A Low-Cost Fluorometer Applied to the Gulf of Saint Lawrence Rhodamine Tracer Experiment

In order to increase the spatiotemporal resolution and accessibility of freshwater and marine data, low-cost in situ fluorometers are required. The devices must be robust, fully submersible, and sensitive in the 1 ppb–1 ppm range for compounds such as Rhodamine water tracer (RWT), a dye used in time of travel, and substance dispersal measurements. In this work, we introduce a deployment-ready, low-cost, small form-factor RWT fluorometer prototype based on the principle of lock-in amplification. Using measurements collected from remotely operated underwater vehicle (ROV) deployments during a dye-tracer experiment in the Gulf of Saint Lawrence, we compare its performance against that of a widely used commercial fluorometer. The results of the prototype’s calibration and deployment show competitive performance against commercial instruments, with a limit of detection (LOD) below 0.2 ppb and for a cost of U.S. $\$ $ 744.70, a fraction of the cost of relevant commercially available in situ fluorometers.


TABLE I CHARACTERISTICS OF SELECT COMMERCIAL FLUOROMETERS
Most contemporary submersible fluorometers follow the Chelsea/Turner model and are configured to target a single fluorophore. Some are configured with multiple channels to target multiple fluorophores independently. These systems can also perform nonfluorometric measurements such as turbidity.
Chelsea Technologies markets their "UviLux" fluorometers for water quality assessment and oil detection, as well as a three-channel "TriLux" fluorometer for HAB detection and monitoring algae. The TriLux has a limit of detection (LOD) of 0.1 ppb chlorophyll-a. In this context, 1 ppb is defined as a concentration of 1 µg/L.
Sea-Bird Scientific (Bellevue, WA, USA) produces the "ECO-FL" line of submersible single-channel fluorometers, having LODs of 0.02 ppb for chlorophyll-a and 0.05 ppb for RWT.
Seapoint Sensors (Exeter, NH, USA) also produces a self-titled line of fluorometers, with an LOD of 0.02 ppb for both chlorophyll-a and RWT.
The characteristics of these fluorometers, along with their most recent available price, are summarized in Table I. A conservative estimate of the market size of these fluorometers is in the tens of thousands.
There is also competing work to produce fluorometers below the U.S. $1000 threshold. Zieger et al. [17] developed a four-channel fluorometer prototype for algal classification for under U.S. $400. However, the design is a flowthrough type known to be less convenient and effective in underwater field use [3]. Flowthrough systems have added complexity with fluid pumping and in dealing with biofouling on enclosed optical windows. Haidekker et al. [18] designed a one-channel in vivo chlorophyll fluorometer for U.S. $300 [18]. The device illuminates samples in free space (i.e., benchtop only) and is not yet ported into a submersible housing, which would add significant cost. While there are reports of low-cost fluorometers, there remains a need to realize submersible and low-cost fluorometers that compare well with commercial instrument performance.
In this work, we present a novel in situ fluorometer ready to use at depth for U.S. $744.70 based on off-the-shelf components, 3-D-printed parts, and an advanced signal processing scheme; the digital lock-in amplifier. The prototype fluorometer uses a sinusoidal modulation of LED light to produce excitation and emission that pulsates at the same modulation frequency so that the fluorescence can be distinguished from sources of noise and interference. The prototype targets RWT fluorescence, but other fluorophores can be targeted by swapping optoelectronics and optical filters with drop-in replacements with no added complexity. A rendering of the prototype and an illustration of in situ RWT fluorometry are presented in Fig. 1.
We calibrated our prototype alongside two representative commercial RWT fluorometers and demonstrated comparable performance. Following calibration, we deployed our prototype along with a commercial fluorometer on an ROV in the Dalhousie University Aquatron Laboratory (Dalhousie University, Halifax, NS, Canada). This platform was then used in a field comparison study in the Gulf of Saint Lawrence of northeastern Canada for the purpose of time of travel and substance dispersal measurements. The work was performed under the Marine Environmental Observation Prediction and Response Network (MEOPAR, Halifax, NS, Canada) and Reseau Quebec Maritime (RQM, Rimouski, QC, Canada) Tracer Release Experiment (TReX), an RWT experiment conducted between September 5 and 9, 2021.
From the laboratory to the field, our novel prototype fluorometer offers comparable performance to commercial systems, at a fraction of the cost.

