TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 1 The use of UAV-derived bathymetric data for hydraulic modelling to inform E-flow assessments Maharaj U.1*, Harvey T.A.1, Pike T.1 and Singh K.R.1 1 GroundTruth, Hilton, South Africa *Corresponding author: udhav@groundtruth.co.za Citation Maharaj, U.; Harvey, T. A.; Pike, T.; Singh, K. R. 2024. The use of UAV- derived bathymetric data for hydraulic modelling to inform e-flow assessments. Colombo, Sri Lanka: International Water Management Institute (IWMI). CGIAR Initiative on Digital Innovation. 24p. INFORMATION Keywords E-flows, HABFLO, SoNAR, LiDAR, hydraulic modelling, habitat assessment Flagship Digital Twin Work package System Monitoring Partners GroundTruth, International Water Management Institute (IWMI) ABSTRACT Environmental flows (E-flows) are crucial for maintaining healthy river ecosystems as an essential part of water resources management, but traditional E-flow assessments that include modelling of hydraulic habitats, often rely on limited, single cross-section data. This study presents a novel approach integrating Sound Navigation and Ranging (SoNAR) and Light Detection and Ranging (LiDAR) data collected using an Unmanned Aerial Vehicle (UAV) to create a high-resolution Digital Terrain Model (DTM) and was carried out for a section of the Olifants River in Southern Africa. The integrated DTM enabled detailed 2-Dimensional (2D) hydraulic modelling using Hydraulic Engineering Centre River Analysis System (HEC-RAS), with the resulting depth and velocity outputs used to visualise the HABitat FLOw (HABFLO) fish and invertebrate habitat classes across the entire reach that was modelled. Additionally, a habitat distribution calculator was developed to determine habitat distributions based on river flows. The longitudinal analysis of habitat distributions for a section of the river revealed variations in habitat class distributions that a single cross-section-based analysis would not highlight, thus providing a more comprehensive understanding of habitat dynamics under varying flow conditions. The successful merging of SoNAR and LiDAR data demonstrates the power of combining UAV-derived remote sensing techniques for characterisation of riverine features. This workflow has the potential to further enhance E-flow assessments, aiding in the development of ecologically sound water management strategies. However, future work should include in-field validation of modelled habitat distributions and the expansion of the methodology to larger areas. INTRODUCTION AND BACKGROUND Rivers are globally exploited for various purposes such as carrying additional increased flows through interbasin transfers, unstainable extractions, assimilation of pollution, etc, often without considering the water needs of the ecosystems themselves (Rai and Jain 2022). For sustainable water resource management, it is crucial to identify a balance between water requirements and the health of the river system for a catchment. E-flows are an important tool to achieve this balance (Rai and Jain 2022). E-flows are defined as water that is intentionally retained in or added to rivers to manage the health of the river; and the specific purpose of E-flows can range from maintaining healthy ecosystems to supporting endangered fish populations (King et al. 2008). Arthington et al. (2018) revised the definition of E-flows to embrace flowing, standing and groundwater dependent ecosystems as well as aquatic systems that alternate between states. The revised E-flow definition is as follows: “Environmental flows describe the quantity, timing, and quality of freshwater flows and levels necessary to sustain aquatic ecosystems, which in turn, support human cultures, economies, sustainable livelihoods, and well-being.” Determination of the flow regime that provides for optimal E-flows is an evolving science, with one approach TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 2 being to model the hydraulic habit as a basis on which to determine the needs of the ecosystem. E-flows are also beneficial for many interconnected components of a river system, including the waterways, nearby wetlands, groundwater, and plants and animals that rely on the entire river network (King et al. 2008). Therefore, E-flow frameworks should be holistic and consider how water flow and other factors interact to impact both the environment and society over entire regions (O'Brien et al. 2018). The current state of practice that is generally used for E-flow assessments in Southern Africa includes the use of single cross-sectional data with habitat distribution models to inform E-flow assessments. In this document an innovative approach is developed to determine habitat distributions using hydraulic modelling and remotely sensed topobathymetric data. There are a variety of methods relating to the quantification of E-Flows and the majority of the methods can be grouped into the following categories: (a) hydrological (Cavendish and Duncan 1986; Milhous et al. 1989), (b) hydraulic rating (Waters 1976; Tharme 1996), (c) habitat simulation (Loar et al. 1986; Dunbar et al. 1997), and (d) holistic methods (Tharme 1996; Arthington 1998). The hydrological approaches for E-flow methodologies generally use historical hydrological data to provide recommendations for E-flows and are generally estimated as a fixed proportion of flow/minimum flow (Cavendish and Duncan 1986; Milhous et al. 1989). A limitation of this approach is that the method is simplistic and does not account for the variability of hydrological processes (King et al. 2008). The hydraulic methods for E-flow assessments generally quantify the relationship between flow and instream resources such as habitats (Tharme 1996). A limiting assumption of this approach is that one hydraulic variable or a group of variables can be used to adequately represent the flow requirements of a particular species (King et al. 2008). The habitat simulation methodologies assess E-Flows using the relationship between flow and biotic response and are sometimes called habitat rating or habitat modelling methods (Loar et al. 1986; Dunbar et al. 1997; King et al. 2008). In these methods, flow (discharge) and changes in microhabitats are modelled using at least one hydraulic variable (Tharme 1996). These variables generally include velocity, depth, or substrate composition. The holistic method was developed in South Africa in the 1990s and the basic premise of the approach is to examine a hydrograph and use expert judgement and available data to determine a flow regime to maintain a river ecosystem in a desired state (King et al. 2008). The outputs from habitat models are generally habitat- discharge curves which can be used to estimate the optimum discharge for target biota (King et al. 2008). It is important to highlight that an aim of E-Flow assessments is to maintain a healthy river, however, it is often difficult to select an appropriate species prior to making a recommendation, and generally very little information pertaining to riverine biota is available in many countries (Richardson 1986; Gan and McMahon 1990). The majority of the methods for simulating habitats assume that it is sufficient to model the reaction of biological responses to discharge using hydraulic variables (Mathur et al. 1985; Shirvell 1986). This level of modelling may not be adequate, and the assumption is limited because these models focus on how the changes in flow affect factors such as water depth and velocity, without accounting for the broader effects on the entire ecosystem. Moreover, when performing ecological studies, the hydraulic data must be interpreted in a manner that is meaningful to ecologists. Many South African studies have specified habitat in terms of habitat classes linked to a range of velocities, depths, and other non- flow-dependent characteristics (Hirschowitz et al. 2007). Poff et al. (1997) proposed that understanding the natural flow regime of a river should be a central goal in efforts to conserve and restore rivers. Therefore, it is important to account for E- flows and E-flows must evolve with human activities to successfully deal with new challenges (Poff and Matthews 2013). Although the techniques and procedures used to advise on E-flows are relatively young (approximately 50 years), there are various different methods for E-flow assessments used globally (King et al. 2008). Hirschowitz et al. (2007) developed the HABitat FLOw (HABFLO) simulation software to provide a working model to automate the prediction of the abundance and composition of fish and macro-invertebrate habitat types. The HABFLO model is used to predict the frequency distribution of hydraulic parameters, viz. depth, and velocity. In practice however, E- flow assessments have relied on measurements taken directly in the river, but these on-site measurements lack sufficient detail to represent the entire river, and the cross section data collected during these surveys might not be easily accessible to everyone who needs it (Singh 2023). Habitat simulation models such as HABFLO follow an instream flow incremental methodology where physical habitat characteristics (depth, velocity, substrate) at a specific point in a river are linked to the suitability of that habitat for target species and life stages (Stalnaker et al. 1995). Traditionally, when performing habitat modelling with the HABFLO model, the model is run based on one or more cross-sections for the site/river. The location of the selected cross sections is generally dependent on conditions at the site when the survey is being conducted. However, the location and number of cross-sections used for the modelling are important when determining the representation and reliability of hydraulic modelling and subsequently the habitat simulations (Bovee and Milhous 1978). Due to limitations associated with obtaining cross section data for 1d modelling and detailed topobathymetric data for 2D modelling, remote sensing techniques for data collection in E-flow assessments are gaining popularity (Singh 2023). TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 3 Stream temperature can cause fish migration and impact fish spawning patterns (Jonsson and Jonsson 2010). Tsang et al. (2021) investigated how climate change will affect fish communities in rivers and they acknowledged that climate change will alter water temperature and flow which will likely change the locations where fish can live in the future. Zhao et al. (2018) found that water flow and temperature affect how quickly pollutants break down in a river and proposed a new method to link fish tolerance to water pollution levels to determine E-flows. The specific aim of this report is to assess the applicability of utilising a Digital Terrain Model (DTM) obtained using data