Deep-Learning-Driven Turbidity Level Classification

Abstract

Accurate turbidity classification is essential for maintaining water quality in various contexts, from drinking water to industrial processes. Traditional turbidimeters face challenges, including interference from colored substances, particle shape and size variations, and the need for regular calibration and maintenance. This paper implements a convolutional neural network (CNN) to classify water samples based on their turbidity levels. The dataset consisted of images captured under controlled laboratory conditions, with turbidity levels measured using a 2100P Portable Turbidimeter. The CNN achieved a classification accuracy of 97.00% in laboratory settings. When tested on real-world water body samples, the model maintained an accuracy of 85.00%. The results demonstrate that deep learning can effectively classify turbidity levels, offering a promising solution to overcome the limitations of traditional methods. The study highlights the potential of CNNs for accurate and efficient turbidity measurement, balancing accuracy with practical applicability in field conditions.

1. Introduction

Water quality is a comprehensive term enclosing various physical, chemical, and biological characteristics of water, determining its suitability for different uses such as drinking [1,2], recreational activities [3], agricultural irrigation [4,5], and industrial processes [6]. Assessing water quality involves measuring a range of parameters to understand the presence and concentration of contaminants and the overall health of the water body [7]. Physical characteristics, such as temperature, color, odor, taste, and turbidity, provide essential information about the conditions within the aquatic environment. Temperature influences aquatic organisms’ metabolic rates and gases’ solubility, while color can reveal the presence of organic material, pollutants, or minerals. Odor and taste often indicate organic compounds, microbial activity, chemical contaminants, and turbidity affecting light penetration and aquatic ecosystems [8,9,10]. Chemical aspects, including pH, dissolved oxygen (DO), nutrients, heavy metals, and organic compounds, are crucial in determining water quality. The pH level indicates the acidity or alkalinity of water, which can influence various biological and chemical processes. DO is vital for the survival of aerobic organisms, and low levels often signal pollution. Nutrients such as nitrogen and phosphorus are essential for plant growth but can lead to eutrophication when present in excess. The presence of heavy metals like lead, mercury, and arsenic, even at low concentrations, poses significant health risks. At the same time, organic compounds, including pesticides, herbicides, and industrial chemicals, can harm humans and wildlife [11,12,13,14]. Finally, some biological factors are critical indicators of water quality. Bacteria and viruses can signify fecal contamination and pose health risks by causing diseases [15]. Algal blooms, often resulting from nutrient pollution, can consume oxygen levels and produce toxins harmful to aquatic life.

The physical characteristic of turbidity is a fluid’s cloudiness or haziness caused by numerous particles that are generally invisible to the naked eye [7]. Turbidity is measured using turbidimeters, which assess the amount of light scattered at a 90-degree angle to the incident light [16]. These particles scatter and absorb light, reducing the transparency of the liquid. The measurement process involves several steps. First, a beam of light, usually from an LED or laser, is directed through the sample with the wavelength chosen according to the size and type of particles expected in the sample. The liquid sample is placed in a transparent container called a cuvette, allowing light to pass through it. As the light moves through the sample, it interacts with the suspended particles, causing scattering, measured by detectors positioned at various angles, typically at 90 degrees to the incident light. The intensity of the scattered light, directly proportional to the turbidity of the sample, is converted by the turbidimeter electronics into a turbidity value, usually expressed in nephelometric turbidity units (NTUs) or formazin nephelometric units (FNU). These steps require knowledge and training in the use of turbidimeters. Besides, turbidity measurements face several challenges. Interference from colored substances in water can absorb incident light, leading to underestimates of turbidity and making accurate readings difficult in colored or stained samples [17]. The assumption of uniform particle size and shape increases complexity, as natural samples often contain varied particles that scatter light differently, causing inconsistencies and inaccuracies in readings [18]. Also, turbidimeters require periodic calibration with standards like formazan and meticulous cuvettes and maintenance of optical components to ensure accuracy, which can be time-consuming and costly [19]. Moreover, these meters have a limited detection range; extremely high particle concentrations can cause multiple scattering events, resulting in nonlinear responses and inaccurate readings at high turbidity levels. External factors such as ambient light, temperature fluctuations, and vibrations can interfere with measurements, necessitating adequate shielding and environmental control to obtain reliable data [20].