A. Novel Fluorometer
The single-channel prototype fluorometer presented in this work is fully submersible to a depth of at least 500 m and has a mass of 400 g and a volume of less than 380 cm 3 . It costs U.S. Normalized spectra/filter responses of the chosen optical source, the excitation and emission of RWT, the solar background, and the chosen excitation and emission filters. Filter responses ensure the rejection of the overlapped source and emission spectra and the rejection of solar spectrum before the optical detector.
$744.70 to assemble at the time of authorship, substantially less than the purchase price of relevant commercial competitors [3], [14], [15], [16], using only two custom components. The prototype targets RWT, which is maximally excited by 555-nm light and exhibits peak emission at 580 nm [7].
The prototype is housed in prefabricated, off-the-shelf cylindrical components by BlueRobotics (Torrance, CA, USA). The interior structural components are 3-D-printed in polylactic acid (PLA) using the CreatorPro 3-D-printer by FlashForge (Jinhua, Zhejiang, China) with default tolerance settings. Black PLA filament is used to absorb stray light inside the fluorometer. The optics are sourced from Edmund Optics (Barrington, NJ, USA) and are mounted and aligned in the fluorometer using 3-D-printed housings.
A black-tinted custom acrylic endcap is hand-milled from 6-mm acrylic stock to match the profile and screw-hole locations of the BlueRobotics clear acrylic endcap.
A custom PCB, repurposed from another project, controls the device. It is interfaced using RS-232 through a six-pin SubConn connector. An architecture diagram of the custom PCB is included in the Appendix. A purpose-driven PCB redesign would be significantly cheaper. The prototype's bill of materials (BOM) is presented in Table II.

B. Optical Design
Our optical design task was to create an optical stack that exposes a sample volume of seawater with excitation in direct view of the optical detector. The stack must use only off-theshelf or 3-D-printed components.
The optical source must overlap the excitation spectrum of RWT, while optical filters must reject the overlap between the source and emission spectra to minimize optical interference (i.e., crosstalk from reflection and backscattering) and must reject background noise (i.e., solar spectrum intrusion). Fig. 2 shows the wavelength spectra and filter responses of the chosen optical stack components, the excitation and emission spectra of RWT from Laidlaw and Smart [7], and the worst case solar spectrum (i.e., neglecting atmospheric and aquatic absorbances). The spectrum of the optical detector used in this work is relatively flat over this wavelength range and is excluded for clarity.
As required in our design task, Fig. 2 shows that the narrowband emission filter passes RWT fluorescence while rejecting most of the solar spectrum and rejects the band of wavelengths where the source and excitation spectra overlap. The excitation filter further minimizes the overlap between the source and the emission spectra in the case where temperature increases or part variations shift the source spectrum to the right. In our design, a green LED optical source is aligned with a collimating lens, and the excitation filter matched to the excitation spectrum peak of RWT. The filtered and collimated excitation transmits through the clear acrylic endcap where it illuminates a sample volume of seawater.
If present, the Rhodamine will fluoresce in response and this emission is collected by a collimating lens. The collected light is filtered by the emission filter matched to the emission spectrum peak of RWT so that only RWT fluorescence passes. The collimated emission is then focused onto a detecting photodiode, with integrated transimpedance amplifier, by a hemispherical focusing lens. The specifications and names of these parts can be found in the BOM, see Table II. Fig. 3(a) presents a sectioned view of the optical system in our prototype. The optics, optical source and detector, sample volume, and optical paths of the excitation and emission are labeled for clarity. Fig. 3(b) shows the prototype photographed in a side view. Fig. 3(c) shows the prototype photographed in a bottom (or sensing end) view, with superimposed sectioning lines used to indicate the view presented in Fig. 3(a).

