Forecasting Cyanobacteria Cell Counts in Lakes Based on Hyperspectral Sensing

Highlights

What are the main findings?

• Hyperspectral reflectance from the HydraSpectra sensor strongly correlates with cyanobacteria cell counts and chlorophyll-a under bloom conditions, and integration with a hydrodynamic-algal growth model enables reliable short-term bloom forecasts in Australian inland lakes.
• Daily variations in cyanobacterial surface concentrations are primarily driven by vertical mixing dynamics rather than temperature changes alone, highlighting the need to consider hydrodynamic- and depth-dependent distributions in bloom assessments.

What is the implication of the main finding?

• Combining continuous hyperspectral monitoring with hydrodynamic and growth models can provide early warnings of cyanobacterial blooms, supporting proactive water management and mitigation strategies.
• Forecast accuracy depends on high-resolution environmental inputs, such as real-time temperature profiles and mixing data, emphasizing that satellite- or surface-only observations may be insufficient without considering water column dynamics.

Abstract

We developed a forecast model for cyanobacteria bloom formation in two contrasting inland lakes in Australia by combining in situ sampling and continuous remote sensing hyperspectral reflectance (HydraSpectra) with hydrodynamic and algal growth models. Cyanobacterial distribution of a buoyant species was simulated with an algal growth model, driven by forecasted meteorological data, and modeled temperature stratification and mixing dynamics from a one-dimensional, vertical k-epsilon turbulence hydrodynamic model. The cyanobacteria model was re-initialized daily with measured cell counts derived from the hyperspectral reflectance data. Simulations of cyanobacterial concentrations (cell counts) reflected the dynamic mixing behavior in the lakes with daily phases of near-surface accumulation and subsequent daily mixing due to wind or night-time cooling. To determine the surface concentration of cyanobacteria on sub-daily time scales, it was demonstrated that the combined use of high-resolution water temperature profiles, HydraSpectra reflectance data, and a hydrodynamic model to quantify the mixing dynamics is essential. Overall, the model results demonstrated a prototype for a cyanobacteria short-term forecast model. Having these tools in place allows us to quantify the risks of cyanobacterial blooms in advance to inform options for lake management.

1. Introduction

The proliferation of cyanobacterial blooms is a major concern in freshwater management because they produce toxic secondary metabolites, collectively known as cyanotoxins [1,2,3,4]. These toxins threaten water safety for humans, aquatic organisms, and livestock, and their treatment is often costly [5,6]. Blooms also compromise aquatic ecosystems, human health, and the usability of limited freshwater resources, leading to significant economic impacts [7,8]. In addition, they obstruct sunlight penetration and, during decomposition, deplete the oxygen essential for aquatic organisms [9]. Large blooms of blue–green algae were linked to tens of thousands of fish deaths along a 30 km stretch of the Darling-Baaka River, Australia, in December 2018, followed by hundreds of thousands in January 2019 [10]. An even more severe event occurred in 2023, with an estimated 20–30 million fish deaths near Menindee, NSW, profoundly affecting both the environment and the local community [11].

In Australia, the Australian National Health and Medical Research Council [12,13] issued guidelines for managing cyanobacterial risks in drinking and recreational water. For recreation, thresholds include “≥10 $\mathsf{\mu }$g/L total microcystins; ≥50,000 cells/mL toxic Microcystis aeruginosa; biovolume ≥4 mm³/L if toxin producers dominate; ≥10 mm³/L for all cyanobacteria where toxins are absent; or persistent scums.” For drinking water, notification may occur at “≥1.3 $\mathsf{\mu }$g/L Microcystis aeruginosa (2000 cells/mL; biovolume 0.2 mm³/L).” While biovolume, integrating abundance and community size structure, is the legislated indicator of cyanobacterial biomass, its measurement is not without limitations. Traditional in situ sampling and microscopy are labor intensive, time-consuming, and limited in temporal and spatial coverage, often failing to capture bloom dynamics [14,15].

Remote sensing offers a promising alternative for monitoring blooms and optical water quality [16,17,18]. Technologies range from satellites [9,15,19,20,21,22,23] to UAVs [24,25,26,27], though they are limited by cloud cover, observation gaps, and reduced sensitivity to vertical bloom structure. Detection is based on the spectral influence of water quality parameters (400–900 nm) [28,29]. Hyperspectral imaging enhances cyanobacteria monitoring by capturing continuous spectra, enabling precise detection, classification, and characterization [30,31]. Distinct pigment compositions produce unique absorption features, and hyperspectral sensors provide the spectral detail needed for species-level classification, unlike multispectral systems with limited bands [32,33].

Above-surface instruments measuring spectral reflectance provide high-temporal-resolution water quality monitoring with minimal atmospheric interference, addressing satellite limitations such as revisit gaps and cloud cover. Traditional spectral devices require multiple sensors for full reflectance coverage, limiting use due to cost [34]. Hyperspectral sensors primarily capture surface reflectance, which may not fully represent the vertical distribution of cyanobacteria, particularly in polymictic lakes with frequent mixing. Despite this, surface measurements provide high-frequency data that can be integrated with modeling to estimate overall bloom dynamics, while future efforts incorporating depth-resolved profiles could further enhance representation of three-dimensional cyanobacterial distribution. HydraSpectra [35] is a low-cost, robust spectrometer that measures hyperspectral reflectance across the visible spectrum using off-the-shelf and custom components. It simultaneously records solar, skylight, and water-leaving irradiance from various viewing geometries with the same patented spectrometer, supported by integrated cameras for qualitative context. Deployed on fixed or floating platforms, HydraSpectra collects reflectance every 15 min and uploads wirelessly, complementing sparse satellite and in situ data with high-frequency observations. Despite its low cost, intercomparisons in Queensland, Australia, and Venice, Italy, confirmed its measurements are comparable to high-end commercial spectroradiometers [36]. In this study, we use HydraSpectra-derived hyperspectral reflectance data (visible to near-infrared) to characterize cyanobacteria, hereafter referred to as “hyperspectral sensing”.