collected with an Unmanned Aerial Vehicle (UAV) with Sonar and Light Detection and Ranging (LiDAR) sensors to perform 2-Dimensional (2D) hydraulic and habitat modelling based on the HABFLO habitat classes for a section of the Olifants River in Southern Africa. AIM AND OBJECTIVES The aim of this report is to use a UAV-derived DTM to perform 2D hydraulic and habitat modelling over a river reach. The following specific objectives are required to achieve this aim: a. Perform Hydraulic Engineering Centres River Analysis System (HEC-RAS) 2D modelling using a DTM generated with LiDAR and Sound Navigation and Ranging (SoNAR) data; b. Develop a methodology to determine and visualise HABFLO habitat classes for an entire river reach (accounting for the spatial distribution of hydraulic parameters); c. Use the developed methodology to visualise HABFLO habitat classes for a river reach using the results generated from the HEC-RAS 2D model; d. Derive a relationship to estimate the percentage distribution of habitat classes over the river reach for different flow values; e. Assess the performance of the developed method compared to the traditional HABFLO model using a single cross-section; and f. Develop a sensor platform to collect water quality data using a UAV and use water temperature data with a habitat simulation model. OVERVIEW OF METHODOLOGY A method was developed to visualise different habitat classes along an entire reach of a river. The process of developing this method has been divided into two phases: Phase 1: Development of a 2-dimensional hydraulic model of a river reach (using HEC-RAS 2D). Phase 2: Development of a 2-dimensional visualisation tool for the distribution of HABFLO habitat classes along the river reach using the results from the HEC-RAS model. The two phases are described (with examples) in further detail in the following sections. In this report, the visualisation of habitat classes for fish and invertebrates, based on depth and velocity raster layers generated from a HEC-RAS model (HEC 2024), was done using an adapted methodology based on the Weighted Usable Area (WUA) concept used in hydraulic models such as River2d and PHABSIM (Steffler and Blackburn 2002; Gard 2009; Roh et al. 2011). To determine fish habitat classes, both the depth and velocity result layers are required, and for invertebrate habitat classes, the velocity layer and substrate type of the reach are required (Hirschowitz et al. 2007). A Channel Index (CI) is defined over the mesh of the area being modelled to evaluate the CI suitability (substrate, cover, etc.) for certain habitats (Steffler and Blackburn 2002), and the WUA is calculated using the habitat Suitability (SI) curve and the CI (Steffler and Blackburn 2002). It is acknowledged that generally an SI curve for depth, velocity, and CI for different species at different growth stages is required, but in this report, the HABFLO depth-velocity (for fish) and sediment-velocity (for invertebrates) classes from Hirschowitz et al. (2007) were used for consistency and comparability with the HABFLO model. The WUA is a suitability index that accounts for the suitability of the entire river reach for a discharge value (Steffler and Blackburn 2002). The WUA should be evaluated at different flows for different species and different life cycles. It is noted that prior to the calculation of the WUA, a steady or transient state solution is required for River2D or PHABSIM (Steffler and Blackburn 2002; Gard 2009; Roh et al. 2011). The WUA is generally calculated using the product, geometric mean or minimum methods using the SI and CI as inputs (Steffler and Blackburn 2002). DATA COLLECTION The data required to achieve the aim and objectives of this study were collected during a site-visit that was undertaken to conduct a topobathymetric survey of the Olifants river downstream of the Balule Bridge as shown in Figure 1 below. When using a UAV for data collection, it is important to highlight that the quantity of data collected depends on a multitude of factors, including but not limited to, weather, safety, landowner permissions, electricity supply, etc. The aim of the surveys was to collect data to generate a DTM together with Red, Green, Blue (RGB) Orthomosaic and multispectral imagery. LiDAR data were collected using a Zenmuse LiDAR module (DJI 2024) with an RGB camera, and the multispectral imagery was collected using a Laquinta sensor (Laquinta 2024). An easily accessible low cost and over TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 4 the counter Ping echosounder and altimeter (Robotics 2024) was used with the UAV to obtain below water depths for the Olifants river with the aim of generating a complete terrestrial and riverbed profile for a portion of the river, as the LiDAR beams do not penetrate water. The Bluerobotics Ping echosounder and an ultrasonic sensor were attached to a flotation device designed by GroundTruth, which was pulled over the water surface using a DJI Matrice (M300) drone (see Figure 2). The data that were collected and the equipment used to collect the data are provided in Table 1. Table 1. Collected data during the field visit Figure 2. SoNAR device integrated with a UAV (Authors' Figure 1. Locality Map of Study Site downstream of Balule Bridge on the Olifants River, South Africa (Authors' Creation) Data Type Instrumentation LiDAR data Zenmuse LiDAR module Mulitspectral imagery Laquinta sensor RGB imagery Zenmuse LiDAR module Below water depths Bluerobotics Ping echosounder and Sontek Acoustic Doppler River discharge MF Pro, and flow gauge reading GPS data Catalyst and Neo 6M GPS module Cross section data Total station and level staff TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 5 Creation) Simultaneously, a research team from the University of Mpumalanga used a SonTek Acoustic Doppler to collect bathymetric data on the same river reach. In addition, a cross sectional survey of the Olifants River was performed using a Total Station survey instrument. The SoNAR devices measure depth relative to the position of the device and the DTM generated by the LiDAR provides points with an altitude above sea level. Therefore, the Global Positioning System (GPS) points associated with the elevation data from the LiDAR survey and the depths from the SoNAR device were linked. The depth below the water surface was then estimated by subtracting the depth from the SoNAR device from the elevation from the LiDAR data. Thereafter, an interpolation procedure was performed to generate an underwater profile of the riverbed. A 25cm resolution DTM was thereafter derived using the information collected by all three surveying techniques. Where the initial resolution was greater from the SonTek and SoNAR devices, these points were interpolated to conserve the final resolution of the LiDAR data. The DTMs were compared to identify the highest level and most accurate representation of the riverbed. In addition, an MF pro and a Global Water flow probe were used to collect depth and flow velocities for points along the Olifants river. The discharge measured by the MF Pro was subsequently verified using the depth data collected from the Department of Water and Sanitation (DWS) gauging Weir (B7H026). During the site-visit, a cross-sectional survey was taken across the width of river using a total station. This information was used to verify the elevations of the processed DTM. The data that were collected were used to set-up and run the HABFLO model and the SoNAR and LiDAR data were merged to generate a DTM of the riverbed and floodplain. The DTM was used for the HEC-RAS 2D hydraulic modelling and the results from the hydraulic modelling were then used to visualise the HABFLO fish and invertebrate classes for the section of the Olifants river that was modelled. In addition to the data collected for Phase 1 and 2, additional water characteristics data were collected for the Lions River in KwaZulu-Natal. The results that were obtained are presented and discussed in the following sections. PHASE 1: DEVELOPMENT OF A HEC-RAS 2D MODEL The HEC-RAS software has been developed by the U.S. Army Corps of Engineers and can simulate unsteady flow through a full network of open channels, floodplains, and alluvial fans (HEC 2024). The unsteady flow component can be used to perform subcritical, supercritical, and mixed flow regime calculations to provide a precise depiction of the site’s properties and hydraulic characteristics (HEC 2024). In a HEC-RAS 2D model, several key inputs are required to simulate the flow of water in a more detailed and spatially distributed manner compared to a traditional 1-Dimensional (1D) model. The main inputs to a 2D model are the terrain data, hydraulic roughness co-efficient, geometric data, and boundary conditions. Further details pertaining to the model inputs for this report are described in the sections below. Terrain Data For any 2D surface flow model, a DTM of the terrain to be assessed is required. For this particular model, GroundTruth carried out an independent UAV (drone) based LiDAR survey of the focused reach of river to be modelled. This survey provided detail on the river floodplains, sandbars, rock ledges, and vegetation, to provide elevation detail and high-resolution RGB imagery for holistic visual information that is not always gathered during a site visit. Additional details regarding the collection of data are provided in the Data Collection section. Hydraulic Roughness Coefficients A hydraulic roughness layer or Manning’s roughness co- efficient (n) layer is required for a 2D HEC-RAS model. This layer is used to describe the hydraulic resistance to the flow of water over different surfaces and is essential to calculate river velocities (HEC 2024). The initial land cover descriptions, which were based off the RGB imagery and general site knowledge from the field work, with the corresponding Manning’s n values (Chow 1959) were developed and have been provided in Table 2 below. These were subsequently refined during the calibration process following the initial model run. Table 2. Table of initial Manning’s n Values Geometry Data Detailed information related to the geometry of the site has to be included in the model, and the geometry file uses the elevation detail of the associated terrain file. A 2D flow area then needs to be established to indicate the area that the model represents. A grid size is established for the 2D flow area, which affects both the accuracy and stability of the model. Within each grid cell, the terrain resolution (derived from the DTM resolution) is fully conserved and thus is not reduced to the 2D grid resolution (Kleinschmidt 2023). The result layers, however, are averaged within each cell, where each cell Description Manning’s n River 0.055 Bare Soil 0.034 Grassed Bank 0.045 