C. Ray-Tracing Simulations
Prior to realizing the prototype fluorometer, we simulated the optical stack with Zemax OpticStudio to optimize its design [19]. The model and the simulation output are shown in Fig. 4.
OpticStudio performed 3-D ray tracing to test the lens-filter stack, simulate the photodiode's response to fluorescence, and optimize feature placement. The ray-tracing model includes the prototype's opaque housing to block stray rays, and a fluorescing volume characterized to match the refractive index and flourescence spectra of RWT dissolved in seawater with a typical [20] quantum yield of 31%. The LED is simulated as a point source of rays, whereas the photodiode is simulated as a subsystem of three parts: a clear lens, a thin aperture, and a square detecting area. The volume of water interrograted by the sensor-the sample volume-is approximately 1.2 cm 3 . Fig. 4(a) offers a side view of the bundle of rays, colored blue, produced by the LED. Fig. 4(b) shows the simulated irradiance pattern detected by the photodiode. The coordinates of each dot represent where a fluorescence ray strikes the simulated photodiode and the grayscale of each dot represents that ray's irradiance. The output of the simulation is taken as the total incident optical power detected by the photodiode, indicated in the red box. The simulations are repeated for a range of RWT concentrations in the sample volume, each requiring approximately three days to complete.
The total incident optical power results from the Optic-Studio simulations are entered into a simple spreadsheet that simulates the photodetector's photocurrent and transimpedance amplifier responses. The spreadsheet produces a simulated calibration curve, plotting amplifier output voltage per sample Rhodamine concentration (molarity). The simulated calibration curve is shown in Fig. 5, illustrating the expected linear (R 2 > 0.99) relationship between fluorophore concentration and detector voltage.

D. Laboratory-Based Calibration and Verification
The prototype, along with two representative instruments of currently available technology to compare to the prototype's performance, was taken to the CERC.OCEAN Laboratory at Dalhousie University and calibrated against known standards of RWT prepared in solution with seawater from the Aquatron Laboratory. The Aquatron houses seawater tanks ranging from small purpose-driven containers to six larger tanks totaling over 2000 m 3 of seawater pumped from the Northwest Arm (Halifax, NS, Canada), making it an extremely attractive location to reliably simulate in situ testing.
Rhodamine WT 400 ppb (Turner Designs, San Jose, CA, USA) standards were each prepared in 2-L beakers by dilution with seawater dispensed from the Aquatron to target concentrations of 0.5, 2, 10, 30, and 60 ppb, as well as a blanking standard of undyed seawater, 0 ppb. These concentrations are chosen from past deployments indicating that the observable range of RWT was between 0.5 and 60 ppb [19].
The RWT standards were measured using three fluorometers: the prototype, a Turner Cyclops-7 (Turner Designs, San Jose, CA, USA), and an AML X2Change powered by Turner Designs (AML oceanographic, Dartmouth, NS, Canada). The prototype and AML devices stream digital data, whereas the Turner device is analog and is measured with a multimeter (Model number 34460A, Keysight Technologies, Kanata, ON, Canada). The commercial devices have three manual gain settings (1×, 10×, 100×) and are set to 10×, whereas the prototype needs no adjustment across its sensitive range.
For each calibration point, the standard in its 2-L beaker was placed in a temperature bath housed inside a light-blocking chamber and kept well-mixed using a magnetic stirrer. The tested sensor was mounted with the sensing end submerged in the solution and pointed down, axially aligned with the beaker at a consistent height across all three devices. A Lauda Eco Gold immersion thermostat (Lauda, Delran, NJ, USA) was suspended in the bath. Each standard is cycled through 5 • C, 8 • C, 11 • C, 14 • C, 17 • C, and 20 • C equilibria, taking approximately 1 h to reach each desired temperature. At each equilibrium, each sensor sequentially collected 5 min worth of data. When the sensors were switched, extra time was taken to allow the bath to return to the desired temperature equilibrium before data collection resumed.