Hydrodynamic modeling is widely applied to simulate vertical stratification and mixing in lakes. The LAKEoneD model [37,38], coupled with a cyanobacteria growth model [39], uses meteorological drivers to simulate water temperature and cyanobacteria abundance at high temporal and vertical resolution. Here, LAKEoneD was tested as a forecasting tool. Initial algal distribution strongly influences forecast accuracy; we demonstrate that algal concentrations retrieved from HydraSpectra provide reliable initial conditions. Details of the retrieval method are presented in Section 2.3.4 and algorithms in Section 2.4.4. The cyanobacteria model was Neumannalized daily with measured cell counts derived from the hyperspectral reflectance data. Combining remote sensing and hydrodynamic modeling has been shown to markedly improve prediction skills in terms of historical simulations and forecasting of cyanobacteria [40].

Given HydraSpectra’s high-frequency observations, direct integration with hydrodynamic modeling could support continuous cyanobacteria forecasts, maintained by frequent re-initialization. Similar to weather forecasts, such outputs would provide empirical evidence of bloom likelihood, enabling early warnings and potential operational applications for water managers.

This study develops a short-term cyanobacteria forecast model using hyperspectral data to set initial conditions. We test its capability in two inland lakes, with results supporting early warning systems based on the World Health Organization’s 2021 algal alert framework [41], guiding management and policy to mitigate harmful blooms.

2. Materials and Methods

2.1. Study Sites

In this study, HydraSpectra units were deployed at two freshwater lakes on Australia’s east coast, namely Lake Hume (LH) and Grahamstown Dam (GTD). While LH is a deep, stratified monomictic hydropower and irrigation reservoir, GTD is a polymictic drinking water supply reservoir. Differences in limnological conditions and operational purposes make the two lakes good candidates for an intercomparison of the success of the approach for inland waters in general. Figure 1 shows pontoons equipped with the HydraSpectra units deployed at the two lakes.

Lake Hume is the largest storage in Australia’s Murray River system, primarily providing water for large-scale irrigated agriculture downstream during the austral spring and summer. The lake plays a critical role in capturing winter and spring rainfall from the Australian Alps and releasing it to regulate the flow of the River Murray, and is also used for flood mitigation and hydroelectricity. Lake Hume was a seeding source of cyanobacteria blooms downstream in the Murray River in 2003, 2005, 2007, 2009, and 2010 [43,44,45,46], and most likely in 2016 [47]. These blooms are often conspicuous due to their occurrence at the surface layer. However, subsurface blooms, which can harbor dense cyanobacterial populations, are generally less visible and often overlooked despite their potential toxicity and frequent occurrence in sensitive water bodies, such as irrigation reservoirs like Lake Hume [48].

Grahamstown Dam (GTD) is the New South Wales Hunter region’s largest drinking water supply dam, classified as an off-river storage facility. From this source, Hunter Water provides water services to a population of nearly 600,000 people in the lower Hunter region on the east coast of Australia. Inputs to GTD are pumped from the Williams River, a perennial tributary of the Hunter River, at Seaham Weir through the Balickera Canal; it collects more water from the surrounding catchment on its way to the pump station from where water is lifted to another canal that feeds the GTD [49]. The dam and the Williams River often experience potentially toxic blue–green algal blooms [50].

2.2. Sampling Techniques

2.2.1. Grab Sampling

In Lake Hume, water samples were collected from approximately 20 cm beneath the surface using opaque plastic sample bottles. These bottles were kept cool and out of the direct sun until further processed in the laboratory several hours later. In the laboratory, the samples were well mixed, and measured aliquots of water were filtered through GF/F filters for the determination of chlorophyll-a (Chl-a), accessory pigments, and phycocyanin by spectroscopic analysis [51,52]. More on the collected data can be found in [53]. Cell counts and derived biovolume were determined from samples regularly obtained by WaterNSW at different locations in the lake. Counts were performed in triplicate using a flow cytometer with cross-validation to microscopic counts performed on two sample days.

Grab sampling results data for cell counts and biovolume of total cyanophytes, Dolichospermum and Microcystis species in Grahamstown Dam were provided by Hunter Water Corporation from their routine water quality monitoring program. Over 340 grab samples were obtained for regular locations in the north (R2), south (R6), and middle of the dam (R12), along with an additional 32 samples at the HydraSpectra location. All samples were preserved in cool boxes and transported to the ALS laboratory in Newcastle for analysis.

2.2.2. Monitoring Equipment

Surface reflectance measurements obtained by the HydraSpectra instruments follow the standard viewing geometry proposed by Mueller et al. [54] to minimize the influence of surface sunlight and non-uniform sky radiance. Designed for fixed and continuous measurement, the sensor deploys angled sky and water sensors to overcome the need for a solar tracker. The instrument employs the conventional methodology to correct for surface reflectance effects [55] and incorporates additional cameras to assess ambient conditions during measurement (Figure 2).

HydraSpectra units were deployed following manufacturer guidelines [35]. In the southern hemisphere, the sensor head was oriented toward 135°–225° azimuth to reduce specular reflection and solar forward scattering, and positioned to maintain an unobstructed 24° field of view across all apertures. Mounting on fixed structures ensured stable geometry, with care taken to avoid shading, obstructions, or bottom reflection effects. These procedures minimize diurnal variability associated with solar angle and view geometry. While full atmospheric correction and view-angle normalization were beyond the scope of this study, the consistent deployment protocol ensured data quality for the comparative purposes of this work.

Once deployed, the instrument operates with minimal maintenance requirements, supports wireless configuration, and delivers high-frequency remote sensing data at customizable intervals. It captures spectral information within the 400–900 nm range at a high spectral resolution of 1 nm. Unlike many satellite-based sensors, such as Sentinel-2 or Landsat, which are constrained by their limited spectral band resolution and are not specifically designed for detailed pigment detection, this instrument offers significantly enhanced capabilities. HydraSpectra’s high spectral resolution enables the detection and differentiation of specific reflectance signatures from various water constituents, including chlorophyll-a, cyanobacterial pigments, colored dissolved organic matter (CDOM), etc. (Figure 2 (right)).

