larger scale (Juan et al., 2017), and so placements of features become wider and less focused. As a result, the outputs from the modelling may become more uncertain, and less representative of real-world scenarios.
Data challenges are being addressed around NFM through calls from the academic community, for example, monitoring pre and post site development for more accurate data for calibration (Black et al., 2021). However, in many cases these efforts are still lacking. Technological advancements for better data gathering, such as LIDAR and 'Structure from Motion', are slowly improving. Gradual technological improvements and data availability will allow for more accurate models. However, with an ever-changing world faced with climate change and other uncertainties around the longevity of NFM sites, it may become more difficult to represent the complexity of site features than first anticipated. DEMs of NFM sites may need to be updated more regularly to ensure that changes, such as those related to climate change, are captured and monitored.

| Model and computational limitations:
Computing power, complex interactions, and model limitations were also identified as sub-themes of this major limitation. 1D-2D coupling-related challenges were also identified as a small theme within this section as there are only few (n = 2) studies that discuss these challenges.
The limitation related to computational power (n = 7) was largely observed in relation to 2D modelling rather than 1D in channel modelling. The problem of computational power with 2D models increases as the simulation scale gets larger, which in turn demands increasing computational requirement and highresolution data to represent complex interactions (Juan et al., 2017). This limitation is compounded with the already existing data limitations and limited evidence base on NFM site-scale processes. In addition, 1D models may have computational power limitations if the computation covers vast river drainage networks over a large domain, but with technological advancements these will likely no longer be a challenge. As more detailed data, and higher resolution DTMs are used in modelling, there will also be a need for more complex approaches or formulations to capture how features work, as well as loss processes, such as evapotranspiration. Improved formulations will ensure that models accurately represent realworld scenarios, as well as correctly reproducing how flows around NFM features work. Capturing system complexity will not only require increased computing power but will also necessitate more efficient data processing. Adopting high-performance computing hardware, with much greater computational power, will help to reduce run times to acceptable levels.
Capturing complex flow processes and interactions (n = 13) also increases the computational challenges of modelling. For example the challenge regards integrating the impact of initial soil moisture (Rosa et al., 2015) and water table on evapotranspiration within models (Zhang & Chui, 2020). These parameters are very important for reducing uncertainties in certain modelling exercises. These uncertainties come from the fact that hydrological processes can be very important in smallmedium scale flood events which impact the effectiveness of NFM sites for holding water and allowing infiltration. In intensive events this might be less important as the soil will be heavily saturated. Representing hydrological processes more accurately will help to reduce uncertainty.
Storage ponds are another feature that is difficult to represent both offline and online in 1D models of pond networks (Nicholson et al., 2019). For example, the turbulence effect of the flows entering the pond is not always considered. Furthermore, the impact of sedimentation on the effectiveness of storage ponds, or ponds experiencing different levels of water through stages of flooding, has only been mentioned in a few papers (Elosegi et al., 2017). Similarly, the hydraulic functionality of leaky barriers in flood scenarios is another challenge that researchers are attempting to understand and model. Leaky barriers do not behave like traditional weirs or dams, and in some models, there are no leaky barrier units to represent this (Wilkinson & Jackson-Blake, 2017). This usually means that leaky barriers are represented using culvert sections, permeable walls, or DEM adjustments, which is not entirely accurate (Senior et al., 2022). The importance of distinguishing between traditional infrastructure and leaky barriers in models is that they interact with flow differently, which impacts the effectiveness of these features for reducing flooding. Compared to some hard engineered structures, NFM features tend to promote greater infiltration. When soil moisture is particularly high there may be less infiltration, and the site may be less effective at reducing flood risk. Whilst it is important, representing these complex interactive processes will inevitably complicate the model formulation and structure, subsequently increasing computational cost and solution uncertainty. In systems where multiple features are modelled, it is important to ensure that as many complex interactions between and within features are captured, to improve the accuracy of the model.
The final computational limitation (n = 12) is around the difficulty to represent certain features or events within models and a wide range of challenges were identified. Krivtsov et al. (2020) for example, highlighted the inability to model high intensity rainfall events, while Yazdi and Khazaei (2019) and Miguez et al. (2009) highlighted the inability to model certain features and parameters, such as land use. These limitations are often related to the model being too bespoke (Welton et al., 2017), without the option to include more general features. Alternatively, models can be too general, and so miss the complexities of modelling specific NFM features. As most models are built for a specific purpose, limitations are often reported and accepted as part of the modelling process.