Light Shrub 0.04 Dense Shrub 0.05 TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 6 contains an averaged depth (m), velocity (m/s), and water surface elevation (m). Therefore, for areas of the terrain that contain inundations, levees, drains, islands, etc., smaller cell sizes are better in terms of accurately representing hydraulic connectivity across such features, rather than greater cells which overlap the feature. The stability of the model is guided by the Courant Number, which is determined by the computation interval (described in further detail below), the geometry cell size, and the modelled wave velocity (HEC 2024). The Courant Number equation is provided below: where: C = courant number, V = wave velocity, ∆T = computation interval or time step, and ∆X = 2D area cell size. A Courant Number of 1 implies that a particle of water will move from one cell centre to the next cell centre in an amount of time equal to the computation interval. For C > 1, there is a risk of numerical diffusion errors, and large diffusion errors lead to instability in the model and inaccurate results (Kleinschmidt 2023). To manage these instabilities, a larger grid size is generally used for the Geometry File and refined for areas with higher variations in the terrain. The refinements are done using break lines and refinement regions. Since, by default, HEC-RAS creates a structured 2D grid and the results are averaged throughout each cell area as mentioned above, these geometry editor functions within HEC-RAS allow the user to restructure the grid to accurately represent the averaged depths and velocities over the terrain. For example, should a large cell size cover the entire cross-section of a levee (Figure 3, top), the averaged results on one side of the levee will be represented throughout the cell, and thus the model will show flow accumulating on the opposite side of the levee, whereas in reality, this may not occur. Therefore, the inclusion of a break line will manipulate and restructure the cell to lie perpendicular to the levee (Figure 3, bottom) to avoid “jumps” in flow and more accurately represent the hydraulic connectivity between cells. Boundary Conditions In HEC-RAS there are five boundary types: flow time series, stage time series, normal depth, rating curve, and precipitation. For this model, a flow time series was used as the inlet flow, and a normal depth was used as the outlet. In this report, the results from the HEC-RAS modelling were used to determine habitat distributions over a reach. Therefore, it was decided that the most efficient way of modelling flow through the HEC-RAS model would be to undertake one simulation with flow increasing with time. Therefore, instead 𝐶 = 𝑉 ∆𝑇 ∆𝑋 Figure 3. Structured grid (top) and manipulated grid (bottom) through the use of a break line from the HEC-RAS model developed for this research (Authors' Creation) of using a steady flow hydrograph which represents one flow and thus will require multiple model simulations for the habitat distribution results, a straight-line hydrograph was inserted into the model (see Figure 4). This hydrograph will relay to the model that the flow entering the system increases incrementally over a prescribed time step. This, however, represents a constant change in flow through the reach of river from the inlet to the outlet until the run is complete. As the habitat distribution study requires the hydraulic characteristics of one specific flow through the entire reach, this would not provide the required results (i.e. at any specific time, the flow rate entering the system would differ from that exiting the system). To develop a more singular representation of flow through the system at any given time, an incremental time step was used to represent the time it would take both low flows and high flows to travel down the river reach. Thereafter, constant flow was entered between the incremental time step until the inlet flow was equal to the outlet flow. The flow was increased incrementally each time the inlet and outlet flow were equal. This then developed a stepped hydrograph (see Figure 5) to ensure that the flow exiting the system matched the flow entering the system before the flow was increased. TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 7 Figure 4. Straight line hydrograph from HEC-RAS to model a range of flows through the system in one simulation (Authors' Creation) Figure 5. Stepped hydrograph from HEC-RAS representing a constant flow entering and exiting the system until the next increment (Authors' Creation) The energy slope associated with the flow boundary is also required when using HEC-RAS. Therefore, the longitudinal water surface slope that was measured on-site was used and verified using the LiDAR data. The outlet boundary condition for the site was taken as a normal depth condition. A normal depth condition implies that the flow will exit the model based on the hydraulic gradient generated through the water surface elevation and the user-defined energy gradient. Model Simulation The model was run using an unsteady flow analysis tool which uses the Conservation of Mass Principal and the Shallow Water Equation or Momentum Conservation Equation, which is then simplified to a Diffusion Wave Equation (Kleinschmidt 2023). The 2D area created in HEC-RAS applies the conservation of mass principal where volume is neither lost nor destroyed and therefore the volume entering each 2D cell equals the volume leaving and storage maintained (Kleinschmidt 2023). The computation then uses the Momentum Conservation Equation which relates changes in velocity to internal and external forces on the fluid caused by hydrostatic pressure, turbulence, and friction. These equations are then vertically averaged throughout each cell to yield 2D Shallow Water Equations. When first applying the model, HEC-RAS makes the following default assumptions: (a) HEC-RAS assumes a low turbulent flow, therefore the eddy viscosity function is not turned on, and (b) there is no acceleration in the system, therefore the Momentum Conservation Equation is reduced to the Diffusion Wave Approximation Equation which HEC-RAS uses by default (Kleinschmidt 2023). Based on the complexity of the river reach and the purpose of the study, it has been assumed that flow acceleration, eddy viscosity, and more detailed flow characteristics are required for E-Flow determination. Therefore, the simulation was first done with the default Diffusion Wave Equation, and thereafter, once the model was finalised, the Full Momentum Equation (Shallow Water Equation) was used and compared. When setting up the unsteady flow analysis, the following computation settings are required: • computation interval; • mapping output interval; • hydrograph output interval; and • detailed output interval. As mentioned in the geometry data section, the computation interval determines the stability of the model, therefore, the computation interval was initially set to 1 second and thereafter reduced until both the volumetric errors and courant numbers were acceptable. The mapping output interval was set to 1 hour because the stepped hydrograph was developed using 1-hour increments. This was done because the result layers required for determining the habitat distribution (documented below) is a singular raster file where the flow entering the system has reached a state of equilibrium (or steady state) to the flow exiting the system. The flow detail until the equilibrium state has been reached is irrelevant to the E-Flow results. Hence, both the output hydrograph intervals were also set to 1 hour. An initial condition was set prior to the flow simulation using an initial time of 4 hours, as this was the estimated time for the lowest flow to reach the end of the system. Therefore, prior to the simulation of the provided unsteady flow, the simulation was set to run for 4 hours of the hydrograph to allow for an initial wetted condition in the reach of river. This was done to avoid an unrealistic case of flow entering a dry/empty system. Subsequently, as mentioned above, the equation was changed from a Diffusion Wave equation to the Full Momentum equation for the final run. PHASE 1: HEC-RAS 2D RESULTS The Diffusion Wave model ran for 29 minutes, without error, and the simulation only generated an overall volume accounting error of 0.000010% and a maximum courant number of 0.76 (lower than 1 as required for the Diffusion Wave equation). The Full Momentum model ran for 1 hour, 37 minutes, without error, and the simulation generated an overall TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 8 volume accounting error of 0.000370% with a maximum courant number of 0.52 (lower than 2 as required for a Full Momentum equation). This indicates that little to no instability occurred during the simulation and the equation results are relatively high in accuracy. The overall representative accuracy however is dependent on the user inputs such as the terrain data, roughness co-efficient and the geometry data. The comparisons between the Diffusion Wave and Full Momentum equation simulations showed very little difference in flow characteristics. There were certain areas around islands that showed slight changes in particle tracing (i.e., eddying), however, in terms of depth and velocity, there were no significant differences. This indicates that the Diffusion Wave equation method is sufficient in accurately representing flow characteristics for habitat distribution classification, which is more feasible than running the Full Momentum method based on faster simulation times and volume accounting errors. Once the model had completed its run, the simulated results are displayed on RASmapper, which, by default display the flow depth, velocity, and water surface elevation (Figure 6, Figure 7, and Figure 8 respectively). RASmapper also allows the user to animate a dynamic representation of the flow through the reach of river over the simulation time at increments specified by the mapping output interval. As a result, the user can visually observe one of the above three result layers at any specific time throughout the hydrograph and observe the flow characteristics such as hydraulic conductivity, water surface gradients, etc. The user can then export these results as a Tagged Image File Format (TIFF) or a shapefile to use in Geographic Information System (GIS) software. The user may also generate additional standard result layers such as shear stress, energy, arrival time, stream power, etc. as well as calculate new result layers using the RASter Calculator to develop comparisons, or in the case of this report, the relationship between depth and velocity ranges. Outcomes During the site visit where surveys were undertaken to establish the terrain data as mentioned above, additional