Water temperature was collected from thermistor chains deployed in each lake. The chains hung vertically in the water column and recorded temperature at 0.5 m depth intervals, generating temperature profiles with an accuracy of up to 0.1 °C at 15 min intervals. The thermistor chains provided high-frequency vertical resolution at representative locations within the lakes, which were selected as critical points for monitoring algal bloom dynamics. Although only single-station chains were available, these measurements captured the key vertical temperature structure required for model forcing.

The internal heat distribution is determined by light absorption, in its simplest form, described via a single absorption coefficient determined from a measured Secchi disk depth.

2.3. Data

2.3.1. Meteorology

The LAKEoneD hydrodynamic models were driven by meteorological time series for solar radiation (measured in W/m²), air temperature (°C), and relative humidity (%) at a height of 2 m, wind speed (m/s) and wind direction (degrees north) at a height of 10 m, and cloudiness (on a scale from 0 to 1).

The current model simulation used historical and forecast meteorological data from the Meteoblue service [57], as it is gap free and covers a large period. An example of meteorological data is shown in Figure 3. For Australia the data are generated on a grid with a spatial resolution of 30 km. Each grid cell represents the average meteorological conditions for that area. The datasets include all the meteorological fields required by the model. These fields are provided at hourly time steps; their gap-free nature ensures continuous and consistent simulation input.

Since the data undergo quality checks by the provider, no additional quality assurance or gap-filling efforts were required for this project. This ensured that the datasets could be directly used without the need for preprocessing or adjustments. The data for Lake Hume covers the entire simulation period 2014–2020, and that for Grahamstown Dam covers 2018–2024.

2.3.2. Bathymetry

To derive digital bathymetric layers for both impoundments, RGB values of pixels were extracted from topographic maps using the Python image library Open CV 4.11.0. Pixel counts for each contour interval were then interpolated to estimate the corresponding areas. The basic geometric parameters—including length, width, surface area, volume—formed the hypsometric curves for the LAKEoneD model. Lake bathymetry was added to the program via a file describing depth versus area on a regular vertical grid.

2.3.3. In Situ Water Quality

• Temperature

At LH, the monitoring station was anchored near the Hume Dam Wall (36°06′27.8″S 147°01′53.4″E). The thermistor was spread from the surface to a depth of about 30 m, consisting of 24 temperature sensors, deployed in February 2017. It remained operational until 7th January 2021. At GTD, the thermistor chains were located at R2 in the north, R12 in the middle, and R6 in the south (Figure 1).

• Algal cell counts

An example of total cyanophyte algal cell counts including Microcystis and Dolichospermum for Grahamstown Dam can be shown in Figure 4.

2.3.4. Hyperspectral Data

The HydraSpectra units were deployed in Lake Hume on 23rd February 2022 at 36°05′27″S 147°03′37″E (north of Bellbridge Bridge, second pier from the eastern side), and at Grahamstown Dam on 4 August 2022 at 32°45′58″S and 151°47′40.1″E, above the water surface (800 mm). With a 24° field of view, the sensor samples from an area of the lake approximately 0.34 m in diameter (0.1 m²).

With acquisitions set for every 15 min during daylight hours, the recorded reflectance spectra were subsequently processed to quantify local chlorophyll-a concentrations and cell counts (HS cellC). This transformation is achieved through band ratio algorithms, which will be discussed in Section 2.4.4. The forecasting model is constructed using these derived indices rather than the full spectral dataset.

2.4. Model Construction

2.4.1. Hydrodynamic Model

The hydrodynamic model employed here, LAKEoneD hydrodynamics, integrates a k-epsilon turbulence model and solves the balance equation for heat energy and momentum in a bounded system, along with a balance equation for turbulent kinetic energy and turbulent dissipation rate [38,58]. Meteorological forces acting on the surface drive the model. This model aligns with similar models based on the same physical equations with slight variations in describing meteorological forcing, turbulence closure, and implementation [59,60]. Extensive studies across different lake settings have compared the outcomes of these turbulence models with various model types, consistently demonstrating comparable competence in resolving temperature stratification over time in most cases [61,62,63]. The vertical turbulence structure can be derived from detailed profiles of the temperature microstructure [64,65]. Horizontal momentum related to the horizontal–vertical velocity vector, $\mathbf{u}$, is driven by wind stress at the surface and vertically distributed diffusion. The change in horizontal momentum can be described by a system of two partial differential equations in the x and y directions, respectively,

$${\displaystyle \frac{\partial \mathbf{u}}{\partial t}}={\displaystyle \frac{\partial }{\partial z}}\left((D_m+D_z){\displaystyle \frac{\partial \mathbf{u}}{\partial z}}\right)+c_b\mathbf{u}\left|\mathbf{u}\right|{\displaystyle \frac{1}{A}}{\displaystyle \frac{\partial A}{\partial z}},$$

(1)

where z runs from 0 at the surface to a maximum depth at the bottom of the lake. The lake’s bathymetry is implicitly considered via the area–depth relation, $A\left(z\right)$. The term $D_m$ is the molecular diffusivity of momentum for water, and $D_z$ is the vertical turbulent diffusivity, which is space and time dependent. The second term on the right-hand side describes the loss of momentum at each depth induced by the lake’s boundaries. This boundary stress term is formulated as a sliding law with a drag coefficient, $c_b$, chosen according to the lake’s settings. The two boundary conditions for the momentum equation are the continuity of shear stress at the surface generated by wind stress, with a velocity-dependent drag coefficient, and vanishing shear stress at the lake’s deepest point [66].