A more specific limitation of the modelling process is associated with uncertainties of the method of either 2D cascade coupling, or 1D-2D coupling (n = 2). In both cases, coupling increases the overall uncertainty of the model. In the case of 1D-2D coupling, flows enter or leave the 1D model as a lumped unit at specific points . However, this is not an accurate representation of how water enters or leaves a channel and does not conserve the momentum of the flow at points along the channel (Betsholtz & Nordlöf, 2017). This can lead to an inaccurate prediction of flow velocities, and a poor representation of the overall flow process. Furthermore, 1D features represented in the channel may have spills round the side of the feature and cannot account for the complexities of flows around the features. In the case of 2D cascade models, the issue of momentum conservation and flows around in-channel features is overcome through 2D hydraulic modelling, but there is still increased uncertainty caused by feeding the two models into each other based on simple assumptions. This uncertainty may also make the model less stable than a fully integrated 2D hydrological and hydraulic model.

| Limited real-world and stakeholder engagement: LID features pre-designed
The final major theme related to model limitations is around the applicability of the modelling to real-world scenarios. This was identified as a major limitation as a model may represent a system, but if this is inaccurate or not reflective of the real world then the model may only be speculative rather than substantive. Stakeholder and practitioner engagement prior to NFM modelling or intervention is particularly important when it comes to ensuring that a project is successful or not (Smith, 2017;Wentworth, 2011). Current practice varies from authority to authority and there is no universal way of approaching stakeholder engagement as each catchment is unique. Commonly, in the UK context, it appears that landowners are consulted when it comes to the locating and feasibility of sites, but prior to this in some of the NFM opportunity mapping stages landowners are not consulted. This can mean that while areas are identified as having potential for NFM, this may not be possible to implement in a 'real-world' situation due to landholder constraints. The importance of stakeholder, landowner, and community engagement is widely understood by flood authorities, but co-production of models with local communities and landowners has been seldom explored.
Real world applicability (n = 11) was highlighted not only as a problem of stakeholder engagement, but also of coarse resolution representation. In scenarios where the resolution was too coarse, it becomes difficult to show the precise locations where NFM sites could be placed. Instead, rough coverage estimates can be used, for example, 50% green roof cover (Her et al., 2017). This can be useful for city level or catchment scale planning more generally but does not provide practitioners with enough detail for a targeted site level approach for catchmentscale FRM. A lack of stakeholder engagement can have a similar effect as poor resolution because it does not allow practitioners to target areas for intervention.
Furthermore, although the articles highlighted in this review do not mention the impact of seasonal change on in-channel and overland vegetation cover, it is important to recognise this is also a challenge and limitation in modelling NFM sites. Typically, modellers will use one set of Manning coefficients to represent roughness when modelling NFM interventions. However, there have been a multitude of overland flume (Bond et al., 2020;Holden et al., 2008), and in-channel (Chien, 1956;Cotton et al., 2006) studies which detail that seasonality can have a diverse impact on flow velocities and channel capacity. In particular, Bond et al. (2020) found that winter vegetation cover allowed a significantly higher runoff velocity than summer vegetation. Modellers should aim to integrate seasonal changes of the site into their modelling strategy, for example, by using different values for Manning coefficient, in order to better represent the effectiveness of NFM measures seasonally.
The final real-world limitation is down to model assumptions. Some models, such as SWMM, use predesigned NFM features which have been pre-calibrated with equations, and set parameters relating to their effectiveness and size. Pre-designed NFM features (n = 13), such as bioretention cells or green roofs, are typically found in 1D models. Modellers have highlighted that predesigned features can be poor representations of the real world, and typically are designed for specific scenarios and climates. Some pre-designed features have been calibrated against real-world scenarios (Zhu and Chen, 2017), but this has been found to be challenging at the small scale as many sites are not monitored at the feature scale. This makes it hard to validate the model outputs against real-world data (Lin et al., 2021). Furthermore, many pre-designed features are parameterised for certain climates. When modelling these features in different climate settings, the accuracy of the efficacy of these physical features may differ more widely when comparing model outputs to monitored data.