E. Large Tank Study and Field Deployment
After calibration, the larger tanks in the Aquatron Laboratory were used to deploy the fluorometers in a controlled environment [19]. The prototype and the AML fluorometer were mounted to a BlueROV2 (BlueRobotics). The ROV was driven through plumes of RWT as it was dispensed, recording the agreement in the RWT concentration reported by the two fluorometers.
Finally, the prototype was deployed along with the Turner Cyclops-7 (standard version of their "Fast" 7F variant) during MEOPAR's TReX, mounting them to the same ROV used in the Aquatron testing. The TReX deployments occurred on September 5-9, 2021. Each day 300 L of 1% RWT was dispensed over a period of approximately 4 min into the Gulf of Saint Lawrence, just north of Rimouski, QC, 48 • 35 ′ 01.0 ′′ N, 68 • 31 ′ 08.8 ′′ W just before noon, except on the 9th where 900 L of 0.44% RWT was dispensed over a period of approximately 6 min. Rhodamine concentration time series were recorded, along with ROV depth, to show how the prototype performs in a head-to-head comparison with an industrial representative, with favorable results.

F. Lock-In Fluorometry
One aspect of this prototype fluorometer is that it uses digital lock-in amplification to extract sinusoidally modulated emission signals from the targeted fluorophore. An all-digital lock-in design has advantages in cost and complexity over the expensive and bulky analog-mixer alternative, as the algorithm is implemented on the same microprocessor that handles device control and data interfacing.
A fluorometer that implements a lock-in amplifier (a lockin fluorometer) has performance advantages compared to the conventional constant-level illumination approach. Fig. 6 presents a simplified block diagram for a lock-in fluorometer.
The signal of interest (i.e., emitted fluorescence) is captured and processed at an intermediate frequency. This separates the signal from low-frequency noise such as amplifier red noise or the in-band component of sunlight that is not rejected by the optical emission filter, preventing signal corruption that is unavoidable in a constant-illumination approach.
The sample volume is illuminated with a modulated excitation signal that is produced by the LED and is in-phase with and identical in frequency to the lock-in reference, V ref,x . A percentage (the quantum yield) of the photons absorbed by the fluorophore in solution results in emission of new photons at a longer wavelength. When an excited fluorophore molecule returns directly to the ground state and emits a photon, this emission is fluorescence [21] and the process occurs over a natural lifetime typically on the nanosecond timescale or 1-5 ns for the Rhodamine family of dyes [22].
Since the natural lifetime τ is much shorter than the modulation period of the excitation signal (0.1-10 ms in our case or 0.1-10 kHz), the emission signal detected at the photodiode is proportional to and time-shifted relative to the excitation signal produced by the LED, whereas only the ac signal components are relevant, if the excitation signal P ex (t) is described by the following: where P is the peak excitation optical power produced by the LED and f l is the frequency of the lock-in amplifier reference; then, the ideal emission signal detected at the photodiode and converted to voltage by its supporting electronics V em (t) is described by the following: where R is proportional to the concentration of the dissolved fluorophore in the sample volume and is the phase shift in the emission signal, relative to the lock-in reference, caused by the natural lifetime of the fluorescence process. The emission signal is mixed with voltage signals in-phase and in-quadrature with the lock-in reference, creating a pair of intermediate signals each containing a sum-frequency component and a difference-frequency (in this case, dc) component. The intermediate signals are then fed through low-pass filters to remove the sum-frequency component, producing an in-phase signal x(t) and an in-quadrature signal y(t). If the amplitude of the references used in the mixers is V ref , then these signals can be described by the following: and From (4) and (5), equations for R and θ directly follow: While this basic treatment considers R and θ to be constants, the analysis and its results are directly applicable to slowly varying cases of either (or both). In practice, background noise and interference, such as solar-spectrum intrusion, reflections and scattering, or incidental fluorescence, is injected into the system in the detection process making R and θ estimates of their ground-truth values.
To illustrate the benefit of the lock-in fluorometer over the conventional dc approach, a lock-in amplifier is simulated in MATLAB. A reference signal is generated at 1 kHz and used to produce a background-and-noise-free emission signal with an amplitude of 0.5 arbitrary units (AU) and a phase shift of 45 • relative to the reference signal. In this case, the ideal values of R and θ are 0.5 • and 45 • , respectively. A pure white Gaussian noise signal is generated with signal power that, in the mean squared sense, is four times that of the emission signal for a signal-to-noise ratio of 0.25 or −6 dB. This is also the signal-to-noise ratio in the comparable dc approach.
The simulated emission signal, both noiseless and with noise, is shown in Fig. 7(a). After mixing and filtering with a simple sliding average filter (chosen for its simplicity and deliberately because of its poor filtering performance), the in-phase and in-quadrature signals x(t) and y(t) are plotted together on a coordinate plane, see Fig. 7(b). The red circled dot illustrates the ideal values of x(t) and y(t) extracted from the noiseless case and falling exactly on the expected 45 • dashed line, whereas the blue dots are scattered due to the presence of noise. Fig. 8 shows the output of the lock-in amplifier. In Fig. 8(a), the real-time computed value of R is presented in the noisy case, the noiseless (ideal) case, and the noisy case once again after filtering with a simple first-order Butterworth lowpass filter. In Fig. 8(b), the original background-and-noise-free emission signal is plotted next to a reconstruction of the emission signal using the average values of R and θ extracted from the noisy emission signal case. The final signal-to-noise ratios on R and θ are 22.7 and 19.7 dB, respectively. This is an improvement of 28.7 dB on R over the comparable dc approach, and an extraction of the phase change θ that a dc approach is completely insensitive to.