The dynamic changes in temperature, $T(z,t)$, can be described as,

$${\displaystyle \frac{\partial T}{\partial t}}={\displaystyle \frac{1}{A}}{\displaystyle \frac{\partial }{\partial z}}\left(A(D_h+{\displaystyle \frac{D_z}{{\sigma }_h}}){\displaystyle \frac{\partial T}{\partial z}}\right)-{\displaystyle \frac{1}{\rho \left(T\right)c\left(T\right)}}{\displaystyle \frac{\partial E}{\partial z}},$$

(2)

where heat is produced by the absorption of short-wave radiation, $E(z,t)$, and vertically distributed by diffusion. In the right-hand side first term, $D_h$ is the molecular of heat in the water, and $D_z/{\sigma }_h$ is the turbulent diffusivity of heat expressed as the turbulent diffusivity of momentum divided by the Prandtl number for heat, ${\sigma }_h$. We neglect heat exchange between the water column and sediments. The second term on the right-hand side describes the absorption of short-wave radiation, where $\rho \left(T\right)$ is the density of water and $c\left(T\right)$ is the specific heat of water, with both being functions of temperature [67]. Short-wave radiation in the water column depends on the incident radiation at the water surface and decreases exponentially with depth, according to Lambert–Beer’s law.

The dynamical distributions of turbulent kinetic energy, k, and turbulent dissipation rate, $ϵ$, are described by the two equations [38,58,68,69],

$${\displaystyle \frac{\partial k}{\partial t}}={\displaystyle \frac{\partial }{\partial z}}\left((D_m+{\displaystyle \frac{D_z}{{\sigma }_k}}){\displaystyle \frac{\partial k}{\partial z}}\right)+P+G-ϵ,$$

(3)

$${\displaystyle \frac{\partial ϵ}{\partial t}}={\displaystyle \frac{\partial }{\partial z}}\left((D_m+{\displaystyle \frac{D_z}{{\sigma }_{ϵ}}}){\displaystyle \frac{\partial ϵ}{\partial z}}\right)+(c_{1}P+c_{3}G-c_2ϵ){\displaystyle \frac{ϵ}{k}}.$$

(4)

where P and G describe the production and loss of turbulent kinetic energy induced by shearing of the flow and buoyancy, respectively. Here, ${\sigma }_k$ and ${\sigma }_{ϵ}$ are the Prandtl numbers for turbulent kinetic energy and turbulent dissipation rate, both assumed to be constant, and $c_1$, $c_2$, and $c_3$ are the coefficients of turbulent dissipation. Their values can be found in [58].

Hourly meteorological data for shortwave radiation, air temperature, relative humidity, wind speed and direction, and cloudiness drive LAKEoneD hydrodynamics. Based on a prescribed bathymetry (hypsometric curve) of a lake, the model solves the set of partial differential equations with a semi-implicit algorithm on a variable vertical grid and for a specified fixed time step, usually around 5 min, and at a 0.5 m depth resolution, ensuring an accurate description of mixing and stratification in the water column. Outputs from LAKEoneD encompass water temperature $T(z,t)$ and turbulent diffusivity $D(z,t)$, presented in regular vertical and time grids. An example of the hydrodynamic model results for Grahamstown Dam is shown in Figure 5. Here, the simulation results are compared against data from the thermistor chains deployed in the lake. The simulation shows good congruence with the measurements.

2.4.2. Cyanobacteria Growth Model

The main drivers of cyanobacteria growth are water temperature, underwater light conditions, and nutrient availability [70,71]. Here, we assume that the lakes are hypertrophic, where nutrient limitation of phytoplankton growth is negligible. While models are valuable tools for understanding and predicting ecosystem dynamics, they often fall short of capturing the full complexity of these systems. This is due to the potential lack of specific parameter values and the general reliance on literature values that may not accurately represent the unique conditions of a particular lake. Additionally, the scarcity of data to calibrate these models further complicates their accuracy and reliability. Our one-dimensional modeling approach avoids those complexities by reducing the system to simple growth relations depending on only physical parameters (light, temperature, and mixing) and, if available, nutrient limitation. The phytoplankton functional groups implemented in the original model differ in light and temperature limitation parameterization and buoyancy/sinking characteristics. Competition for light between phytoplankton functional groups can then be described by a reaction–advection–diffusion model [70,71] as in,

$${\displaystyle \frac{\partial C_i}{\partial t}}=p_i(I,T(z,t))C_i-L_i\left(T(z,t)\right)C_i+v_i{\displaystyle \frac{\partial C_i}{\partial z}}+{\displaystyle \frac{\partial }{\partial z}}\left(D(z,t){\displaystyle \frac{\partial C_i}{\partial z}}\right).$$

(5)

where the population density (cells per unit volume) of species i, $C_i$, is assumed to change via light intensity ($\mathsf{\mu }$molm²s⁻¹), I; temperature, T; dependent growth, $p_i(I,T)$; and a species-specific loss rate, $L_i$, which is a function of temperature. Cells can either rise and sink (e.g., buoyant cyanobacteria) or sink (e.g., diatoms and green algae) with a constant velocity $v_i$ (m/s). The last term describes the mixing of the cells via turbulent diffusivity, $D(z,t)$. A full description of the model and a glossary with a detailed explanation and physical meaning of the parameters can be found in [39].

2.4.3. A Structured Modeling Approach

The modeling of the lake follows a structured approach that integrates hydrodynamic and biological processes to simulate cyanobacteria distribution and growth. The schematic process chain is illustrated in Figure 6. The process begins with the initialization of the model using a predefined time series of inflow data, including flow rate, water temperature, and cyanobacteria cell count. These parameters establish the initial conditions required for accurate simulation. Once initialized, the LAKEoneD hydrodynamic model is employed to simulate thermal stratification and mixing. The interaction of meteorological factors, lake geometry (such as hypsometry and alignment for fetch impact), and flow characteristics determine the distribution of water temperature, stratification, and turbulence within the system.

Following the hydrodynamic simulation, the results are stored in a depth–time matrix for both temperature and turbulent diffusivity. This dataset forms the basis for subsequent modeling of cyanobacteria distribution and growth. Important hydrodynamic outputs, including surface and bottom water temperatures and thermocline depth, are also recorded. The last time-step is stored as vectors, encompassing water temperature, momentum, turbulent kinetic energy, and turbulent dissipation rate. This information can initialize subsequent simulations or forecasting applications, ensuring continuity in model cyanobacteria growth.