data were collected at points along the river reach to develop a good indication of the depth and velocity values for the specific flow rate during the time of the survey. The MF Pro was used to measure the flow rates, however, as an additional check, the depth at the gauging weir directly upstream of the study site (weir B7H026) was measured and converted to a flow rate using the rating table obtained from the DWS. This provided an accurate estimation on the observed flows during the time of the survey. The rating table indicated that the flow during the time of the survey was between 4.208 m³/s and 4.476 m³/s as the flow depth measured at the weir was approximately between 0.24 m and 0.25 m (see Figure 9). The MF Pro measured an averaged flow of 4.528 m³/s across the cross Figure 6. Depth Result Layer from HEC-RAS (Authors' Creation) TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 9 Figure 7. Velocity Result Layer from HEC-RAS (Authors' Creation) Figure 8. Water Surface Elevation (WSE) Result Layer from HEC-RAS (Authors' Creation) TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 10 section surveyed. This indicates a strong correlation between the MF Pro measurements and the gauging weir, and therefore, the measured flow in the modelling was taken as 4.5m³/s. At this flow, the modelled depth and velocities were compared to the point data collected during the survey to ensure that the model was representative of the river. The results of which have been described below. Figure 9. Gauging weir B7H026 (Authors' Creation) Several iterations were done using the geometry and the Manning’s n layer to obtain the most accurate and realistic representation of the observed data. The accuracy of the results was computed following the final model run. This was done by selecting the measured flow as mentioned above and comparing the corresponding depths and velocities between the simulated results and the on-site observed data. Where the DTM evidently matched the terrain on site (this was determined based on the channel profile and what could be seen through imagery), the differences between the depths and velocities were on average between 0.02 m and 0.2 m/s respectively. Based on the HABFLO ranges, these differences are small enough not to have a large effect on the habitat class determination. These correlations made up 40% of points collected. The other 60% of the points collected differed substantially, with the greatest difference = 0.44 m. This difference could result in an inaccurate representation of habitat class classification and distribution over the river reach. When investigated, it is concluded that the inaccurate points were mostly due to inaccurate bathymetric data. Along the measured cross-section, the points which correlated well lay on the cross-section where the DTM matched the manual survey. However, where the DTM differs from the manual survey, the depth and velocity results are inaccurate. The comparison between the DTM and manual survey can be seen in Figure 10 below. Figure 10. Difference between total station survey and cross- section of DTM (Authors' Creation) For results that differed from the observed data and were not close to the cross-section, it was seen on the imagery that the bathymetric survey did not pick up certain features such as sediment plumes or inundations. As shown in Figure 11 below, Point 51 represents a good correlation between the DTM and what the terrain is evidently showing on the imagery. The depth and velocity results from Point 51 also matched well between the modelled and measured results with a difference in depth of 0.07 m. Points 49, and 50, however, do not represent the terrain accurately when visually compared to the imagery. This inaccuracy shows strongly in the depth and velocity results, where the depths differed by 0.21m and 0.26m respectively. These outcomes indicate that with accurate bathymetric data and real-time site data, a 2D HEC-RAS model can accurately represent the flow characteristics (depth and velocity in particular) of a reach of river. The feasibility of collecting accurate bathymetric data, however, will determine the feasibility of carrying out the approach described in this report. This will depend on the cost of the equipment required, the man-hours and effort required to obtain the data, and the overall scale of the project and available budget. Table 3 below represents a feasibility matrix of the above factors for three survey options, namely, Green LiDAR; SoNAR/ Acoustic Doppler etc.; and manual surveys using Real Time Kinetic (RTK) systems/Total stations etc. Although it has been acknowledged that the application of the SoNAR bathymetric survey provided limitations to the accuracy of the HEC-RAS results, for the purpose of the continuation of the study, these results were used to develop the proof of concept that a suitable and feasible system can be developed for modelling habitat distributions over a reach of river using a 2D-hydraulic model (on the basis that feasible detailed bathymetric data can be obtained). The following sections provide the approaches used to develop HABFLO habitat distributions of the river reach using HEC-RAS result layers as an input to the process. TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 11 PHASE 2: GENERATING A VISUAL REPRESENTATION OF HABFLO CLASSES OVER A REACH Fish Habitat Classes In this section of the report, an overview of the process used to create a visual representation of fish habitat classes and the results from the HEC-RAS 2D model for the reach is described. Two approaches, i.e. GIS and hydraulic modelling, were applied. The background, limitations and results for each approach are detailed below. Geographic Information System (GIS) Approach An adaptation of the WUA method was generated to utilise the outputs from HABFLO (Hirschowitz et al. 2007) and HEC- RAS (HEC 2024) in Quantum GIS (QGIS 2024) using the steps detailed below: 1. Load the velocity and depth raster files for a specific river flow rate (Q) into QGIS and filter out any “No Data” cells. 2. A new raster was created to identify the overlaps within the HABFLO-specified depth and velocity ranges shown in Table 4. 3. The output of the raster calculator is a binary raster with values of 0 (no overlap) and 1 (overlap). Therefore, the outputs were reclassified to represent each HABFLO class. Thereafter, all seven layers that represented each of the seven HABFLO fish habitat classes shown in Table 4 were merged. 4. The percentage distribution of the different classes for the reach was subsequently calculated as shown in Table 5. Figure 11. Comparison of stitched DTM elevations (left) to those seen through Imagery (right), showing the depth, which was accurately surveyed around Point 51, and the areas which were not covered by the SoNAR survey at Point 49 and Point 50. It is clear that the invert elevation at Point 51 should have continued to Point 50 as the channel bed slopes from Point 50 and Point 51. It can also be seen that Point 49 is higher than what it should be based on the adjacent islands seen in the Image (Authors' Creation) Survey Method Cost of equipment Man-hours required Scale of project suited to method Accurate bathymetric data obtained Green LiDAR High Low High High SoNAR or similar Medium High Medium Medium Manual Survey Low Medium Low Low Table 3. Feasibility Matrix for Survey Methods TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 12 Table 4. HABFLO classes for Fish (Hirschowitz et al. 2007) A limitation of this approach is that the overlap between the fish habitat classes could not be accounted for dynamically in a GIS environment. Therefore, the HABFLO classes were separated based on velocity. As an example, the percentage distribution of HABFLO classes is shown in Table 5. Table 5. Percentage distribution of HABFLO classes Class Name Distribution (%) Distribution SvS 27 0 SS 59 0 SD 14 0 FvS 0 27 FS 0 53 FI 0 9 FD 0 11 Velocity (m/s) Depth (m) HABFLO Min Max Min Max Description 0.0 0.3 0.0 0.1 SvS (Slow very Shallow) 0.0 0.3 0.1 0.5 SS (Slow Shallow) 0.0 0.3 0.5 10.0 SD (Slow Deep) 0.3 10.0 0.0 0.1 FvS (Fast very Shallow) 0.3 10.0 0.1 0.2 FS (Fast Shallow) 0.3 10.0 0.2 0.3 FI (Fast Intermediate) 0.3 10.0 0.3 10.0 FD (Fast Deep) As noted previously, a limitation of this approach is that the overlap between classes is not dynamically accounted for. Therefore, the method was subsequently refined to adopt a more dynamic approach. The methodology and results from the refined approach are provided in the next section. Hydraulic Modelling Approach The GIS approach described in the previous section was refined to be carried out using HEC-RAS to generate the output for a range of different river discharge/flow rate (Q) values for each depth and velocity pixel of the river reach as described in the HEC-RAS 2D section above. A Visual Basic script was written for the Fish habitat classes and applied using the RASter calculator tool within HEC-RAS to assign a HABFLO class to each pixel for the reach using the specific depth and velocity of each pixel for an entire hydrograph (i.e. for different Q values at different time steps). This calculated result layer is shown in Figure 12 below. Figure 12 shows a single time step of the flow hydrograph representing the calculated fish habitat (HABFLO output) distribution over the river reach at a specific flow (Q) for a specific timestep. The model displays a dynamic representation of the distribution of the different habitat classes and how they change over time as the flow increases incrementally. As HEC-RAS requires the user to provide each depth-velocity range a value, this layer shows each habitat class as a single value. Figure 12. Visual distribution of habitat classes for Fish on HEC-RAS where Class: SVS (1-Red); SS (2-Orange); SD (3-Yellow); FVS (4-Blue); FS (5-Turquoise); FI (6-Dark Green); FD (7-Light Green) (Authors' Creation) TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 13 The HABFLO outputs calculated from HEC-RAS were then used to determine the percentage distribution of each HABFLO class for the reach using the following steps: 1. A range of Q values was selected based on the data collected during the site visit and the information obtained from the Department of Water and Sanitation (DWS 2023) gauging weir at the site (see HEC-RAS Result Outcomes section). 2. The ‘No Data’ cells were filtered out from the raster files. 3. Thereafter, the percentage distribution of each HABFLO class for each selected Q value was estimated. The distributions generated in this step are termed the ‘observed’ case in this document. 