Water bodies with sparse plant and animal life have turbidity levels less than 0.1 NTU. Drinking water should also have turbidity levels below 0.1 NTU. Typical groundwater usually has turbidity levels of less than 1 NTU. Water bodies with moderate plant and animal life range from 1 to 10 NTU. Those enriched with nutrients support sizable planktonic life plumes ranging from 10 to 50 NTU. According to [21], during winter storm flows, creeks and rivers can experience turbidity levels between 20 and 1000 NTU.

High precision in turbidity assessment is essential for maintaining healthy aquatic ecosystems and ensuring safe drinking water. Advances in machine learning offer promising solutions to enhance the accuracy and efficiency of turbidity measurements, addressing traditional challenges and limitations in this critical field. For instance, the regression model $R^2$ with RGB sensors for turbidity achieved an accuracy of 91.23% within the range of 0.02 to 60 NTU. However, environmental factors such as temperature, light, and pH can influence sensor performance, introducing measurement variability [22]. This system necessitates meticulous calibration and testing to maintain accuracy, which can be time-consuming and resource-intensive. Contrarily, the recursive neural network (RNN2) model achieves a prediction accuracy of 94.10% for turbidity values obtained from the Qingtan Weir in Taiwan, utilizing underwater sensors and UAV image data [23]. However, UAV operations are susceptible to weather conditions, potentially affecting data acquisition consistency. The gradient boosting decision tree (GBDT) model exhibits an accuracy of 88.00% in the range of 0.94 to 103.43 NTU but requires careful tuning of hyperparameters and a substantial amount of training data to attain high accuracy [24]. The transfer learning method achieved a classification accuracy of 96.86%, with a noted overfitting risk, which could impair performance with new data [25]. The self-organizing multichannel deep-learning system (SMDLS) for river turbidity monitoring, evaluated on a comprehensive dataset from nine locations over a year, enhances accuracy by approximately 14.27% compared to other models based on PSNR, MSE, and NMGE metrics [26]. Nonetheless, its computational efficiency is limited due to the necessity of training seven subnetworks and amalgamating their outputs. Implementing CNNs for turbidity and total suspended solids (TSS) measurements achieved 98.24% accuracy for TSS and 97.20% for turbidity using white light; however, precise control of illumination distance and camera positioning is imperative, which may be challenging in field conditions without specialized equipment [27]. Additionally, a study employing a CNN to accurately classify the turbidity of water samples, involving a relatively small dataset of 200 images, attained an overall accuracy of 97.5% in classifying turbidity from 0 to over 250 NTU, underscoring the need for a larger dataset to enhance the model’s accuracy and generalization [28].

This document describes the outcome of a comprehensive tailored dataset for training a CNN. The dataset is carefully balanced to ensure an even distribution of samples across five distinct Classes. Each image is meticulously chosen to be clear and noise-free, with accurate labeling to provide high-quality training data for the CNN. Every sample is associated with a specific NTU value, precisely measured using a turbidimeter. The innovation of this study lies in its simple classification process, which demands lower computational resources due to the characteristics of the proposed CNN and dataset. In contrast to prior studies, this research collects water samples with NTU values ranging from 200 to 800 and achieves higher accuracy in their classification. The images are captured using an experimental setup to minimize external interference and ensure correct classification, as shown in Figure 1. Although simple, this setup is crucial for achieving reliable results. The classification process achieves an impressive turbidity classification accuracy of 97.00%, demonstrating this approach’s potential for accurately selecting Classes in turbid water applications.