The final simulation step involves executing the cyanobacteria growth model, which utilizes the stored hydrodynamic data to simulate the distribution and proliferation of cyanobacteria. The model generates cell count concentrations as a depth–time matrix alongside aggregated time-series outputs for surface cell concentration and total cell count. The system simulates hydrodynamics independent from growth. However, growth is driven by water temperature, turbulent diffusivity, and the prevailing external meteorological conditions. This assumption allows for a “serialization” of the two models [39,72].

2.4.4. A Short-Term Forecast System

We use a simple forecast system. In the first step, the hydrodynamic model was run using the baseline parameterization to establish initial temperature stratification and mixing conditions. This model was then re-initialized daily using these conditions and forecasted meteorology to predict water temperature and diffusivity. The meteorological forecast data from MeteoBlue service included daily minimum and maximum temperatures, hourly wind speed and direction, and three-hourly cloudiness, all provided at non-equidistant time points. For each simulation day, a seven-day hourly forecast was generated by applying linear interpolation to the cloudiness and wind speed data and sinusoidal fitting for the minimum and maximum temperature data.

In the next forecasting step, the cyanobacteria growth model was run based on the hydrodynamic model’s outputs (temperature and diffusivity) and the forecasted weather and inflow rates. This model was initialized with observed cell counts from the forecast start date, serving as a prototype for operational cyanobacteria forecasting; see schematic in Figure 6. The output is predicted cyanobacteria growth under the forecast conditions.

Accurate seeding conditions are essential for modeling species behavior, as uncertainties rapidly arise without precise calibration using measured cyanobacteria biomass. We derived cyanobacteria indices over time from the high-frequency (daytime, 15 min temporal resolution) HydraSpectra surface spectral reflectance measurements.

In previous research [73], we evaluated several band ratio formulations using above surface spectral reflectance data and simultaneous chlorophyll data obtained over New South Wales inland water bodies, with the best outcome ($R^2$ = 0.91, RMSE = 3.53 $\mathsf{\mu }$g/L) achieved using a three-band algorithm,

$$Index=({\displaystyle \frac{1}{R_{rs}\left(684\right)}}-{\displaystyle \frac{1}{R_{rs}\left(711\right)}})\times R_{rs}\left(750\right),$$

(6)

where $R_{rs}\left(\lambda \right)$ stands for in situ remote sensing reflectance at wavelength $\lambda$ in nanometers.

These wavelengths were selected based on their sensitivity to phytoplankton abundance and have been validated in previous studies [74,75]. The algorithm is grounded in prior knowledge of the absorption characteristics of chlorophyll pigments and reflectance peaks associated with these pigments and is highly sensitive to cyanobacterial abundance [76]. Specifically, 684 nm corresponds to the chlorophyll-a absorption peak, 711 nm is near the reflectance shoulder associated with the chlorophyll-a pigment, and 750 nm lies in the near-infrared region where water-leaving reflectance is typically minimal but corrects for residual sediment reflectance in turbid waters, thus enhancing contrast. By leveraging these well-established spectral features, we ensure that the reflectance-based index used in this study is biologically meaningful and optimized for detecting abundance in cyanobacteria-dominated inland waters.

Here, the time series of indices were compared to the simulated surface appearance of cyanobacteria and the closest in situ cell count samples in the time spectrum to establish a linear relation, which was then used to convert the cyanobacteria index to cell count estimates. Daily cell counts were then estimated from HydraSpectra cyanobacteria indices. While this three-band chlorophyll algorithm has been shown to perform well in mesotrophic and eutrophic waters [77], its sensitivity decreases under low-biomass conditions (<10 $\mathsf{\mu }$g/L), which may reduce accuracy in oligotrophic systems [78]. Despite this limitation, we retained the algorithm in both study lakes to ensure methodological consistency and enable direct cross-comparison across trophic states. This consistency is critical for evaluating the framework’s applicability under contrasting ecological conditions. Further refinements, such as locally calibrated optical models or adaptive band-ratio approaches, are discussed in Section 4.

We ran growth simulations with cyanobacterial cell concentrations from the estimated values at a specific date and time. This estimated cell count value was initially distributed over a specific depth range, and the growth was simulated for the following days. Vertical changes in cyanobacterial concentration due to buoyancy were also simulated, where cyanobacteria accumulate over the day in the upper water layers and are mixed down during the evening/night when nighttime cooling starts to mix the daily mixing layer. However, the wind will also change the mixing structure. Thus, the mixing depth will vary with the system’s mixing state.

The cyanobacteria model was initialized to maintain forecast precision with measured cell counts on an actual day. Thus, this model is the prototype of an operational forecast model for cyanobacteria as in day-by-day weather predictions.

3. Results

3.1. Blue–Green Algal Growth Monitoring

The results from both model simulations and hyperspectral-derived indices for October 2019 to April 2020 in Lake Hume are presented in Figure 7.

The correlation between hyperspectral-derived indices and grab-sample cell counts is shown in Figure 8 and

Table 1. Cyanobacteria hyperspectral derived indices performance measured by RMSE, MAPE, $R^2$ (COD), and Pearson’s r in Lake Hume and Grahamstown Dam.

Lake	RMSE (Cells/mL)	MAPE (%)	${\mathit{R}}^2$ (COD)	Pearson’s r
Hume	8817	17.04	0.31	0.56
Grahamstown Dam	696	27.83	0.52	0.43

. Using the three-band algorithm (Equation (6)), relationships were approximately linear. For Lake Hume, Pearson’s $r=0.56$ indicates moderate correlation, with $R^2=0.31$ and adjusted $R^2=0.30$, consistent with the variability reported in previous studies [80,81]. For Grahamstown Dam, correlation was lower ($r=0.43$), but $R^2=0.52$ suggests reduced scatter.

Variability reflects short-term mixing dynamics, mismatched sampling times, and limitations of assuming simple linearity between reflectance and biological processes. Nevertheless, hyperspectral-derived indices reproduced overall cyanobacteria growth patterns, capturing seasonal increases and declines.