4. The relationship between Q and percentage distribution for each HABFLO class was determined using regression models to develop a generalised model to estimate the percentage distribution of the HABFLO classes for the river reach. The basic form of the relationship is as follows: Distribution (%) of HABFLO fish habitat class over the reach = f(Q) 5. The model between Q and the percentage distribution was normalised and the results obtained in Step 3 was compared to the results from the models/distribution curve outputs developed in Step 4. The distributions estimated in Step 4 are termed the ‘simulated’ distributions in this document, whereas the distributions estimated in Step 3 are termed ‘observed’ distributions. 6. A distribution calculation tool was created based on the regression equations for the different HABFLO habitat classes. Therefore, if any reasonable Q value is inserted to the calculator, a distribution (%) of each fish habitat class for the reach is calculated. 7. The error between the outputs mentioned in Step 5 were compared using the average absolute error, Nash- Sutcliffe Efficiency (NSE) (Nash and Sutcliffe 1970), correlation coefficient (R²) (Schulze et al. 1995), Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE) (Legates and McGabe 1999; Ritter and Munoz- Carpena 2013; Lal et al. 2016). The selected flow rates, 0.5, 2.5, 4.5, 6.5, and 8.5 m3s-1, and associated distribution layers were exported from HEC-RAS and used to create regression models as highlighted above (Step 3). The HEC-RAS-derived distributions of HABFLO classes for these chosen flow rates have been provided in Table 6. Table 6. Modelled distribution of HABFLO classes across the reach using HEC-RAS layers Flow rates of 11 and 12.5 m3s-1 were then entered into the regression model to simulate the HABFLO distribution at this discharge (Step 5). These simulated distributions have been provided in Table 7 below. The simulated values for these two flow rates were then compared to the actually modelled distributions from HEC-RAS and the average absolute error between the modelled and simulated values were 4 % and 6 % respectively. Table 7. Simulated distribution of HABFLO classes across the reach using regression model It is interesting to note that the regression models were developed using Q values of 0.5, 2.5, 4.5, 6.5, and 8.5 m3s-1, but the models performed well (Errors < 10 % and R2 > 0.8) in terms of estimating the distribution for a reach for Q values of 11 and 12.5 m3s-1 as shown in Figure 13. Figure 13. Comparison of modelled and simulated habitat distribution values for discharges of 11 and 12.5 m³/s (Authors' Creation) A statistical comparison between the modelled and simulated values for all the discharge values is provided in Table 8. Q (M3.S-1) HABFLO Distribution SvS SS SD FvS FS FI FD 11 10% 44% 3% 1% 5% 10% 28% 12.5 8% 41% 3% 1% 5% 10% 33% Q (M3.S-1) HABFLO Distribution SvS SS SD FvS FS FI FD 0.5 36% 56% 5% 2% 1% 0% 0% 2.5 31% 45% 8% 2% 2% 3% 9% 4.5 28% 41% 6% 2% 5% 2% 16% 6.5 23% 40% 5% 1% 5% 6% 19% 8.5 17% 43% 4% 1% 6% 8% 2% TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 14 Table 8. Performance of the model that was developed An NSE and R2 of one and a RMSE close to zero generally indicate good model performance. As shown in Table 8, the correlation between the modelled and simulated values is good (R2 and NSE ≈ 1) and the error between the modelled and simulated data are low (MAE and RMSE ≈ 0). Therefore, the method presented above may be used to generate distributions of HABFLO fish-habitat classes for an entire reach using any discharge value. A comparison between the results generated from the traditional 1D HABFLO model based on the cross-section measured on-site, and the hydraulic modelling method presented in this section is provided in Table 9, below. The comparison between the visualisation of these results has also been shown in Figure 14. It should be noted that the accuracy of the results for this comparison is based on the confidence of the results in the generation of the DTM and the hydraulic modelling as described in the sections above. Nonetheless, when the results obtained (from HABFLO and the hydraulic modelling approach to estimate habitat distributions) were compared to the HEC-RAS model within the areas which correlate well to the observed data, it can be concluded that, for the Olifants River, the 1D HABFLO model Performance Statistic Value Correlation coefficient (R2) 0.953 Mean Absolute Error (MAE) 0.016 Nash Sutcliffe Efficiency (NSE) model coefficient 0.948 Root Mean Square Error (RMSE) 0.035 does not account for certain habitat classes which are clearly apparent in the 2D model over the reach of the river for certain depths of flow. It has also been found that the percentage distribution of a specific habitat class may be low along the 1D channel profile, and high within the surrounding areas of the reach, or vice versa. This is due to large variations in the channel as observed in the Olifants river. In a case such as this, these results may affect an E-Flow determination study when considering fish species, invertebrates, geomorphology, riparian vegetation, etc., whilst a more holistic understanding of the habitat distributions over a river reach may provide a higher confidence level for the specialists involved in the E- Flow decision making. Overall, it was evident that the habitat distributions estimated using both the methods correlate well for low flows. This can be attributed to the possible similarities in channel characteristics at a low flow depth. Therefore, for a uniform channel, one may find that 1D modelling using HABFLO may be sufficient in representing habitat distributions for E-Flow determination. However, as the flow increases, and the more complex river features (sandbars, rock ledges, islands, etc.) become activated within the flow area, the habitat distributions begin to differ significantly between the 1D and 2D hydraulic modelling results, thus concluding that 1D modelling may not be representative for a non-uniform river reach. Additional to this, in a complex river system such as the Olifants, the 2D hydraulic model provides further insight into fish migration opportunities for breeding patterns. In the next section, the GIS and hydraulic based methodology was applied to generate a visual representation of invertebrate habitat classes for the same river reach. Q (m3.s-1) Fish Habitat Distribution Comparisons Average difference across classes SvS SS SD FvS FS FI FD R ea ch C ro ss - Se ct io n R ea ch C ro ss - Se ct io n R ea ch C ro ss - Se ct io n R ea ch C ro ss - Se ct io n R ea ch C ro ss - Se ct io n R ea ch C ro ss - Se ct io n R ea ch C ro ss - Se ct io n 0.5 36% 33% 56% 58% 5% 0% 2% 3% 1% 5% 0% 1% 0% 0% 2% 2.54 32% 14% 46% 53% 6% 0% 2% 7% 3% 11% 2% 8% 9% 7% 7% 4.53 27% 15% 41% 40% 6% 0% 2% 12% 4% 8% 4% 10% 16% 14% 6% 6.6 22% 10% 40% 35% 5% 2% 2% 11% 5% 10% 6% 13% 20% 18% 6% 8.3 18% 8% 42% 30% 3% 4% 1% 10% 6% 12% 8% 9% 22% 26% 6% 11.38 10% 6% 52% 24% 0% 7% 0% 10% 5% 10% 13% 10% 20% 33% 10% 12.56 7% 5% 56% 23% 0% 7% 0% 9% 4% 10% 14% 12% 18% 34% 11% Avg. difference across flow 9% 13% 5% 8% 5% 4% 6% Table 9. Comparison between modelled and simulated distributions and HABFLO distributions TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 15 Invertebrate Habitat Classes In this section of the report, an overview of the process followed to create a visual representation of invertebrate habitat classes over the reach is described. The results are based on a GIS and Hydraulic based approach to determine the habitat suitability of a reach for varying flow conditions as detailed below. Geographic Information System (GIS) Approach The invertebrate habitat classes from HABFLO (Hirschowitz et al. 2007) provided in Table 10 were visualised using the velocity raster layers for different Q values generated using HEC-RAS (HEC 2024) in QGIS (QGIS 2024) using the steps detailed below: 1. A range of discharge values was selected based on the data collected during the site visit and the information obtained from the Department of Water and Sanitation (DWS 2023) gauging weir at the site. 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Pe rc en ta ge Depth of Flow (m) Fish Habitat Availability as a Percentage of Total Area SvS SS SD FVS FS FI FD Figure 14. HABFLO generated habitat classes (top) using the distribution Lookup Table, and HEC-RAS generated habitat classes (bottom) using the distribution calculator (Authors' Creation) TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 16 2. Load the velocity raster files (for different discharges) into QGIS and filter out any “No Data” cells. 3. The sandbars from the imagery were classified as fine sediments and all other areas of the reach were classified as coarse sediments. 4. The sediment layers were classified as follows; fine sediment cells were assigned a value of 1 and coarse sediment cells were assigned a value of 2. 5. A new raster was created to identify the overlaps between the fine and coarse sediment layer and the HABFLO velocity ranges for invertebrates shown in Table 6. 6. The percentage distribution of the different classes for the reach was subsequently calculated. 7. The relationship between discharge and percentage distribution for each HABFLO invertebrate class was determined to develop a generalised model to estimate the percentage distribution of the HABFLO invertebrate habitat classes for the river reach. The basic form of the model is as follows: Distribution (%) of HABFLO class over the reach = f (discharge) 8. The model between discharge and the percentage distribution was normalised and redistributed and the results obtained in Step 3 were compared to the results from the models developed in Step 4. The distributions estimated in this step are termed the ‘simulated’ distributions in this document. 9. A distribution calculation tool was created based on the regression equations for the different HABFLO habitat classes. Therefore, if any reasonable Q value is inserted into the calculator, a distribution (%) of each invertebrate habitat class for the reach is calculated. 