Turbidivision: a machine vision application for estimating turbidity from underwater images

Introduction

In fields ranging from environmental science to public health, assessing water quality is vital. These assessments of water quality often begin with measuring turbidity, which can impact water clarity (Davies-Colley & Smith, 2001) and potability (LeChevallier, Evans & Seidler, 1981). Turbidity is an optical measure of water clarity, measured by the scattering of light by particles suspended in water, contributing to a murky or cloudy appearance. In nature, many of these particles are agitated sediment, such as clays and soils or suspended organic matter (e.g., plant debris or microscopic organisms) and largely depend on surrounding land use (Moreno Madriñán et al., 2012). Pollutants contained in industrial and agricultural runoff can also be linked to turbidity (Rügner et al., 2013; World Health Organization, 2017).

Increasing turbidity can have negative impacts on aquatic life, since it often shares an inverse correlation with dissolved oxygen levels, creating an inhospitable environment for many taxa (Talke, de Swart & De Jonge, 2009). In addition, increasing turbidity can cause nonlethal effects on aquatic taxa through altered predator-prey interactions (Abrahams & Kattenfeld, 1997; Ferrari, Lysak & Chivers, 2010), decreasing light penetration and reduced photosynthesis (Moore, Wetzel & Orth, 1997), and physical stress such as gill deformities (Lowe, Morrison & Taylor, 2015). Beyond affecting aquatic ecosystems and taxa, overly turbid water is unsuitable for consumption by humans (Muoio et al., 2020) and livestock (Umar et al., 2014), as well as acting as an indicator of bacterial contamination (Gharibi et al., 2012). For water treatment plants, this necessitates filtering suspended particles to bring turbidity within acceptable levels of 5 Nephelometric Turbidity Units (NTU) or lower, depending on use and filtration method (World Health Organization, 2017). Furthermore, monitoring turbidity can be helpful in tracking sediment runoff and pollution in bodies of water, providing valuable insights for environmental management and conservation efforts (Owens et al., 2005).

Typically, turbidity is measured with a turbidimeter which shines a source of light–either white or infrared–into the fluid and a probe which measures the resulting scatter of light. Meters that measure in Formazin Nephelometric Units (FNU) shine an infrared light into the solution and measure the scatter at a 90-degree angle of incidence. Other common units for turbidity include NTU, which are measured with white light at a 90-degree angle of incidence, and Formazin Attenuation Units (FAU), which are measured with infrared light at a 180-degree angle of incidence (Anderson, 2005). Agency standards are typically based on method of measurement rather than accuracy thresholds. For example, U.S. Environmental Protection Agency standards require 90-degree hatchure and visible radiation, with equipment tested by the U.S. Geological Survey (USGS) having 5% error or less. These results are similar to the International Standards Organization (ISO) 7027 which require a back-scatter angle of 90 degrees. Testing by USGS found these instruments to also have less than 5% error at turbidity above 40 NTU and greater than 10% below 20 NTU (Wilde & Gibs, 2008).

Turbidimeters are generally expensive, costing hundreds or even thousands of dollars. For example, the LaMotte 2020i turbidimeter used during this project has an MSRP of $1,449 USD (https://lamotte.com/2020i-portable-turbidity-meter#specifications; LaMotte Company, Chesterton, MD, USA). This expense can be a significant barrier to citizen science initiatives, or laboratories with a limited budget, decreasing the ability of laypeople to contribute to water quality monitoring efforts. The converse can also be true, that lowering costs and creating easy and consistent methods of data collection increase the quality of data and participation in data collection (Zheng et al., 2018; Lee, Lee & Bell, 2020). Measuring turbidity with a turbidimeter can also be time-consuming, as they necessitate the transport of the meter itself, or the collection of individual water samples for later measurement.