| Improving set-up and calibration data
Improving the input data is one way to reduce the uncertainty and improve the overall modelling outputs. Modern techniques could be employed within the data collection process to ensure the best data possible is collected. 'Structure from Motion' (SfM) is one way of improving the resolution of the DTMs inputted into the model. SfM uses images collected from the flight of UAVs to build high-resolution (<10 mm) DTMs of sites. Airborne LiDAR tends to be less accurate, and is available at a lower resolution (Rogers, 2017). These high-resolution data are particularly important when representing NFM features, as many sites contain small scale features that can be <1 m in width. This sort of data improvement will be particularly useful for 2D models that simulate overland flow processes. Combining this strategy with GPS points for validating channel cross sections, pond depths, and water levels will also help to aid 1D models that rely on cross sectional data. Furthermore, if urban settings are mapped to provide highly accurate measurements of features, the LID modules in models, such as SWMM, can be better represented in these settings where the data may be currently unavailable.
The accuracy of land use data is also important when modelling the impact of NFM at larger scales. Enhancing drainage through interventions, such as tree planting or afforestation, is a typical NFM method, and so we must improve the accuracy of land use maps to reflect changes. A way to do this is through thorough ground truthing of land use at the site, and post-site construction assessments. Furthermore, existing public land use datasets should be regularly updated by trusted organisations, such as public bodies or NGOs and universities, in order to keep more up-to-date records of land use change. Private firms could also be a trusted source, but this is dependent on the prior bias of why the data were collected. However, an issue with traditional ground truthing is that it can be labour-intensive and costly (Hanuschak & Mueller, 2002). Modern methods offer a potential solution, Yan and Ryu (2019) combined deep learning algorithms with Google street view images to automate the process of land use ground truthing speeding up the process and reducing the costs. Land use data are particularly useful for 2D modelling as for overland flows the type of land use will impact the roughness coefficient, and in turn the amount of runoff generated from these areas. This may also be particularly useful in any models looking at water quality or sediment loading as land use change and accurate prediction of flow processes will impact sediment load as well.
Finally, to improve the calibration and validation data a targeted site-scale approach is required. Firstly, practitioners should aim to monitor sites in a similar fashion to the Eddleston catchment (Black et al., 2021) where preand post-NFM development is monitored, and so a baseline can be developed, helping to realise the efficacy of NFM interventions. However, the baseline monitoring for this site is short. Ideally, practitioners should aim to put in place monitoring systems for many years prior to site intervention to have a thorough understanding of the hydrological processes and variations between seasons and different time periods, but this may not be possible without long term financial arrangements. In terms of monitoring at the site-scale the use of trapezoidal flumes by Q-NFM (Chappell, 2022), or thin-plate weirs by Protect NFM (Evans, 2018), have shown to be effective ways of gauging small-scale sites <1 km 2 . If impossible to implement, due to the site already been constructed, then using the salt-slug dilution method (Moore, 2005), or other similar gauging methods, is still cost-effective for gauging these small-scale sites. Although this does not solve the issue of lacking pre-implementation data, it will still provide the researcher with some flow data at the site-scale to support calibration and validation of both 1D and 2D models.

| High-performance large-scale models
In recent years, the advancement of high-performance computing technologies for scientific computation has allowed for modelling to take place at much larger scales and much higher resolutions, for example, achieved by using modern graphic processing units (GPUs) to speed up the model run times (Hu & Song, 2018). Although there are other modern methods for model acceleration, such as integrating machine learning through artificial neural networks into the simulation process, these are still in their infancy stages and must be further developed before they can become widely useable in simulating NFM features, and the underlying complex flow interactions (Jamali et al., 2021). The recent development in high-performance physically based hydraulic or hydrodynamic models have allowed flood modelling to be applied in entire catchments at unprecedented high resolutions to capture the complex topographies, flow processes, and their interactions. An example of one of these models is the High-Performance Integrated Hydrodynamic Modelling System (HiPIMS) (Liang et al., 2016;Smith & Liang, 2013) which solves the full 2D SWEs using modern numerical methods on multiple high-performance GPUs. HiPIMS is able to predict full-scale flooding process from rainfall to overland flow and inundation over large catchments or cities. In an application to the 2500 km 2 Eden catchment in the UK, the model was able to reproduce the full process of a large flood event in real time at a resolution of 5 m (Xia et al., 2019). These highperformance models could be part of the solution to breaking the current <100 km 2 limit (Wingfield et al., 2019) that fully integrated 2D models have met when modelling NFM at a catchment scale. Harnessing the advantages of the latest high-performance technology, these models can represent features and simulate flow processes and interactions at very high resolutions. This is key when trying to integrate the NFM features of different size and form, such as leaky barriers, into models, and should ensure better accuracy when modelling these sites, while showing impacts of NFM sites on catchment wide flood risk. Furthermore, by modelling these features at the catchment scale the knock-on impacts of NFM to flood peaks can be realised, and such the siting of NFM features as to not increase flood risk can be explored further (Reaney, 2022). 4.3.3 | Complex processes at small scale NFM sites are complex due to the integration of features of different size and form, and the complex processes that occur are difficult to represent in models. For example, leaky barriers may come in many configurations and sizes, and are also subject to constant change on sites due to accumulating debris and sediment activities. Therefore, there is no standard way of representing them in models. Leakey et al. (2020) conducted lab-based experiments and attempted to use a 1D hydraulic model to investigate and simulate the performance of leaky barriers to identify optimal leakiness. Further work on using more sophisticated CFD modelling to directly consider more complex configurations is being completed to represent these features more accurately (Ahadi et al., 2020). Using such highly detailed results to inform parameterisation of catchment-scale models will improve the overall simulation accuracy of these large-scale models.