10. The error between the outputs mentioned in Step 5 were compared using the average absolute error, NSE (Nash and Sutcliffe 1970), correlation coefficient (R2) (Schulze et al. 1995), MAE and the RMSE (Legates and McGabe 1999; Ritter and Munoz-Carpena 2013; Lal et al. 2016). Table 10. HABFLO invertebrate habitat classes (Hirschowitz et al, 2007) A limitation when performing the mapping of the invertebrate habitat classes is that the level of spatial information for the sediments was not detailed in this report. Therefore, the distribution of the HABFLO invertebrate classes for the different discharges may not be representative of physical reality. However, this was performed as a proof-of-concept study to determine the distribution of invertebrate habitat classes across a reach using the outputs from HEC-RAS. This was done to verify whether the methodology can be applied for future E-Flow studies should detailed sediment information be available. In Table 11, the modelled results that were generated using discharges of 0.5, 2.5, 4.5, 6.5 and 8.5 m3s-1 are provided. Similar to the Fish Habitat Classes above, regression models were fitted to the discharge and distribution data for each HABFLO invertebrate class. Thereafter, habitat distributions were simulated for discharges of 11 and 12.5 m3s-1 as presented in Table 12. The HEC-RAS distribution results were then compared for these flows and the average absolute error between the modelled and simulated values were 4 and 6 % respectively. The regression models were not developed with discharges of 11 and 12.5 m3s-1, but the model performed well (Errors < 10 % and R2 > 0.8) when tested using these discharge values as shown in Figure 15. Velocity (m/s) Sediment HABFLOMin Max 0 0.1 FS VSFS (Very Slow Fine Sediment) 0.1 0.3 FS SFS (Slow Fine Sediment) 0.3 0.6 FS FFS (Fast Fine Sediment) >0.6 FS VFFS (Very Fast Fine Sediment) 0 0.1 CS VSCS (Very Slow Course Sediment) 0.1 0.3 CS SCS (Slow Course Sediment) 0.3 0.6 CS FCS (Fast Course Sediment) >0.6 CS VFCS (Very Fast Course Sediment) Q (m3.s-1) HABFLO Distribution VSFS SFS FFS VFFS VSCS SCS FCS VFCS 0.5 4% 1% 0% 0% 69% 23% 2% 0% 2.5 13% 4% 0% 0% 27% 41% 14% 1% 4.5 13% 8% 1% 0% 17% 36% 23% 2% 6.5 11% 13% 1% 0% 10% 35% 27% 3% 8.5 7% 18% 1% 0% 6% 32% 32% 5% Table 11. Modelled distribution of HABFLO invertebrate habitat classes across the reach TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 17 Figure 15. Comparison of modelled and simulated habitat distribution values for discharges of 11 and 12.5 m³/s A statistical comparison between the modelled and simulated values for all the discharge values is provided in Table 13. Table 13. Performance of the model that was developed An NSE and R2 of one and a RMSE close to zero generally indicate good model performance. As shown in Table 13, the correlation between the modelled and simulated values is good (R2 and NSE ≈ 1) and the error between the modelled and simulated data are low (MAE and RMSE < 0.1). Therefore, the method presented above may be used to generate distributions of HABFLO invertebrate habitat classes for an entire reach using a discharge value. A comparison between the results generated from the traditional HABFLO based on a cross- section and the method presented has not be shown for the invertebrate habitat classes due to the sediment layer being an assumed distribution of fine and course sediment. The HABFLO model accounts for variation in water flow depth and velocity along a cross section, however, there are other factors such as temperature that could also impact habitat suitability. Zhao et al. (2018) stated that both water quality and quantity should be considered when performing e-flow assessments and found that water velocity and temperature influenced the pollution degradation rate in rivers. Tsang et al. (2021) found that temperature changes due to climate change Performance Statistic Value Correlation coefficient (R2) 0.871 Mean Absolute Error (MAE) 0.031 Nash Sutcliffe Efficiency (NSE) model coefficient 0.838 Root Mean Square Error (RMSE) 0.057 will impact and change stream classes for different habitats. Therefore, for this project, a device was constructed and used to collect temperature and river characteristics data using a UAV as described in the next section. INNOVATIVE RIVER MONITORING WITH THE DEVELOPMENT OF UAV-TOWED SENSING PLATFORM AND ADOPTING GREEN LIDAR TECHNOLOGY The development of a UAV-towed sensing system Traditional point-sampling methods limit comprehensive assessments of river health over large spatial areas. This challenge was addressed by GroundTruth by developing a drone-towed sensor platform for spatially extensive water quality data collection. The aim of developing the unit was to improve the data collection process and to further develop the linkages between river flow regimes and water quality parameters. This data could ultimately benefit E-flow assessments. The sensor platform that was built was equipped with the following devices: • Three temperature sensors to capture thermal variations at multiple depths and across the width of the river. • A temperature-calibrated Total Dissolved Solids (TDS) sensor to monitor total dissolved solids, providing insights into potential pollution sources. • A SoNAR device to map variations in the riverbed profile. • A turbidity sensor to measure water clarity, indicating areas of high sediment load. • A GPS module for precise geotagging of all sensor readings. The data logging occurred at 2-second intervals, ensuring the capture of changes within the river environment. All the sensors and microcontrollers required to process the collected data were mounted on a specially built boat that was manoeuvred using a UAV as shown in Figure 16. The UAV- based approach enabled a safe and efficient survey of a river segment. This highlights the potential for expansion in terms of data collection capabilities compared to traditional point sampling methods. The boat also allows for LiDAR and below Q (m3.s-1) HABFLO Distribution VSFS SFS FFS VFFS VSCS SCS FCS VFCS 11 4% 22% 1% 0% 7% 24% 35% 7% 12.5 3% 22% 2% 0% 8% 20% 38% 8% Table 12. Simulated distribution of HABFLO invertebrate habitat classes across the reach TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 18 water surveys to be done by a single pilot with one UAV. The entire system was developed and programmed by the GroundTruth team, and the system was tested in KwaZulu- Natal. The resulting dataset offers unique insight into riverine characteristics for a segment of the Lions River in KwaZulu- Natal. As shown in Figure 17, the temperature varied along the river reach and temperature variations can reveal zones of groundwater influx or thermal stratification of the water column. As shown in Figure 18, the TDS fluctuations may correlate with land-use practices within the catchment and the SoNAR data that were collected are vital for complementing hydraulic and riverbed morphology studies. The mapping of turbidity, as shown in Figure 19 can be linked to sediment transport dynamics. This multi-parameter water quality data can potentially be integrated with E-flow assessments and has the potential to identify areas where flow alterations cause significant changes in water quality and quantify the water quality degradation. This knowledge can be vital for targeted and effective catchment and E-flow management strategies. The temperature data that were collected in this study were subsequently used with the HABFUZZ model and the results were compared to the results obtained from the HABFLO model as detailed in the next section. Figure 16. Custom drone towed sensor platform (left) and during an autonomous flight (right, Authors' Creation) Figure 17. Interpolated temperature maps of a section of the Lion’s River derived using data from the custom drone towed sensor platform (Authors' Creation) Figure 18. Interpolated TDS maps of a section of the Lion’s River derived using data from the custom drone towed sensor platform (Authors' Creation) Figure 19. Interpolated turbidity maps of a section of the Lion’s River derived using data from the custom drone towed sensor platform (Authors' Creation) TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 19 Further enhancement using Green LiDAR Green LiDAR bathymetry is a further innovative river bathymetry technology. Green LiDAR technology is a laser based bathymetric scanning system that’s uses two lasers in the light range: infrared and green. An infrared pulse reflects off the surface of water or land, while a green pulse penetrates the water and reflects off the bottom of a water-body and off the land (Quadros et al. 2008). Singh (2023) reviewed a variety of optical and acoustic sensors to undertake remotely sensed bathymetric surveys. Based on the findings of the literature review and with the objective to select a robust approach that performs well in multiple environments, perform repeatable measurements and to conduct surveys in difficult to access river reaches, the green LiDAR optical approach was chosen for use. The green LiDAR payloads are compatible with many commercial UAV’s which facilitates topobathymetric measurements in non-navigable areas. The findings of the green LiDAR trial highlighted that the UAV-based green LiDAR was able to detect depths of up to 85 m and 2 Secchi Depths under ideal conditions and provides high resolution and accuracies. Given these characteristics the system can be viewed as a cost-effective solution over large land and coastal zones making it an attractive tool for the creation of a digital twin. The primary limitation of the approach is the decreased performance in turbid areas and areas with riffles. An example of a generated DTM generated using a Green LiDAR session is shown in Figure 20. Green LiDAR pulses are able to penetrate water and as such Figure 21 shows cross sections of DTMs generated with and without elevation points that penetrated the surface of the water body to showcase the detail and value that using a Green LiDAR system provides. Figure 20. Example of a DTM generated by GroundTruth using a Green LiDAR Sensor (Authors’ Creation) Figure 21. Cross section of an elevation model with water surface elevation points (top), and without water surface elevation points showing the riverbed profile (bottom, Authors’ Creation) ENHANCING HABITAT MODELLING WITH FUZZY LOGIC AND UAV-COLLECTED DATA Hydrological, geomorphological (above and below the water surface), and water quality data were collected during this study using a range of sensors as detailed in the previous section. The data were collected for the Lions River in KwaZulu-Natal and were used to run the different habitat modelling processes, i.e. HABFLO and HABFUZZ. The aim of the investigation was to compare the performance of the HABFUZZ model to the HABFLO model that is widely used in South Africa. The HABFUZZ model uses fuzzy inference processes and Bayesian joint probability inference methods to calculate the instream habitat suitability based on flow velocity, water depth, substrate type and temperature of a hydraulically simulated river reach (Theodoropoulos et al. 2016). HABFUZZ introduces the power of fuzzy logic to model habitat preferences across a broader spatial scale, and fuzzy logic is useful to account for the inherent uncertainties and variations found in ecological systems. The conventional models such as HABFLO rely on definitive boundaries (e.g., a fish species only exists above a specific depth), however fuzzy logic uses membership functions to represent degrees of suitability. This is done in HABFUZZ by assigning fuzzy membership values to different habitat