A. Concentration and Temperature Calibration
The results of the laboratory-based calibration and verification are given in Fig. 9. In Fig. 9(a), the prototype sensor output is plotted against standard concentration, and curves are plotted for each of the equilibrium temperatures indicating strong linear relation (R 2 > 0.99) between sensor amplitude and RWT concentration as well as between slope (mV/ppb) and temperature. The linear fit of the temperature-dependent calibration curves is described by the following: where C S is the known RWT standard concentration, S(T ) is the slope of the calibration curve at known equilibrium temperature T , V S is the voltage reading directly from the sensor at the known standard concentration and temperature, and V offset is the offset (vertical intercept) of the linear fit. The slope S(T ) of each fit is extracted and plotted against temperature in Fig. 9(b). This plot is also fit to a linear equation that yields the temperature-corrected relationship between sensor output and concentration (inset equation, top right) as described by the following: where S 0 and S 1 are the parameters of the linear fit. Fig. 9(c) shows the offset (intercept) of each fit plotted in Fig. 9(a). Fig. 9(d)-(f) shows the equivalent set of calibration curves corresponding to the AML fluorometer, whereas Fig. 9(g)-(i) corresponds to the Turner fluorometer.
As the parameters of each fit are themselves random variables, the offsets provided in Fig. 9(c), (f), and (i) are subject to variance due to temperature-dependent noise and component parameters (i.e., temperature-dependent spectrum shifts), as well as component nonlinearities, quantization noise, and device alignment deviations in the calibration chamber between calibration trials. To compensate for this variance across temperatures, the mean offsetṼ offset is computed.
The coefficients of these sequential linear fits, along with the mean offsets, are then used to form temperature-corrected equations for each fluorometer, and the generic equation for each is described by the following: where C is the temperature-corrected measurement of RWT concentration in ppb and V raw is the voltage reading directly from the sensor at the sample concentration and temperature. Equation (10) is used to process the subsequent precision, LOD, and deployment data. The temperature dependence of fluorescence is known [7] to follow the exponential relationship described by the following: where F S is the fluorescence of a standard at standard temperature T S , F is the fluorescence of a sample at sample temperature T , and n is the temperature exponent. Notably, the temperature dependence exhibited in Fig. 9 shows good agreement with the literature. For RWT, the temperature exponent is reportedly within the range of −0.027 • C −1 [23] to −0.026 • C −1 [7]. To a first approximation, the temperature exponent can be obtained from the linear fit parameters with the following equation: The approximate temperature exponents are −0.017 • C −1 , −0.016 • C −1 , and −0.017 • C −1 for the prototype, AML device, and Turner device, respectively. These results show remarkable concordance and are within 50% of reference values.
The results obtained for the prototype sensor are also in reasonably good agreement with their Zemax OpticStudio simulations. Qualitatively, both exhibit the expected linear response. Quantitatively, the half-amplitude output of the lock-in amplifier can be scaled by two to peak-to-peak amplitude to compare more directly to the dc-based OpticStudio output (a fair comparison when referred to the excitation LEDs output range). The simulated calibration curve achieves an output voltage of 300 mV at a concentration of 100 nM (approximately 56.7 ppb for RWT). The measured calibration curve achieves an output range between 200-and 320-mV peak-to-peak for the same concentration, depending on temperature. The simulated result is therefore accurate to within a factor of 2.
The quantitative discrepancy between the measured and simulated amplitudes, as well as between measured and referenced temperature exponents, can be attributed to unmodeled confounding factors such as the decaying quantum yield [7] over time in the presence of the sodium and potassium chlorides in seawater. Correcting the model for these minor factors is outside the scope of this work and is not strictly relevant to the calibration nor to the field testing as they demonstrably affect each of the three fluorometers evenly.