The rest of this paper is structured into five sections. Section 2 details the methodology for evaluating water turbidity using laboratory-prepared samples, encompassing the setup with a 2100P Portable Turbidimeter and Nikon D3300 camera in a controlled environment. Section 3 discusses the rationale behind selecting the CNN for classification, its performance in controlled settings, and its application to real-world aquatic environments. It provides insights into the dataset used, CNN architecture specifics, and performance metrics such as accuracy and confusion matrix analysis across various turbidity Classes, extending the evaluation to natural water bodies. Section 4 synthesizes the study’s findings, critically evaluating the performance, strengths, and limitations of the CNN. It compares the model’s accuracy in controlled versus natural environments and explores its practical implications for field applications. Finally, Section 5 summarizes the study’s results and explores the practical impact of the CNN model’s performance in turbidity classification across diverse environmental conditions.

2. Comprehensive Framework for High-Quality Data Collection

This part of the study evaluates water turbidity by analyzing images of laboratory-prepared samples. Various turbidity levels were intentionally created in the samples to cover various measurements. Each sample was explicitly prepared to achieve certain turbidity levels and was then measured using a 2100P Portable Turbidimeter (HACH Company, Loveland, CO, USA) with a specialized sample cell. The 2100P Portable Turbidimeter was calibrated using standard values of <0.1, 20, 100, and 800 NTU, as recommended by the manufacturer, to ensure accurate and reliable measurements [29]. The samples were classified into five intervals, as indicated in

Table 1. Water turbidity categories and their NTU ranges.

Class Number	NTU Range	Turbidity Level
1	200–320	Low
2	320–440	Moderate
3	440–560	Intermediate
4	560–680	High
5	680–800	Very high

.

A high-resolution digital camera, the Nikon D3300 (Nikon Corporation, Tokyo, Japan) digital camera, equipped with a 35 mm fixed lens, was chosen for its high-resolution capabilities and adjustable settings, ensuring precise capture of water sample images across varying turbidity levels. Employed within a Puluz photo lightbox (PPL), designed to minimize external light interference, diffused LED lighting uniformly illuminated each sample to prevent shadows, glare, and reflections. Standardized camera settings, including a fixed ISO of 100, maintained consistency throughout image capture, with each sample placed in a transparent cylindrical cell at the PPL center.

The images undergo meticulous labeling based on NTU measurements obtained from the Hach 2100P Portable Turbidimeter, categorizing each image into predefined turbidity Classes essential for training deep-learning models. Consistent and accurate labeling across all images prevents bias in the training data, ensuring each image is correctly assigned to its respective turbidity Class and maintains uniform standards. This precise labeling is crucial for training a robust model capable of accurately classifying turbidity levels in new, unseen data. Some images are depicted in

Table 2. Dataset of experimental laboratory samples.

Class 1	Class 2	Class 3	Class 4	Class 5

.

3. Classification Schema for the NTU

Selecting an appropriate model architecture is crucial in developing an effective deep-learning system for turbidity classification, significantly influencing the model’s ability to classify images based on turbidity levels accurately. For this study, a CNN was chosen because it can automatically and adaptively learn spatial hierarchies of features from input images. This capability makes CNNs especially well-suited for handling visual data tasks, as they excel in feature extraction, manage high-dimensional data efficiently, and provide robust performance in image classification, as highlighted in

Table 3. Advantages of CNNs over machine-learning and deep-learning techniques in image classification, based on Le Cun et al. [30].

Aspect	CNN	Machine Learning	Deep Learning
Feature extraction	Automatic, highly effective	Manual, requires domain knowledge	Automatic, effective
Handles high-dimensional data	Excels, especially with images	Not well	Very well
Complexity	High for image tasks	Moderate	High
Training time	Long, but worth it	Short	Long
Performance	Excellent for images	Good	Excellent
Applications	Image/video processing	Various tasks	Wide range
Generalization	Excellent with enough data	Good	Very good
Adaptability	Needs retraining	Easily adaptable	Adaptable
Robustness	High	Moderate	Moderate