With the expansion and increasing accessibility of machine vision models their use for water quality has become an emerging field of research and offers potential to reduce time, costs, and increase accessibility of data collection. Current approaches include machine vision in conjunction with existing analytical tools (Yan et al., 2024), model development and image analysis from controlled environments (Nazemi Ashani et al., 2024), and remote sensing (Leeuw & Boss, 2018). The focus of machine vision research for water quality has largely been in economically important fields such as aquaculture (Li & Du, 2022) and wastewater treatment (Mullins et al., 2018). However, there has been little work with model development from in situ images for specific water quality parameters, such as turbidity. Furthermore, while historical underwater image datasets exist (e.g., Peng, Zhu & Bian, 2023) we have been unable to identify any historical datasets of underwater images that include associated turbidity measurements, making it impossible to retroactively assess water quality for historical datasets.

Our primary goal was to develop a machine vision model capable of estimating turbidity from underwater images that could be made publicly available and easily accessible. Our model offers a cost-effective alternative to traditional turbidimeters, making water quality monitoring more accessible for those who may not have the necessary funds to purchase expensive equipment. Given the affordability and widespread availability of waterproof digital cameras, including smartphones, this approach has the potential to democratize non-critical water quality assessment (e.g., Zheng et al., 2022). Using images can simplify the process for field sampling by requiring less equipment and eliminate sample processing. It also allows existing underwater image data without turbidity readings to be retroactively analyzed. We also evaluate the feasibility and accuracy of machine vision for predicting turbidity levels in water bodies. By assessing how well the model performs, this initiative could pave the way for innovative applications in research and citizen science, as well as gaining insights into historical data, fostering greater engagement and participation in environmental monitoring and conservation efforts.

Materials and Methods

Results

The YOLOv8 training process generated a confusion matrix to visualize the performance of the model when inference was run on the validation set. An ideal model would have a perfect correlation, where each predicted class matches the actual class. For our model, predictions from the validation set were tightly clustered around the true classes, with 100% of predictions within one class of the expected class, and 84% of predictions matching the expected class (Fig. 1).

By comparison, the regression had an R² of 0.975, with the parameters ( $x$) and coefficients ( $\mathrm{\beta }$) of the regression equation of form $y={\mathrm{\beta }}_0+{\mathrm{\beta }}_1{x}_1+\dots +{\mathrm{\beta }}_{11}{x}_{11}$ shown in Table 1, with additional regression metrics in Table 2. While the accuracy of the regression model is not significantly greater than the classification model, the numerical, rather than categorical, estimates provide a more broadly useful result for further analysis (Fig. 2). Given that our model predicts turbidity values in a range from 0 FNU to 55 FNU, assessing its accuracy based on the root mean square error (RMSE) does not make sense, as an error of approximately 2 FNU at the high end of the range is considerably less impactful than an error of 2 FNU at the low end. This is why relative root mean square error (RRMSE) and relative standard deviation (RSD) were calculated. While there is no accepted rule for RRMSE or RSD, the lower it is the more accurate the model is, and our 18.46% could generally be considered good. We found that from 0 to 2.5 FNU, 95% of our model’s predictions fell within ±0.7 FNU of the true values, and that from 2.5 to 55 FNU, 95% of our model’s predictions fell within ±33% of the true values. In addition, there was no change in model accuracy when trained based on metadata, such as water body type or Secchi disk presence allowing us to use all images together in the final dataset. We also visually compared subsets of the predicted and actual values against image appearance to confirm there were no patterns between image background and accuracy (Fig. 3).

Discussion

In this study, we evaluated the ability of a machine vision model to estimate turbidity and the accuracy of that model. While the accuracy of the model is lower than a physical meter, it is comparable to field test kits typically used in citizen science programs, free, and a more accessible way to measure turbidity for those willing to accept the reduced accuracy compared to a traditional turbidimeter as anyone with a smartphone can use the Android or web app. In addition, this model is accurate enough to provide insights on turbidity conditions from historical data-underwater image datasets that never had turbidity recorded alongside the images-which would otherwise be impossible to go back and obtain. As an open-source program, the model can also be integrated into other systems which could be used for any variety of other potential data processing tasks related to underwater imagery. Even though the linear regression model is no more accurate than the classification, a single numerical value with a known confidence level is often more useful for analysis.