B. Precision and Noise
Since the error bars are small relative to the full-scale calibration curves presented in Fig. 9, the standard deviations (SDs) of Fig. 9(a) can be found in Table III, similarly for Fig. 9(d) in Table IV and for Fig. 9(g) in Table V. The SDs are relatively stable and insensitive to both temperature and concentration, in all but a few anomalous points (indicated in red). This finding indicates an absence of heteroskedasticity  III  MEASUREMENT SD, PROTOTYPE CALIBRATION DATA   TABLE IV  MEASUREMENT SD, AML X2CHANGE CALIBRATION DATA   TABLE V MEASUREMENT SD, TURNER CYCLOPS-7 CALIBRATION DATA and therefore justifies the choice of line-of-best-fit linear regression.
The calibrations were performed in a light-blocking chamber, so the solar spectrum does not contribute variance to the measurement data. In integrated CMOS transimpedance devices such as TSL257, thermal noise is known [24] to be the dominant contributor of noise in the amplifier and therefore in the electronic subsystem. The Johnson-Nyquist thermal noise equation for resistance is given in the following equation: where k B is Boltzmann's constant, T is the absolute temperature of the device in kelvin, and R is the resistance, which indicates that per root Hz of bandwidth and the SD due to thermal noise is proportional to the square root of the device's absolute temperature. However, as this proportionality is not meaningfully observed in any of Table III,  Table IV, or Table V, other temperature-sensitive effects (i.e., spectral shifts in the optical source and detector) and temperature-insensitive sources of noise and interference (optical crosstalk and scattering of fluorescent emission) drown out the variance of thermal noise alone. To assess this, assuming that the variation in SD across concentrations is at least approximately normally distributed, Fig. 10 shows chi-squared confidence intervals (CIs) on the prototype variance data Fig. 10. Chi-squared 80% CIs for the prototype measurement variance across concentrations, per temperature. Due to the overlap between CIs, it cannot be said that the measurement variance monotonically increases significantly with temperature.
(SD squared). Even at the 80% confidence level used, all intervals overlap except those with anomalous entries, and those intervals overlap if the anomalous entries are excluded. Therefore, it cannot be said that the variance monotonically increases significantly with temperature. The same analyses are performed for the AML and Turner data and reach the same conclusion but are omitted for brevity.  The prototype calibration curve offsets have a mean,Ṽ offset , of 4.69 mV and a standard error of 0.098 mV or 2% of the mean. This is superior to the AML device withṼ offset of 14.3 mV and a standard error of 1.24 mV or 8.7% of the mean and comparable to the Turner device with aṼ offset of 26.2 mV and a standard error of 0.24 mV or 1% of the mean. This indicates that the prototype has calibration repeatability that is within the range of commercially available fluorometers.