For Lake Hume, agreement between HydraSpectra indices and simulated surface cell counts was moderate, with RMSE = 8817 cells/mL and MAPE = 17.04% over a range of ≈75,000 cells/mL. Discrepancies likely arose from spatial heterogeneity, model input uncertainty, and local mixing. Despite this, simulations tracked seasonal and interannual dynamics, including bloom onset and decay, and aligned with multi-site species counts. High cell count outliers were driven by very small cyanobacteria contributing little to total biovolume.

At Grahamstown Dam, hyperspectral estimates followed seasonal trends but diverged from simulations during bloom events, with RMSE = 696 cells/mL and MAPE = 27.83% over ≈2500 cells/mL. This polymictic system is shaped by frequent mixing, wind-driven turbulence, and variable inflows [82,83], which complicate vertical distributions of buoyant species like Microcystis. These dynamics partly explain the discrepancies between surface-detected HydraSpectra signals and depth-integrated model outputs. Calibration with annualized initial biomass values, while keeping the physiological parameters from laboratory data [84], improved model skill and yielded close alignment during several bloom events (Figure 9 with arrows).

Importantly, the model captured bloom timing and trends, which is the study’s primary objective. For example, Figure 7b shows agreement between hyperspectral-derived, simulated, and in situ counts at the 25 February 2020 bloom, consistent with the official WaterNSW alert. This demonstrates the potential for integrating hyperspectral data with hydrodynamic models to deliver operational early warnings.

3.2. Cyanobacteria Bloom Forecasting

Forecasting results for both lakes are shown in Figure 10. The model was initialized daily using peak cyanobacteria concentrations derived from HydraSpectra reflectance indices, distributed across predefined depth ranges (0–1 m, 0–2 m, or full depth). Growth was then simulated forward for several days, with forecasts compared against a baseline simulation calibrated with year-round hyperspectral data.

Forecast simulations generally reproduced seasonal growth dynamics, with normalized RMSE values of 23% for Lake Hume (LH) and 26% for Grahamstown Dam (GTD), indicating moderate skill. Agreement reflects the influence of mixing depth on surface concentrations: nighttime cooling increased the LH mixing depth from ≈1 m to 5–7 m and fully mixed the ≈9 m GTD water column, diluting surface counts. Discrepancies occurred when high initialization values were rapidly redistributed to depth, producing lower-than-expected surface concentrations. These errors may stem from uncertainties in hyperspectral estimates, misrepresentation of short-term mixing processes, or biases in meteorological forcing (e.g., modeled winds vs. observed calm conditions). Cross-validation with chlorophyll fluorescence or real-time meteorology would help reduce these uncertainties.

Forecasts were sensitive to high initial cell counts. For example, between 15 and 22 February 2020 at LH, inflated HydraSpectra values propagated into exaggerated bloom projections. Adaptive averaging or filtering of outliers could improve forecast robustness. Forecasts also assumed that the daily maximum cell count corresponded to the mixing depth, estimated from hydrodynamic simulations. Multiple depth scenarios were tested, with the most consistent selected for each forecast. However, in polymictic GTD, transient stratification events (e.g., 11–16 February 2020) were poorly captured when fixed mixing depths were assumed, underestimating bloom intensity. These findings highlight the need for real-time mixing diagnostics in forecast frameworks.

Scenario simulations (Figure 11b) showed that ±4 °C air temperature changes over seven days had negligible effects on surface abundance, reinforcing the dominant role of mixing processes in shaping vertical cyanobacteria distributions.

These results underscore that reflectance-based indices mainly capture surface concentrations, which may not represent total abundance. Figure 11 illustrates five years of LH simulations: surface counts (0–1 m) varied strongly, while depth-integrated counts followed a smoother trend. A detailed two-week period captured the emergence and decline of a surface bloom, with nearly all cells concentrated in the upper 1 m on 24 February 2020, corroborated by WaterNSW’s red alert. Yet two days later, CSIRO sampling at the pontoon showed no bloom, reflecting spatial and temporal heterogeneity.

Overall, the forecasting model captured bloom timing and magnitude with moderate skill but was sensitive to initialization errors and mixing assumptions. These limitations emphasize the need for uncertainty quantification in hyperspectral retrieval, sensitivity analyses of initial conditions, and improved coupling with hydrodynamic diagnostics.

4. Discussion

4.1. Discussing Our Findings

Our results highlight the contrasting cyanobacterial bloom dynamics in a monomictic (Lake Hume) versus polymictic (Grahamstown Dam) system. HydraSpectra provided high-frequency hyperspectral reflectance data that, through a three-band Chl-a algorithm, enabled the estimation of cell counts and integration into a forecasting framework. In Lake Hume, model–data agreement was moderate ($R^2$ = 0.31, RMSE = 8817 cells/mL, MAPE = 17%), while in Grahamstown Dam, correlations were lower ($R^2$ = 0.52, RMSE = 696 cells/mL, MAPE = 27.8%) due to frequent mixing and low biomass conditions. These findings are consistent with previous observations of increased uncertainty in dynamic or oligotrophic systems [80,81,85].

Hydrodynamic simulations reproduced seasonal bloom timing, but sub-daily mixing variability (e.g., diel stratification cycles and wind-driven turbulence) was a major source of error, particularly in polymictic GTD. Forecast sensitivity analysis showed that initialization errors from high outlier HydraSpectra values propagated into inflated bloom predictions (e.g., Lake Hume, 15–22 February 2020). These results emphasize that while the integrated hyperspectral–hydrodynamic framework is effective for short-term bloom forecasting, reliability is constrained by retrieval errors, initialization quality, and mixing depth uncertainty.

4.2. Limitations and Uncertainties

The predictive framework demonstrated moderate skill, but performance was constrained by several sources of uncertainty. These uncertainties arise primarily from hyperspectral retrieval, hydrodynamic mixing representation, and initialization quality, and each can be quantified based on our results.