For ponds, future modellers should aim to include water level monitoring, in addition to flow rate, in the calibration of their models, as this will be key in estimating the effectiveness of the ponds at the time heavy rainfall events occur. Another area that should be explored further is sediment transportation in CFD and 2D models. An investigation of its impact on the flow dynamics in and around storage ponds should be conducted (Guan et al., 2018). Sonnenwald et al. (2018) have already performed CFD modelling using lab-scale vegetation, as well as different pond configurations to reveal the changes in performance between pond design and vegetation. It was found that vegetation changes and pond shape will affect the rate of flow and roughness. Therefore, it is important to carefully consider these aspects when developing a model. Important flow phenomena, such as turbulence, are not normally considered in current modelling practice, either 1D or 2D. Some of these flow phenomena may affect sediment transport and the near-field flow dynamics around and through the features, which could have great impact on the simulation results and the following assessment of the performance of NFM features.
Overall, lab-based studies will be useful for validating fully 3D CFD models which, in turn, can be used to parameterise complex processes in 1D or 2D models, such as turbulent flows in ponds or around leaky barriers (Wang, 2017). Including these complex flow processes and other relevant factors, such as the effect of soil moisture content and proper representation of feature details, may help improve the simulation accuracy in large-scale modelling.

| Co-production of modelling approaches
Stakeholder engagement is considered to be an essential part of NFM modelling as a majority of NFM interventions require building sites on land owned by a range of different actors from farmers to local communities. This complexity of implementation requires not only communication and negotiation, but also the integration of appropriate partnerships from the outset of a project (Samuels, 2022). When modelling is completed without considerations of where sites can be placed or purely based on recommended areas identified for NFM sites, the solutions may be impractical for real-world application. Co-production of sites and models with government, industry professionals, and local communities has proven to be particularly effective for designing effective flood alleviation schemes (Clarke et al., 2016;Fitton & Moncaster, 2018). In particular, the importance of community knowledge of areas which become inundated during floods and areas of land that cannot be developed due to particular local social values has been recognised and can provide insight at many different stages of the design and implementation of features (Fitton, 2015). McEwen (2011) suggested that an approach that integrates a combination of expert knowledge and lay knowledge is an effective way of designing solutions which are beneficial to both local communities and professionals. Furthermore, modelling also provides an effective way to improve communications with different levels of stakeholders and their inputs can help improve model design and output visualisation to better serve this purpose.
NFM projects, in the UK, particularly affect smallrural communities and farmers as it is their land which is typically being used to implement these projects. A suitable approach may involve co-production of the modelling process with landowners, where they may be able to more effectively target specific portions of their land which may be more effective at storing flood water. Further to this, local knowledge can help to inform modellers on the areas that flood regularly on their land, what processes are important to capture in the flood model, for example, upwelling or Saturation Excess Overland Flow (SEOF) zones, and may also suggest appropriate locations to monitor for flows and water levels. Outside of the UK, a combined top-down and bottom-up participation approach with landowners and communities being engaged in the modelling process may be more appropriate. This is important for ensuring appropriate solutions that can be backed by community support are selected during the modelling process, as some communities may not want certain NFM features being created for social or cultural reasons (Drosou et al., 2019). Local knowledge is key for developing a fit-for-purpose model where it not only communicates a clear message of flood risk, but also provides a better reflection of the reality of flooding for communities.