. This section describes, first, a preliminary case study designed to rigorously evaluate the effectiveness of the developed deep-learning model in a controlled laboratory environment. This case study involves creating various turbidity levels in laboratory-prepared samples, capturing high-quality images of these samples, and assessing the model’s performance using these controlled data. The objective is to ensure the model can accurately classify turbidity levels when the conditions are tightly regulated, and the data are consistent. The following section explores the practical application of the developed CNN in real-world aquatic environments, emphasizing the transition from laboratory conditions to natural settings. This scenario involves deploying the model in a single aquatic body, such as a lake, and adapting the data collection methods to accommodate the variability and unpredictability of natural environments. The goal is to evaluate the model’s robustness and adaptability, ensuring it can perform accurately, even when faced with the complexities and challenges of real-world data. This comprehensive approach validates the model’s efficacy under controlled conditions and demonstrates its practical utility in real-world scenarios.

3.1. Preliminary Case Study: Controlled Examination

The study utilizes a dataset of images with specific turbidity levels, encompassing five distinct classes ranging from 200 to 800 NTU and containing approximately 700 images, which can be found in [31]. The methodology systematically names each image with the corresponding NTU value provided by the turbidimeter. This systematic labeling ensures an accurate association between the images and their respective turbidity measurements, facilitating precise model training and evaluation. The CNNs have been optimized with Algorithm 1, which describes the steps to optimize the hyperparameters of a general CNN through multiple trials and performance evaluations based on validation data. The process begins with the initialization of the best model and accuracy, followed by iterative sampling of hyperparameters and training of the CNN model for a specified number of epochs. During each trial, the model’s validation accuracy is calculated, and the best-performing model is updated accordingly. This iterative process ensures the selection of an optimized CNN model after all trials are completed.

The CNN implemented in this study was designed to classify water turbidity levels based on image data. The CNN architecture begins with an input layer of 656 × 875 pixels directly extracted from the camera, representing a grayscale image. The first convolutional layer employs 32 filters of size 3 × 3, each applying a rectified linear unit (ReLU) activation function to extract features from the input image. Subsequently, a max pooling layer with a 2 × 2 pool size reduces the spatial dimensions while retaining the most significant features.

Following the initial layer, the network progresses to a second convolutional layer with 64 filters of the same 3 × 3 size and ReLU activation, further enhancing feature extraction. Another max pooling layer with a 2 × 2 pool size follows to continue spatial reduction. The third convolutional layer intensifies feature extraction with 128 filters of 3 × 3 size and ReLU activation, culminating in a third max pooling layer of 2 × 2 size.

Algorithm 1 Iterative hyperparameter optimization for the CNN using multiple trials.

Require: N (number of trials), M (number of epochs), H (hyperparameter space),                 D (training data), V (validation data) Ensure: Optimized CNN model   1:  Initialize best_model ← None   2:  Initialize best_accuracy ← 0   3:  for$i=1$ to N do   4:         Sample hyperparameters $h_i\in H$   5:         Initialize CNN model with hyperparameters $h_i$:   6:               Set model architecture (number of layers, types of layers, etc.)   7:               Set activation functions (ReLU, sigmoid, etc.)   8:               Set regularization methods (dropout rates, L2 regularization, etc.)   9:         Initialize optimizer (e.g., SGD, Adam) with learning rate from $h_i$ 10:         Initialize loss function (e.g., cross-entropy loss) 11:         for $j=1$ to M do 12:              Train model on training data D for one epoch 13:              Evaluate model on validation data V 14:              Compute validation accuracy $a_j$ 15:              if $a_j>\mathrm{best\_accuracy}$ then 16:                    best_accuracy ←$a_j$ 17:                    best_model ← model 18:                 end if 19:         end for 20: end for 21: return best_model

After the convolutional layers, a flattening layer converts the 2D output from the last convolutional layer into a 1D vector, preparing it for input into a fully connected layer. This fully connected layer comprises 100 units with ReLU activation, facilitating high-level feature learning and abstraction. Finally, the output layer, comprising five units corresponding to the five turbidity classes, utilizes an appropriate activation function to classify input images into their respective turbidity levels based on the learned features extracted by the preceding layers. The model was trained on 70% of the dataset (Figure 2) and tested on the remaining 30%, utilizing 30 epochs and the Adam optimizer.