C. Limit of Detection
The LOD and limit of quantification (LOQ) are determined from the noise levels in each system using data collected in the blanking concentration (0 ppm) step of calibration. The literature [25], [26], [27], [28], and regulatory [29] standard "three-sigma" method is used, taking the LOD and LOQ to be three times and ten times the SD of the blanking test, respectively [30]. SDs are converted from volts to ppb by using (10) and settingṼ offset to zero.
Since higher temperatures yield weaker fluorescence for the same concentration, as described by (11), the LOD and LOQ are fundamentally increasing functions of temperature. In simpler terms, to obtain a voltage signal that is strong enough to exceed the LOD or LOQ threshold in voltage terms, the concentration must be higher for higher temperatures.
The LODs at each temperature can be found in Table VI, whereas the LOQs at each temperature can be found in Table VII. These limits are relatively monotonic in temperature as expected as the relatively temperature-insensitive measurement SDs are converted to concentrations using the temperature-compensation calibration equation.
The Turner device has the lowest LOD, in a range of 150%-200% of the specification provided in Table I. The prototype has the highest LOD, though notably, it is also within 150%-200% of the Chelsea TriLux chlorophyll-a LOD of 0.1 ppb, relevant as RWT and chlorophyll-a tend to have similar LODs. This discrepancy can also be attributed to the use of seawater-RWT solutions instead of the usual Milli-Q preparations. The interference from other ions and the degradation of the dye in seawater [7] causes a lower apparent concentration in the standards, which increases the LOD and LOQ.

D. Large Tank Study
The prototype fluorometer was mounted to a BlueROV2 as shown in Fig. 11, along with the AML fluorometer. Both fluorometers are aimed beneath the ROVs ventral side [19].
The sensor integrated ROV was deployed in Dalhousie's Aquatron Laboratory, a 285-m 3 tank of seawater. While the ROV was in the tank, approximately 7 mL of 20% RWT dye (Turner Designs, San Jose, CA, USA) was injected into the tank. The ROV was then arbitrarily driven through the plume and time-series data from the fluorometers were collected. The ROV was then removed from the tank and the tank was cycled . The clustering of points at 6 ppb shows that the sensors return to good agreement after the tank is cycled to achieve a homogenous RWT concentration. At 13:40, the second dose is added, and the sensor readings diverge, but over time (12 min), both sensors settle toward agreement at 12 ppb, with RMSRE dropping to 0.062.
such that the RWT concentration was uniform throughout. The ROV was then reintroduced to the tank and a second dose of 7 mL of 20% RWT was introduced. As before, the ROV was driven through the plume and the time-series data from the fluorometers were collected.
The time-series results of this test are presented in Fig. 12 along with unity curves illustrating the level of agreement between the two sensors. The first dose time series and the unity curve are presented in Fig. 12(a) and (b), respectively, while the second dose time series and the unity curve are presented in Fig. 12(c) and (d), respectively.
A trend is readily observable from the time series; as the ROV is driving through the plume, the sensor readings diverge significantly, whereas in the homogenous solution (i.e., 0 ppb before the first dose, 6 ppb after the first dose is cycled, and 12 ppb after the second dose has had time to disperse), the sensors agree.
This trend is not an unexpected result: the sensors are spaced approximately 11.5 cm apart and each sensor has a sample volume on the order 1 cm 3 . Since the viscous RWT plume can have high concentration gradients and the sensors' sample volumes do not overlap, the concentrations recorded by each sensor can diverge accordingly, although this divergence diminishes over time as the RWT plume disperses and homogenizes throughout the tank. This is supported by the sharply changing individual readings of each sensor beginning at 13:40 in Fig. 12(c) when the second RWT dose was added. Prior to this time point (i.e., 13:36 to 13:40), the first RWT dose had been cycled in the tank and settled to a uniform concentration of 6 ppb. This is represented by the well-correlated cluster of points at 6 ppb in Fig. 12(d).
By the 13:45 time point, the second RWT dose had settled to an approximately uniform concentration of 12 ppb and the two sensors return to excellent agreement, represented by the well-correlated cluster of points at 12 ppb in Fig. 12(d). This second clustering at 12 ppb is to be expected from the result of cycling the tanks between doses. Each dose contained the same titrated volume of 20% RWT, so the total amount of RWT in the tank after the second dose is twice the amount following the first.