Table 2 summarizes the key sources of uncertainty, their mechanisms, quantified impacts, and potential mitigation strategies. In brief, we have the following:

• Hyperspectral retrieval: The three-band Chl-a algorithm performed well in eutrophic Lake Hume but was less reliable in oligotrophic conditions (Chl-a < 10 $\mathsf{\mu }$g/L, typical of GTD, see Figure A1 for reference). Driven by wavelengths in the red and near-infrared region (680 to 750 nm), this algorithm has been shown to provide high correlation to chlorophyll for particular application to the higher chlorophyll concentrations experienced in inland water bodies [77], and was calibrated for Australian inland waters [73]. The algorithm using red/NIR reflectances introduces additional uncertainty in low-chlorophyll regimes, where signal-to-noise ratios are lower and absorption features are weaker. Retrieval error increased normalized RMSE to 28% in GTD compared with 12% in LH. This highlights the need for refined algorithms tailored to low-biomass waters.
• Hydrodynamic mixing depth: Forecast skill was highly sensitive to mixing depth representation. In GTD, frequent diel mixing and inflows caused errors when fixed depths were assumed. Strong mixing can carry buoyant cells to lower than the surface mixing depth, and dilute the initialized-model cell count value, which was estimated from surface-derived hyperspectral reflectance. On the other hand, reduced mixing depth can promote cyanobacterial abundance from the higher availability of photosynthetically active radiation [86]. The real-time assimilation of thermistor-chain data would improve forecast accuracy in such polymictic systems. Our results show that the forecast model initiated by hyperspectral-derived cell count works better for a monomictic lake than a polymictic lake unless there is a real-time measurement of accurate mixing depth supply, which is generally a difficult task in many remote regions.
• Simplifications in the 1D model: Hydrodynamic modeling is a widely used method for assessing vertical temperature stratification and mixing dynamics within the water column of lakes [87,88,89]. This study employed a one-dimensional (1D) vertical process model to simulate water column dynamics, assuming lateral uniformity across large lake areas. Our computationally efficient one-dimensional model has certain limitations, such as assuming lateral uniformity and excluding nutrient dynamics, as well as reduced skill in more heterogeneous systems, for example, the 1D model struggles to capture the 3D processes of seasonal changes in flow regime, resulting in overly warm mixed water in autumn, as compared to actual measurements (e.g., Figure 5 in [79]). While three-dimensional (3D) models offer greater spatial detail, their high computational and storage demands make them less practical for long-term operational use. In contrast, one-dimensional models provide a computationally efficient alternative, requiring fewer historical data inputs than data-driven models [90,91,92].
  Moreover, phytoplankton growth is influenced by physical factors such as temperature, light, and mixing, but additional biological and chemical processes also play a role. Jöhnk et al. [39] developed a simplified, generalizable model for cyanobacterial growth by integrating the one-dimensional hydrodynamic model LAKEoneD [37,38,72] with an algal competition framework. This model primarily considers light and temperature as limiting factors, assuming nutrients are non-limiting. It effectively simulates algal growth in eutrophic lakes or during short-term nutrient pulses but struggles with long-term dynamics without incorporating a life cycle component. The single-species growth model was parameterized using growth data for a specific species, which might not adequately reflect the behavior of other species and the complex competition process of multiple species in Grahamstown Dam, in particular. However, it allows for the focus on toxin-producing cyanobacterial species such as Chrysosporum ovalisporum, a dominant species in 2015–2016, 2018–2021, and Microcystis in 2016–2018, in Lake Hume. It also allows for the easy deployment of systems with largely unknown biogeochemical characteristics and complex food webs. Other limitations include the inability to represent nutrient pulses from decay from viral lysis [93], wind-driven resuspension, and the effects of UV radiation [94]. The position of cyanobacterial colonies in the water column is critical to their light exposure, nutrient acquisition, and ultimately to their ability to dominate the phytoplankton community and to produce toxic blooms. Future work should also target more sophisticated ecological models of phytoplankton dynamics, such as those described by Hense and Beckmann [95, 96], for enhanced predictive capability.
• Initialization: Forecasts initialized with HydraSpectra-derived outliers led to exaggerated bloom projections (e.g., LH, 15–22 February 2020). In LH, an ≈+30% initialization error increased MAPE from ≈17% to ≈25%. Adaptive filtering of input data would mitigate these effects.
• Additional factors: Mismatch in sampling times and locations introduced discrepancies (e.g., WaterNSW red alert vs. no detection at CSIRO pontoon).

Table 2. Summary of key uncertainties in the hyperspectral–hydrodynamic bloom forecasting framework. The table explicitly quantifies how errors propagate from (i) hyperspectral retrieval, (ii) hydrodynamic mixing representation, and (iii) initialization of forecasts, with additional sources of variability noted. Reported values are based on sensitivity analyses and validation statistics from Lake Hume (LH) and Grahamstown Dam (GTD), illustrating the different uncertainty profiles of monomictic versus polymictic systems. Proposed mitigation strategies highlight future directions for improving model robustness and operational forecasting.

Uncertainty Source	Mechanism	Quantified Impact	Mitigation Strategy
Hyperspectral retrieval (Equation (6))	Three-band Chl-a algorithm less accurate in low-biomass waters; weak absorption at 673 nm increases noise [85]	At Chl-a < 10 µg L−1 (e.g., GTD), retrieval errors increased normalized RMSE to 28% (vs. 12% for LH)	Develop/refine algorithms for oligotrophic conditions; calibrate with local optical datasets; use adaptive band ratios; use additional spectral bands or machine learning–based approaches
Hydrodynamic mixing depth	Errors in simulated mixing depth propagate into surface concentration predictions, especially in polymictic systems	GTD forecasts underestimated bloom intensity (11–16 February 2020)	Assimilate high-frequency temperature profiles (thermistor chains); explore real-time buoy data; consider 3D modeling in polymictic lakes
Initial conditions (HydraSpectra-derived)	Outlier cell count estimates inflate forecast initialization	In LH, +30% initialization error increased MAPE from 17% to 25% (3-day horizon)	Apply adaptive filtering/averaging to remove outliers; assimilate depth-integrated in situ counts when available
Temporal/spatial mismatch	Misalignment of grab samples, hyperspectral data, and meteorological forcing introduces bias	Leads to discrepancies between modeled and observed peaks (e.g., false negatives at CSIRO pontoon vs. WaterNSW alert, February 2020)	Align sampling schedules; integrate multi-source data (fluorescence sensors, meteorological stations) for validation
Model simplifications	1D model assumes lateral uniformity; excludes nutrient pulses, viral lysis, or inter-species competition	Overly warm mixed waters in autumn; reduced accuracy in GTD multi-species periods	Incorporate nutrient dynamics; extend to multi-species models [95,96]; test 3D models in heterogeneous systems

Table 2. Summary of key uncertainties in the hyperspectral–hydrodynamic bloom forecasting framework. The table explicitly quantifies how errors propagate from (i) hyperspectral retrieval, (ii) hydrodynamic mixing representation, and (iii) initialization of forecasts, with additional sources of variability noted. Reported values are based on sensitivity analyses and validation statistics from Lake Hume (LH) and Grahamstown Dam (GTD), illustrating the different uncertainty profiles of monomictic versus polymictic systems. Proposed mitigation strategies highlight future directions for improving model robustness and operational forecasting.