variables. These membership functions are often derived from expert knowledge or field observations. HABFUZZ then combines these fuzzy values across multiple variables, creating a composite measure of habitat suitability for each point within the modelled section of a river. In addition to the use of fuzzy logic, HABFUZZ also accounts for the temperature of water in the reach being modelled. TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 20 Temperature is one of the drivers of aquatic ecosystems, because it influences fish metabolism, growth rates, spawning, and the overall distribution of species. Changes in flow can significantly alter water temperature regimes, impacting habitat suitability. Therefore, it may be important to include water temperature when performing habitat suitability modelling. For this study, temperature data were collected using temperature sensors mounted to a boat that was towed using a UAV. For this comparison, the HABFUZZ model was trained using five combinations of flow velocity and water depth and eight classes of substrates. It was assumed that the habitat suitability was best for the lowest flow velocities (0.05 to 0.1 ms-1) and water depths (0.1 to 0.15 m) and water temperatures of 15 ○C. Five different habitat suitability classes (from bad to high) were defined for the habitat suitability classes as shown in Table 14. These rules can be changed for any specific organism being investigated. Table 14. HABFUZZ habitat suitability definitions For the cross section where data were collected, the K values from HABFUZZ were between 0 and 0.2 as shown in Table 15 indicating that the habitat suitability would not be ideal for an organism that prefers slow moving shallow waters. These results correlate with the HABFLO results where majority of the fish and invertebrate habitat distributions were not in the slow/shallow segment of the total distribution. The habitat suitability values from HABFUZZ can be used to produce a continuous map of habitat suitability values across a river reach. This spatial representation of habitat suitability can potentially reveal more information on how varying flow scenarios might alter the suitability of a river for different habitat classes. This spatial perspective can also help to identify areas most sensitive to flow alterations and subsequently guide the design of flow releases to optimise habitat availability for target species throughout their life stages. In the next section, the assumptions, limitations, and recommendations based on the study are provided. K class meaning Values High > 0.8-1 Good > 0.6-0.8 Moderate > 0.4-0.6 Poor > 0.2-0.4 Bad 0-0.2 ASSUMPTIONS, LIMITATIONS AND RECOMMENDATIONS In this section of the report, a list of assumptions, limitations and recommendations are provided to guide further research in this field of study. • The results presented in this report are only for the segment of the Olifants river where data were available. However, irrespective of the area modelled, it is envisaged that the methodology can be replicated for other river reaches provided that adequate data are available. Therefore, if more data are collected and processed more efficiently (for example, with a green LiDAR sensor) at any other site, the methodology can be easily replicated using the steps documented in this report. • The depth and velocity classes from HABFLO were used in this report but it is acknowledged that different depth and velocity classes or suitability curves could be used for habitat modelling. Tape (m) Velocity Depth Substrate Temper- ature Habitat suitability 2 0.00 0.00 0.02 20.50 0 2.4 0.20 0.23 0.02 19.61 0.01 3.2 0.40 0.27 0.02 19.29 0.01 4 0.40 0.24 0.02 19.09 0.01 4.7 0.40 0.31 0.03 19.74 0.01 5.7 0.50 0.32 0.03 20.65 0.01 6.7 0.70 0.31 0.03 19.14 0.01 6.9 0.80 0.06 0.03 19.72 0.01 7.1 0.50 0.36 0.03 19.48 0.01 7.7 0.50 0.37 0.07 20.36 0.01 8.4 0.50 0.41 0.07 19.37 0.01 9.2 0.60 0.38 0.07 20.31 0.01 10.2 0.50 0.37 0.07 20.91 0.01 10.9 0.80 0.34 0.07 19.77 0.01 11.7 0.70 0.37 0.07 19.46 0.01 12.5 0.70 0.32 0.07 19.95 0.01 12.8 0.90 0.30 0.07 20.09 0.01 13.1 0.90 0.05 0.07 20.10 0.01 13.3 0.40 0.41 0.03 19.55 0.01 14 0.50 0.33 0.03 19.08 0.01 14.7 0.80 0.35 0.03 20.69 0.01 15.5 0.80 0.36 0.03 20.22 0.01 15.8 0.70 0.30 0.03 19.90 0.01 15.9 0.30 0.00 0.02 19.04 0 Table 15. HABFUZZ inputs and outputs for the cross section TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 21 • The DTM used for the hydraulic modelling was the most accurate representation of reality that could be derived using the data available at the time of the study. It is acknowledged that the accuracy of the outputs from the hydraulic model will vary based on the level of accuracy of the DTM. Therefore, the level of accuracy of the habitat distributions is dependent on the accuracy of the DTM and the depth and velocity outputs generated from the hydraulic model. • A limitation of this study is that the DTM used for the HEC-RAS model was developed using LiDAR data and SoNAR data which were stitched together. The application of the SoNAR device did not cover the entire surface area of the riverbed due to time constraints and did not collect data at a range of angles other than vertically downwards. Therefore, detail around sandbars, crevices, rocky ledges, and accumulated sediment plumes were not collected and thus detailed bathymetric data was not collected to develop an accurate hydraulic model. Should this process be used in future, it is believed that with more precise application the developed methodology will provide accurate results and can be used to develop an accurate digital twin of a river and floodplain. • An additional limitation is that there are uncertainties when using the distribution model to determine the distributions of HABFLO classes out of the range of discharges used to develop the distribution curve. For this report, a limited number of Q values were used for the reach. The performance of the model can therefore be improved if additional Q values are used based on the requirements of the user and the study. • Assumptions were made regarding the type of sediments found across the reach because no detailed sediment mapping was conducted during the site visit. However, the methodology to estimate the distribution of invertebrate habitat classes show that the method can be beneficial if detailed on-site sediment data are collected and mapped for a river reach. CONCLUSION AND RECOMMENDATIONS The results presented in this report highlight the successful integration of SoNAR and LiDAR datasets to generate a high- resolution DTM of a length of river. The successful integration of SoNAR and LiDAR data demonstrates the value of combining complementary UAV derived remote sensing techniques for a comprehensive characterisation of riverine habitats. This approach offers greater detail and spatial coverage than either method (SoNAR or LiDAR) alone. The generation of a DTM with an underwater river profile should theoretically enable more nuanced hydraulic modelling in HEC-RAS. However, as highlighted previously, the accuracy from HEC-RAS is dependent on the accuracy of the DTM, which relates to the accuracy of the survey application. Nevertheless, the derived depth and velocity outputs from the HEC-RAS model provided the basis for a longitudinal visualisation of the HABFLO fish and invertebrate habitat classes. In this study, the analysis of habitat distributions was extended to an entire river reach. This addresses a common limitation of traditional habitat assessments that are done for a single cross section. The results obtained revealed variations in habitat class distribution that would be undetectable using a single cross-section approach. The longitudinal perspective/ visualisation can prove to be crucial for understanding habitat availability and connectivity within the river system. The spatially explicit representation of habitat classes as a function of flow enables a more realistic assessment of the potential of the river to support diverse species and life stages for different flows and it can highlight areas of particular importance for conservation or restoration efforts. It is recommended that field sampling of fish and invertebrate populations be performed to validate the modelled habitat class distribution, thereby strengthening the results obtained in the study. In addition, the incorporation of a range of flow scenarios can be used to explain how habitat class availability changes throughout the year which could potentially aid in the development of seasonally adjusted E-flow recommendations. It is further recommended that the methodology be applied to larger river reaches or entire catchments. Overall, it is concluded that the use of a 2D hydraulic model is effective in modelling habitat classes for E-Flows for a specific reach of river compared to a single cross-section. When comparing the results for the entire reach to a 1D model, it was evident that the habitat distribution at a single cross-section is not representative of an entire reach of river. However, the feasibility of such a model is dependent on the feasibility of the data collection to develop a 2D model. This may not yet be feasible in terms of collecting high-resolution bathymetric data required for a long reach of river for a small-scale project. However, if a study were to focus on smaller segments of a reach, it would be feasible to collect sufficient data in terms of time and cost. As represented in Table 3 of this report, it can be concluded that the traditional 1D cross-section approach would be the least costly in terms of equipment, however, this approach is human and time-based resource intensive depending on the level of detail required. As shown in the results, the information obtained from a 1D cross-section will also not provide a detailed representation of a river reach with more complex flow characteristics. The integrated approach of using a SoNAR device, or similar, with a UAV (as done for this study) will require larger equipment costs, yet fewer resources. The time required to apply the survey accurately may be more than that to undertake a single cross-section, however, the level of detail obtained over the study area is greater and may be TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 22 used in a 2D-hydraulic model for a more accurate representation of the river reach. Furthermore, this report presents a novel workflow for linking advanced remote sensing techniques with 2D hydraulic and habitat modelling. The upscaling from a single cross-section to a reach-level analysis provides valuable insights into the dynamic nature of riverine habitats and their response to varying flow conditions. The resulting visualisations can potentially be used to effectively communicate the ecological