Furthermore, there are many concerns around who has to pay for the implementation and maintenance of NFM sites and whether landowners should be compensated, and if so by how much. A recent survey on the views and expectations on NFM from many different stakeholders by Bark et al. (2021) found that, although many groups agree that conventional FRM techniques are failing, there is still a need for more partnerships and cross-sector debate around the responsibility of NFM projects in order to move towards large-scale NFM intervention. However, there are examples of small-catchment scale NFM interventions being successful through coproduction.
One example of good practice is that of the work on the Eddleston catchment (69 km 2 ). In this catchment, a locally based NGO, Tweed Forum, was funded by government to lead a NFM trial, bringing to the project local knowledge and credibility within the local community. The hydrological analysis was considered to benefit from effective coordination between academics, local government, residents, and farmers (Black et al., 2021). Landowners were also offered the opportunity to construct the features on their land meaning they are paid for the construction of these features, and the sites were sensitive to the needs of their own land. An improved system of 'Payments for Ecosystem Services' (PES) whereby landowners are paid fairly for providing downstream users with ecosystem services, as well as the maintenance of NFM sites to a standard that ensures they are still effective could be more equitable for service providers and users. However, this does get more complex when upscaling as there are more farmers and bodies who will need to be included in the decision-making process and 'PES' system, but the premise of co-production from the start of a long-term project and payments to farmers for NFM site construction seems to be particularly effective in this case.

| CONCLUSION
To conclude the following key considerations have been identified: • Of the research items identified from the search, 60% referred to 1D models, 20% to 1D-2D models, and 20% to 2D models, with the number of 2D models identified increasing in recent years, which is likely due to data and computational improvements. • All 2D modelling has been completed at a scale of <100 km 2 (Lane, 2017;Wingfield et al., 2019), with a majority being focused on assessing the impact of NFM sites on locations directly downstream of sites rather than the whole catchment. • Major model limitations identified from this review are widely related to data, computational power, and realworld applicability. • This review gives a clear picture of approaches, limitations, and potential solutions to some of the challenges facing NFM modellers today.
In particular, the use of modern high-performance computing technology (e.g., GPUs) for speeding up models (Hu & Song, 2018) and 'Structure from Motion' to get highly accurate DEM or DTM data for use in models (Rogers, 2017) are examples of contemporary approaches available to improve current modelling practice. But some of these individual examples may not have been widely demonstrated or accepted. This review has also highlighted the importance of collaboration and coproduction in the modelling space whereby without discussion between academic bodies, industry experts and local communities' projects are largely idealistic and more speculative than useful for practitioners.
Future modellers investigating NFM should aim to find solutions to specific challenges rather than simply improving the knowledge base. Specific limitations around site scale gauging and data limitations could be solved through modern methods for small-scale data collection. Furthermore, modellers could focus on improving the modelling process by using modern computing methods combined with co-production with local landowners to target sites more effectively when deciding their locations within the catchment. Catchment scale models will be key to improving the evidence base of the effectiveness of NFM for CBFM and hopefully help to support a shift towards practitioners implementing CBFM schemes using multiple NFM sites within a catchment.
There are a few limitations of this review which derive from the screening and eligibility criteria set at the start of the review process. By not including uncalibrated models for instance means that the number of models that represented NFM features was smaller even if the modelling process was similar to models that were included as there was no way of discussing the accuracy of the model to real-world results. Another limitation is that the review only focuses on models for fluvial and pluvial flooding but not coastal flooding. Although some systems are affected by coastal flooding, many NFM interventions focus on reducing the impact of flooding from rainfall events at the catchment scale. Furthermore, the type of modelling assessment conducted is very different for coastal versus riverine flooding. Some publications which entirely focused on lab-based studies were also excluded as although these are important for understanding the processes which take place at the feature level many of them did not include any aspect of 1D or 2D modelling. Finally, the review was limited to only include more natural features rather than more engineered features. For example, studies that only focused on green roofs, rain barrels, permeable pavements and other more engineered SUDS features were not included as these do not fit the definition of NFM and stray more towards green infrastructural features. This means that many papers were identified as non-relevant as they focused purely on green infrastructure features and a review that focuses on modelling purely green infrastructure features or SUDS features may need to be conducted in the future.