Uncertainty Source	Mechanism	Quantified Impact	Mitigation Strategy
Hyperspectral retrieval (Equation (6))	Three-band Chl-a algorithm less accurate in low-biomass waters; weak absorption at 673 nm increases noise [85]	At Chl-a < 10 µg L−1 (e.g., GTD), retrieval errors increased normalized RMSE to 28% (vs. 12% for LH)	Develop/refine algorithms for oligotrophic conditions; calibrate with local optical datasets; use adaptive band ratios; use additional spectral bands or machine learning–based approaches
Hydrodynamic mixing depth	Errors in simulated mixing depth propagate into surface concentration predictions, especially in polymictic systems	GTD forecasts underestimated bloom intensity (11–16 February 2020)	Assimilate high-frequency temperature profiles (thermistor chains); explore real-time buoy data; consider 3D modeling in polymictic lakes
Initial conditions (HydraSpectra-derived)	Outlier cell count estimates inflate forecast initialization	In LH, +30% initialization error increased MAPE from 17% to 25% (3-day horizon)	Apply adaptive filtering/averaging to remove outliers; assimilate depth-integrated in situ counts when available
Temporal/spatial mismatch	Misalignment of grab samples, hyperspectral data, and meteorological forcing introduces bias	Leads to discrepancies between modeled and observed peaks (e.g., false negatives at CSIRO pontoon vs. WaterNSW alert, February 2020)	Align sampling schedules; integrate multi-source data (fluorescence sensors, meteorological stations) for validation
Model simplifications	1D model assumes lateral uniformity; excludes nutrient pulses, viral lysis, or inter-species competition	Overly warm mixed waters in autumn; reduced accuracy in GTD multi-species periods	Incorporate nutrient dynamics; extend to multi-species models [95,96]; test 3D models in heterogeneous systems

Addressing these limitations requires targeted improvements: (i) refining spectral algorithms for low-biomass waters; (ii) assimilating real-time high-frequency thermal data into hydrodynamic models; (iii) applying adaptive filtering for initialization; and (iv) integrating depth-resolved and multi-species data. Together, these enhancements would enable more robust operational forecasts of cyanobacterial blooms, especially in polymictic systems.

5. Conclusions

This study integrates remote sensing using data from an above-surface sensor to enhance water quality monitoring and cyanobacterial bloom forecasting. Timely detection and forecasting are essential for minimizing risks, yet traditional monitoring methods are often slow and costly. This study used hyperspectral reflectance measurements from the emerging HydraSpectra instrument and incorporated hydrodynamic modeling coupling with algal growth simulations to forecast cyanobacteria blooms. The framework was applied to two Australian reservoirs: Lake Hume, a monomictic reservoir used for irrigation and hydropower, and Grahamstown Dam, a polymictic reservoir supplying drinking water. Both these lakes have experienced a growing occurrence of toxic algal blooms. Cyanobacterial blooms pose a significant challenge in many freshwater systems. This study highlighted their environmental, social, and economic impacts, underscoring the need for improved forecasting to guide mitigation actions, inform water quality policies, and support effective management. The major findings are as follows:

• This study demonstrated the strong correlation between hyperspectral-derived spectral reflectance, chlorophyll-a concentrations, and cyanobacteria cell counts under bloom conditions, with the highest agreement observed using a three-band chlorophyll-a index algorithm. When integrated into a growth model coupled with a hydrodynamic model, this data proved effective for forecasting blooms. While our one-dimensional hydrodynamic model did not account for all lake dynamics, it proved a valuable tool for understanding vertical mixing processes and their role in cyanobacterial distribution.
• Hyperspectral reflectance-based cyanobacterial indices predominantly represent surface conditions, which may not fully account for variations throughout the water column. Our findings indicated that daily fluctuations in cyanobacterial surface concentrations are closely linked to changes in mixing layer depth, underscoring the importance of high-resolution environmental data and well-calibrated models for accurate forecasting. To address this limitation, surface cell count values were assumed to represent a specific depth range, determined through hydrodynamic modeling, and calibrated using in situ measurements. Comparing simulations across different depth scenarios will improve mixing dynamic estimates and enhance overall forecasting accuracy. Additionally, these results suggest that satellite-based cyanobacteria detection may be unreliable without considering mixing processes.
• The forecasting model developed in this study provided robust short-term predictions of surface cyanobacteria concentrations. However, improving forecasting accuracy necessitates a more comprehensive system that integrates continuous weather monitoring, hyperspectral sensing for surface cell count observations, and high-resolution thermistor chain data to track mixing dynamics, especially in polymictic systems. Combining these components with hydrodynamic and growth models would refine localized forecasts, strengthening early warning systems for effective water quality management.

Global freshwater availability and quality continue to decline due to climate change, land overuse, drought, bushfires, and pollution, while anthropogenic demand continues to rise. The effective monitoring and management of water resources are essential for sustaining ecosystems and meeting human, agricultural, and industrial needs. Despite the United Nations’ Sustainable Development Goal 6 [97], aiming for universal access to clean water by 2030, progress remains inadequate. The forecasting tools developed as part of this study strongly support the early warning of emerging water quality issues, which is vital for mitigating pollution and ensuring sustainable resource management.