Based on the confusion matrix provided in

Table 4. Confusion matrix analysis of laboratory-controlled examination with the proposed CNN.

	Predicted Class
Real Class	1	2	3	4	5
1	1.00	0.00	0.00	0.00	0.00
2	0.00	1.00	0.00	0.00	0.00
3	0.00	0.05	0.93	0.02	0.00
4	0.00	0.00	0.13	0.85	0.02
5	0.00	0.00	0.00	0.03	0.97

, the described CNN model demonstrates an accuracy of 97 % in laboratory-controlled examinations. Each class shows strong performance, with the model correctly classifying all instances of classes 1 and 2, as indicated by the values of 1.00 on the diagonal. Class 3 is also accurately classified with a high accuracy of 0.93, while class 4 has a slightly lower accuracy at 0.85, and class 5 at 0.97, still indicating robust performance. The instances of misclassification are minimal; for example, there are a few misclassifications between classes 2 and 3 (0.05) and between classes 3 and 4 (0.02). Similarly, class 4 has some instances misclassified as class 3 (0.13) and class 5 (0.02). These misclassifications are relatively minor, suggesting that the model can distinguish between different classes, even though there may be some overlap or similarities in the features of certain classes.

3.2. Application of NTU Classification to Aquatic Body Environments

The proposed CNN in Figure 2 was tested in a water body. In this sense, the Mexican standard that explains how to measure the turbidity of a water body is NMX-AA-038-SCFI-2001 [32]. This standard establishes the method for determining turbidity in natural, residual, and treated residual waters. The samples were taken to establish the standard shown in Figure 3. These samples were collected from different in situ sampling spots within the water body and measured with a turbidimeter to obtain the necessary data for testing the proposed CNN model, the dataset can be found in [31].

According to the described results of Section 3.1, the CNN model demonstrates superior performance compared to the model presented in [28], which also uses a laboratory sample. The developed base data for our model have advantageous characteristics, including good size, composition, Class levels, balance, and dimensionality. These factors contribute to the simplicity of the CNN layers relative to those in [28]. In the water body, 60 samples were collected, all classified as Class 1. The same camera was used for all captures, with standardized settings: an aperture of f/2, ISO 100, and a shutter speed of 1/600 s to eliminate motion blur. Manual white balance was adjusted to align with the LED light color temperature in the PPL, ensuring uniform color across all images.

Based on the characteristics of the water body samples, when the CNN was tested with samples from a natural water body, the accuracy decreased to 85%, despite these samples being within the scope of the original dataset. The details of the CNN classification are summarized in

Table 5. Evaluation of CNN accuracy for aquatic bodies using a confusion matrix.

	Predicted Class
Real Class	1	2
1	50.0	10.0
2	00.0	00.0

. When the actual Class is 1, the model correctly predicted 50.0 samples as Class 1 but incorrectly predicted 10.0 as Class 2. The similar turbidity levels between the upper end of Class 1 and the lower end of Class 2 make it easy to confuse the two Classes, contributing to the observed decrease in accuracy when dealing with natural water body samples. Each sample was placed in a transparent, cylindrical cell at the center of the PPL. The comparison between Class 1 turbidity in the laboratory and in situ samples is illustrated in

Table 6. Comparison of Class 1 turbidity: laboratory samples vs. in situ samples.

Class 1—Laboratory Samples	Class 1—In Situ Samples

.