Future directions

One potential improvement for future research could be changing the underlying model. We are predicting a continuous variable (turbidity), but instead of directly predicting a continuous variable, we first predicted a discrete variable, a class, and then used that to run a secondary regression model. If we were instead to create a neural network that ends in a fully connected layer which outputs a single continuous variable, from 0 to 55, we could effectively run regression directly on the images. However, such a process would require much more data and the creation of a custom convolutional neural network. The exact amount of data needed to train such a model would depend upon the architecture of the model itself, along with the desired accuracy. In addition, since regression is continuous, and not based on classification, direct model comparison is not possible. However, if we assume that quantizing turbidimeter measurements based on the reported accuracy of the meter provides an adequate analogue to number of classes, we can then calculate a minimum recommended amount of images needed for training such a regression model.

If we consider a goal to match the accuracy of the LaMotte 2020i turbidimeter, which is among the most accurate available (meters range from 2–5%; Wilde & Gibs, 2008), then the ideal would be within ± 0.05 FNU between 0–2.5 FNU (LaMotte Company, Chesterton, MD, USA). For this we can say that there are approximately 25 unique ranges of distinct accuracy within this range (2.5/0.1); and since it has a reported accuracy of ± 2% from 2.5–100 FNU, we can split it into 79 unique ranges of distinct accuracy between 2.5 and 55 FNU ( $2.5\times {1.04}^{count}>55$, solved for count). The sum of these two values, 104, gives us a rudimentary analogue to the number of classes needed. If we assume such a regression model would require a similar amount of data as a classification model and a recommended minimum of 150 images per class (Shahinfar, Meek & Falzon, 2020), we can use our analogue class count multiplied by 150 pictures per class to estimate an absolute minimum of 15,600 images, ideally with the turbidity values of these images evenly distributed between 0 and 55 FNU. Such a model would not be able to make use of the development speed increases of transfer learning, as all weights would need to be trained from scratch. This training would require a larger dataset, potentially larger than mentioned above, as well as more time required for model training. If such a model were developed with a similar architecture to Yolov8, processing times for the end user should be comparable to the model created in this project.

Within the existing modelling framework, more images, but fewer than those needed for a custom convolutional neural network, would improve the training process and should be able to achieve greater accuracy. More images would allow the model to learn the features of each class better, and a greater variety of images, such as in location, water quality, and lighting conditions, should help the model gain more resilience against features not related to turbidity. A larger dataset could also allow for the creation of more classes (smaller bins), which would allow the model to make more granular predictions. Another potential approach would be to train our model (or a similar model) to above-water images and assess accuracy. If a model could be trained on above-water images, even if more training is required, it would increase accessibility and the applicability to historical images dramatically.

Conclusions

Measuring turbidity can be critically important for contexts ranging from human health to food webs. Our goal was to determine efficacy, accuracy, and precision of a machine vision model for measuring turbidity from underwater images. We successfully demonstrated the potential of modern machine vision techniques as viable for estimating the turbidity of natural bodies of water, with an accuracy comparable to commercial test kits up to 55 FNU and down to near 0 FNU (e.g., LaMotte Turbidity Test Kit #7519-01). The model we created offers an accessible alternative to traditional turbidimeters and can provide turbidity measurements within an acceptable margin of error in many applications.

The application developed as part of this research project is also the first photo-based turbidity measuring tool accessible to the public. To make the model widely accessible, we implemented it as a free, user-friendly web application that would enable anyone with a modern web-enabled device to make use of the tool. Additionally, the web application is designed to allow the data to be processed on the user’s device, keeping their data secure and avoiding unnecessary use of internet bandwidth by preventing the need to upload their files to a server. The app is compatible with a wide range of devices and has a simple user-interface, allowing anyone to easily benefit from the results of this research. In addition, its use of images as an input allows users to retroactively gain insights on turbidity from historical underwater image datasets and understand past trends.