implications of E-flow recommendations to both scientists and stakeholders. These spatial representations can be incorporated into a digital twin which can lead to more informed decision making. The results obtained from the sensor platform demonstrates the potential of low-cost, adaptable sensor platforms for water resource monitoring. Further refinement could include real- time data transmission and an expanded suite of sensors with further testing. The scalability and ease-of-use of this approach shows promise for citizen science initiatives which can empower communities with data to advocate for the health of their local rivers. Furthermore, the applicability and performance of a green LiDAR technology is very promising in practice being able to perform well in multiple environments, perform repeatable measurements and to conduct surveys in difficult to access river reaches. The system can be viewed as a cost-effective solution over large land and coastal zones making it an attractive tool for the creation of a digital twin. HABFLO is advantageous over HABFUZZ because the model is simple to use and there is detailed literature available to guide users of the model, but the model does not have the capability to determine habitat suitability for an entire river reach. However, HABFUZZ, can be used to determine habitat suitability for an entire river reach and can be setup using data collected with a UAV as shown in this report. Therefore, based on the results obtained in this study, it is recommended that both HABFUZZ and HABFLO be used in conjunction with each other if training and temperature data are available for HABFUZZ. REFERENCES Arthington A. 1998. Comparative evaluation of environmental flow assessment techniques: Review of holistic methodologies. LWRRDC Occasional Paper 26/98. Land and Water Resources Research and Development Corporation (LWRRDC): Canberra. Arthington AH, Bhaduri A, Bunn SE, Jackson SE, Tharme RE, Tickner D, Young B, Acreman M, Baker N and Capon S. 2018. The Brisbane declaration and global action agenda on environmental flows (2018). Frontiers in Environmental Science 6 45. Bovee KD and Milhous RT. 1978. Hydraulic simulation in instream flow studies: theory and techniques. Department of the Interior, Fish and Wildlife Service. Cavendish MG and Duncan MI. 1986. Use of the instream flow incremental methodology: a tool for negotiation. Environmental Impact Assessment Review 6 (4): 347-363. Changsen Z, Shengtian Y, Junguo L, Changming L, Fanghua H, Zhonggen W, Huitong Z, Jinxi S, Simon MM, Richard PL. 2018. Linking fish tolerance to water quality criteria for the assessment of environmental flows: A practical method for streamflow regulation and pollution control, Water Research: 141: 96-108. Chow VT. 1959. Open channel hydraulics. McGraw-Hill, Illinois, USA. DJI. 2024. A LiDAR + RGB Solution for Aerial Surveying. [Internet]. Available from: https://enterprise.dji.com/ zenmuse-l1. [Accessed: 13 March 2024]. Dunbar MJ, Gustard A, Acreman M and Elliott CR. 1997. Environment Agency Project W6B (96) 4 Overseas approaches to setting river flow objectives Draft Interim Technical Report. DWS. 2023. Hydrological Services - Surface Water (Data, Dams, Floods and Flows). [Internet]. Department of Water and Sanitation (DWS), RSA. Available from: http://www. dwa.gov.za/Hydrology/default.aspx. [Accessed: September 2023]. Gan K and McMahon T. 1990. Variability of results from the use of PHABSIM in estimating habitat area. Regulated Rivers: Research & Management 5 (3): 233-239. Gard M. 2009. Comparison of spawning habitat predictions of PHABSIM and River2D models. International Journal of River Basin Management 7 (1): 55-71. HEC. 2024. Hydrologic Engineering Center's (CEIWR-HEC) River Analysis System (HEC-RAS). [Internet]. US Army Corps of Engineers Hydrologic Engineering Center. Available from: https://www.hec.usace.army.mil/software/ hec-ras/documentation.aspx. [Accessed: January 2024]. Hirschowitz P, Birkhead AL and James CS. 2007. Hydraulic modelling for ecological studies for South African rivers. Jonsson B and Jonsson N. 2009. A review of the likely effects of climate change on anadromous Atlantic salmon Salmo salar and brown trout Salmo trutta, with particular reference to water temperature and flow. Journal of Fish Biology: 75, 2381–2447 King J, Tharme R and MS dV. 2008. Environmental Flow Assessments for Rivers: Manual For The Building Block Methodology. WRC Report No TT 354/08. Water Research Commission, Pretoria, South Africa. TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital 23 Lal M, Mishra S, Pandey A, Pandey R, Meena P, Chaudhary A, Jha RK, Shreevastava AK and Kumar Y. 2016. Evaluation of the Soil Conservation Service Curve Number Methodology using Data from Agricultural Plots. Hydrogeology Journal 25 (1): 151-167. Laquinta. 2024. Multispectral UAV and ground vehicle camera. [Internet]. Available from: https://arborcarbon.com. au/arbor/laquinta/index.html. [Accessed: 13 March 2024]. Legates DR and McGabe GJ. 1999. Evaluating the use of “Goodness-Of-Fit” Measures in Hydrologic and Hydroclimatic Model Validation. Water Resources Research 35 (1): 233-241. Loar JM, Sale MJ and Cada GF. 1986. Instream flow needs to protect fishery resources. Oak Ridge National Lab., TN (USA). Environmental Sciences Div. Mathur D, Bason WH, Purdy Jr EJ and Silver CA. 1985. A critique of the in stream flow incremental methodology. Canadian journal of fisheries and aquatic sciences 42 (4): 825-831. Milhous RT, Updike MA and Schneider DM. 1989. Physical habitat simulation system reference manual: version II. US Fish and Wildlife Service. Nash JE and Sutcliffe JV. 1970. River Flow Forecasting Through Conceptual Models Part I - A Discussion of Principles. Journal of Hydrology 10 (3): 282-290. O'Brien GC, Dickens C, Hines E, Wepener V, Stassen R, Quayle L, Fouchy K, MacKenzie J, Graham PM and Landis WG. 2018. A regional-scale ecological risk framework for environmental flow evaluations. Hydrology and Earth System Sciences 22 (2): 957-975. Poff NL, Allan JD, Bain MB, Karr JR, Prestegaard KL, Richter BD, Sparks RE and Stromberg JC. 1997. The natural flow regime. BioScience 47 (11): 769-784. Poff NL and Matthews JH. 2013. Environmental flows in the Anthropocence: past progress and future prospects. Current Opinion in Environmental Sustainability 5 (6): 667-675. QGIS. 2024. Open Source Geospatial Foundation Project. [Internet]. QGIS Geographic Information System. . Available from: http://qgis.org. [Accessed: January 2024]. Quadros ND, Collier PA and Fraser CS. 2008. Integration of Bathymetric and topographic LiDAR: A preliminary investigation. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII.B8: 1299–1304. Rai N and Jain A. 2022. Environmental Flow Assessment for River Valley Projects. Water and Energy International 65 (1): 26-32. Richardson BA. 1986. Evaluation of instream flow methodologies for freshwater fish in New South Wales. Stream protection, the management of rivers for instream uses. Water Studies Centre, Chisholm Institute of Technology, Caulfield 143-167. Ritter A and Munoz-Carpena R. 2013. Performance Evaluation of Hydrological Models: Statistical Significance for Reducing Subjectivity in Goodness-Of-Fit Assessments. Journal of Hydrology 480: 33-45. Robotics B. 2024. Ping2 Sonar Altimeter and Echosounder. [Internet]. Available from: https://bluerobotics.com/store/ sonars/echosounders/ping-sonar-r2-rp/. [Accessed: 13 march 2024]. Roh KB, Park SC, Jin YH and Park MO. 2011. Study on ecological instream flow estimation using River2D model in the Seomjin River. Journal of Korean Society on Water Environment 27 (6): 822-829. Schulze RE, Lynch SD, Smithers JC, Pike A and Schmidt EJ. 1995. Statistical Output from ACRU. In: ed. Schulze, RE, Hydrology and Agrohydrology: A Text to Accompany the ACRU 3.00 Agrohydrological Modelling System, Chapter 21, AT21-1 - AT21-35. Water Research Commission, Pretoria, RSA. Shirvell C. 1986. Pitfalls of physical habitat simulation in the instream flow incremental methodology. Department of Fisheries and Oceans, Fisheries Research Branch. Singh K. 2023. Current Trends in River Bathymetry using UAV-borne Technology to Inform E-flow Assessments in Southern Africa. Stalnaker, C, Lamb, B, Henriksen, J, Bovee, K and Bartholow, J. 1995. The instream flow incremental methodology: a primer for IFIM. Biological Report 29. US Department of the Interior, National Biological Service. Steffler P and Blackburn J. 2002. Two-Dimensional Depth Averaged Model of River Hydrodynamics and Fish Habitat. University of Alberta, USA. Tharme RE. 1996. Review of international methodologies for the quantification of the instream flow requirements of rivers. Water Law Review: final report for policy development, South African Department of Water Affairs and Forestry. Freshwater Research Unit: University of Cape Town: Pretoria. Theodoropoulos C, Skoulikidis N and Stamou A. 2016. Habfuzz: A tool to calculate the instream hydraulic habitat suitability using fuzzy logic and fuzzy Bayesian inference. Journal of Open Source Software 1 (6): 82. Tsang Y, Infante DM, Wang L, Krueger D & Wieferich D. 2021. Conserving stream fishes with changing climate: Assessing fish responses to changes in habitat over a large region. Science of The Total Environment: 755. TECHNICAL REPORT CGIAR Initiative on Digital Innovation | on.cgiar.org/digital This publication has been prepared as an output of the CGIAR Initiative on Digital Innovation, which researches pathways to accelerate the transformation towards sustainable and inclusive agrifood systems by generating research-based evidence and innovative digital solutions. This publication has not been independently peer reviewed. Responsibility for editing, proofreading, and layout, opinions expressed, and any possible errors lies with the authors and not the institutions involved. The boundaries and names shown and the designations used on maps do not imply official endorsement or acceptance by the International Water Management Institute (IWMI), CGIAR, our partner institutions, or donors. In line with principles defined in the CGIAR Open and FAIR Data Assets Policy, this publication is available under a CC BY 4.0 license. © The copyright of this publication is held by IWMI. We thank all funders who supported this research through their contributions to the CGIAR Trust Fund. 24 Waters BF. 1976. A methodology for evaluating the effects of different streamflows on salmonid habitat. Proceedings of the symposium and specialty conference on instream flow needs, 254. Zhao C, Yang S, Liu J, Liu C, Hao F, Wang Z, Zhang H, Song J, Mitrovic SM and Lim RP. 2018. Linking fish tolerance to water quality criteria for the assessment of environmental flows: A practical method for streamflow regulation and pollution control. Water Research 141: 96-108.