E. Field Deployment
The prototype and Turner fluorometers were mounted to the same BlueROV2 as in the Large Tank study and field-deployed in a dye-tracing measurement application during MEOPAR's 2021 TReX [19].
The TReX is an RWT field experiment focused on time of travel and dispersion data collection for modeling and forecasting. The TReX project is cofunded by the MEOPAR and RQM organizations and is a multidisciplinary effort that integrates inputs from government, academia, and local communities.
In addition to the broad scope of dye-tracer experiments and their applications [8], the MEOPAR website for the TReX project [31] lists ten research works in which the TReX data have been used, with applications, including eddy simulation of surface layer mixing, Langmuir circulation surface drift and dispersion, radar-data estimation of dye dispersion, deeplearning contaminant dispersion, and assessment of oceanic drift prediction models. Fig. 13 presents the data acquired on the 5th. Fig. 13(a) shows a plot of two time series: the prototype sensor's fluorometry and the Turner analog sensor's fluorometry. Inset onto the plot are time-lapse images from the ROVs' front-facing camera (top left to top center). The images are captured at the start of each minute. For a sanity check, the images were compared to the time-series data; when the fluorometer data peaked, the camera recorded the vivid orange plumes of the RWT dye. When the fluorometer data flattened, the camera recorded images of green-blue seawater. Fig. 13(b) shows a unity curve of the data presented in Fig. 13(a). Fig. 13(c) presents the time series of depth and ambient temperature recorded by the ROV. Fig. 13(d) shows a map image indicating the ROVs' starting coordinates.
Ultimately, the results of the TReX deployment validate the performance of the prototype as a submersible fluorometer and the prototype achieves good agreement with the Turner Cyclops-7 fluorometer.

IV. CONCLUSION
This work presents a prototype low-cost submersible fluorometer that applies a digital lock-in signal processing scheme to obtain the performance that is competitive with commercial devices such as the Turner Cyclops-7F fluorometer at a fraction of the cost. The device, when configured for Rhodamine fluorescence, can be assembled for U.S. $744.70 (2022), and with appropriate changes to the custom PCB controller, the price can be further reduced and the performance can be further increased. The LOD of the prototype is below 0.2 ppb, which is comparable to commercially available competitors.
The prototype was successfully mounted to a BlueRobotics BlueROV2 along with its representative competitors for a comparative study in the Dalhousie Aquatron Laboratory and an application to the MEOPAR TReX dye-tracer experiment.
While the initial design of our novel fluorometer targets Rhodamine fluorescence, swapping optoelectronics and optical filters with drop-in alternatives would allow it to target other fluorophores with no added complexity. The demonstrated low-cost fluorometer will facilitate improving our spatiotemporal observation capacity of aquatic environments.

APPENDIX
The system architecture block diagram of the custom PCB that has been repurposed to control the prototype fluorometer is presented in Fig. 14. The controller is printed on a six-layer round circuit board with dimensions matched to the inner diameter of the fluorometer's cast acrylic tube housing.
The controller is based on the PIC24F microcontroller. The integrated 16-bit delta-sigma ADC is used to sample the output of the TSL257 photodiode, whereas the integrated 10-bit DAC and OPAMP form a feedback loop to drive the LED525-33 optical source with the lock-in reference sinusoid.
Power to the device is supplied externally through the six-pin SubConn connector. Of the remaining four pins, three are dedicated to RS-232 data I/O and the last is unused. The LED and photodiode are not placed on the PCB itself; they are interfaced using hook-up wires and are mounted to the 3-D-printed optical stack housings, as shown in Fig. 3(a).
The custom PCB is repurposed from a submersible phosphate analyzer [31] and contains several unused components. The unused sensors, connectors, additional LEDs and photodiodes, and motor driver circuits add cost and layout complexity with no benefit. A functionally identical controller PCB can be realized by simply omitting these components from its design.