4. Critical Discussion and Comparative Analysis

This study evaluates the performance of a CNN in classifying water turbidity, comparing its efficacy in controlled laboratory settings and real-world aquatic environments. The findings highlight the robustness of CNNs in controlled conditions and their adaptability to natural settings, albeit with reduced accuracy. The CNN demonstrated exceptional performance in laboratory environments, achieving an overall accuracy of 97% across five distinct turbidity Classes. The confusion matrix revealed high classification capabilities, with perfect accuracy (1.00) for Classes 1 and 2 and high accuracy for Classes 3, 4, and 5, indicating precise feature extraction and pattern recognition abilities with minimal misclassification between adjacent Classes. These results can be analyzed in terms of epoch, accuracy, and computational time. The proposal dataset + CNN reaches higher accuracy in both test and train in a shorter time. At 20 epochs, it achieves a training accuracy of 0.97 and a testing accuracy of 0.93 in just 22 s, shown in

Table 7. Comparative performance analysis of CNN models on different datasets for turbidity classification.

Proposal Dataset + CNN				Dataset + CNN from [28]
Epoch	Time [s]	Accuracy Train	Accuracy Test	Epoch	Time [s]	Accuracy Train	Accuracy Test
20	22	0.97	0.93	-	-	-	-
30	42	1.00	0.97	-	-	-	-
50	40	1.00	0.93	50	59	0.88	0.87
100	60	1.00	0.96	100	119	0.92	0.90

. Conversely, when tested with samples from natural water bodies, the CNN’s accuracy decreased to 85%. Although lower than in the laboratory, this accuracy remains acceptable for practical applications. Samples from natural water bodies, all classified as Class 1, were collected and imaged under standardized conditions to ensure consistency. Despite these efforts, the model misclassified ten samples, indicating challenges in adapting to real-world variations.

The CNN’s performance in real-world conditions remains competitive compared to other models in the literature. Ref. [22] utilized a test bench and calibration method for classification, achieving 91.23% accuracy with a fine K-nearest-neighbor classifier, highlighting the importance of controlled experimental conditions for reliable results. Ref. [23] implemented a more complex classification method integrating regression and RNN2 to manage different weather conditions, resulting in RMSE values of 20.89 during training and 36.11 during testing, alongside high R-squared values of 99.30% for training and 94.10% for testing, indicating substantial predictive accuracy despite inherent complexity and computational demands. Ref. [24] focused on data acquisition using Sentinel-2 satellite images, ensuring image quality through pre-processing steps to remove clouds and correct atmospheric effects. This enabled accurate spectral reflectance data collection, which is essential for reliable turbidity estimation. On the other hand, Ref. [26] relied on a large river turbidity monitoring dataset comprising 11,681 data points from nine fixed locations over a year, improving accuracy by 14.27% over other models, though facing computational inefficiency due to the sheer volume of data and processing requirements. Finally, in [27], a dataset using liquid samples illuminated with LED lights and recorded with a smartphone camera achieved 97.20% accuracy using a CNN architecture (AlexNet), benefiting from diverse lighting conditions and high-quality image capture. In contrast, the proposed CNN achieves an accuracy of 97.00% with a simple classification process and a carefully balanced dataset featuring specific NTU values ranging from 200 to 800, emphasizing lower computational resource demands while maintaining high accuracy. The controlled experimental setup minimized external interference, underscoring the potential for efficient and accurate turbidity classification in practical applications.

Given the above analysis, this study demonstrates the proposed CNN’s effectiveness and practicality across various settings, underscoring its competitive advantage and potential for real-world applications in turbidity estimation.

5. Conclusions

The proposed CNN demonstrates competitive performance compared to other models in the literature, such as the regression model with RGB sensors, RNN2, GBDT, transfer learning method, and SMDLS, while maintaining high accuracy and practical applicability. Although some models exhibit higher accuracy rates, they often require extensive calibration, precise control of environmental conditions, or significant computational resources. The methodology employed to construct the dataset of turbidity measurements enabled the CNN to achieve high accuracy rates, with a notable 97% accuracy for laboratory measures and 85% for real-world samples. This rigorous dataset construction is essential to the CNN’s success. Further, the proposed CNN’s simplicity and lower computational requirements make it a viable option for field applications, particularly in less controlled environmental conditions. However, further tuning and adaptation are necessary to enhance its robustness against environmental variations. Overall, the proposed CNN offers a promising solution for turbidity measurement in real-world conditions, effectively balancing accuracy and practicality.