IoT‐Based Monitoring and Prediction of Water Quality Using Swin‐Transformer and Depthwise Separable Convolutional Neural Network Optimized by Circulatory System‐Based Optimization

ABSTRACT

The degradation of water quality, one of the major global environmental challenges, demands advanced and intelligent monitoring and prediction systems. This study uses data from the Rohri Canal in the Pakistani city of SBA to develop a combined IoT‐based advanced deep learning (DL) model that can accurately forecast the water quality index (WQI) and water quality class (WQC). The sensors that are used under IoT are turbidity, temperature, pH, and total dissolved solids (TDS) to measure real‐time parameters of water quality. All the acquired data are preprocessed with the help of a quality‐aware fuzzy min–max neural network (QAFMMN) to reduce noise efficiently and increase the quality of the data. A combination of the addax optimization algorithm and the Nadam optimizer (AOA‐NO) is used to select features, keeping only the most important ones, which makes the method computationally efficient. In this regard, the model known as the Swin‐Transformer–depthwise separable convolutional neural network, trained with the circulatory system‐based optimization method (ST‐DSCNN‐CSBO), is established. Indeed, one can see the incredible results of the proposed experimental research because the R ² score obtained therein equals 0.999, the mean absolute error (MAE) equals 0.008, and the overall water quality classification accuracy is equal to 99.96%.

Summary

• IoT sensors enable real‐time monitoring of turbidity, pH, temperature, and TDS.

• QAFMMN preprocessing improves data quality by removing noise in sensor data.

• AOA‐NO feature selection reduces complexity and improves prediction efficiency.

• ST‐DSCNN‐CSBO model accurately predicts WQI and classifies water quality.

• Framework supports reliable smart water monitoring for environmental management.

This study uses data from the Rohri Canal in the Pakistani city of SBA to develop a combined IoT‐based advanced deep learning (DL) model that can accurately forecast the water quality index (WQI) and water quality class (WQC). The sensors that are used under IoT are turbidity, temperature, pH, and total dissolved solids (TDS) to measure real‐time parameters of water quality.

1. Introduction

Because clean and safe water is essential for agriculture, human health, and the sustainable growth of all ecosystems, it is considered one of the fundamental human necessities. However, population increase, industrialization, urbanization, and climate change all contributed to significant water damage on the planet (Hemdan et al. 2023; El‐Shafeiy et al. 2023; Shams et al. 2024). Water pollution caused by agricultural runoffs, industrial emissions, and domestic wastewater constantly poses the risk of unsafe usage of these resources because their purity remains questionable. In such a way, modern challenges to maintain drinking water supply and protect aquatic life appear to be complex, and the problem of water quality assessment emerges as one of the key ecological and health concerns (Adeleke et al. 2023; Bi et al. 2023).

Water bodies are also dynamic, which further complicates the exercise of monitoring. Some of the factors that cause the change in turbidity, temperature, pH, and TDS, among other water quality variables, include seasonal changes, rainfall, sedimentation, and biological activities. Such variations should be monitored constantly to obtain suitable and up‐to‐date information (Jeong et al. 2024; Chen et al. 2023; Gao et al. 2023). The conventional approaches to monitoring can involve manual sampling, laboratory tasks, and periodic reporting, which are laborious and costly and would not be capable of providing real‐time data. Therefore, the systems that would constantly, efficiently, and consistently monitor the water quality are highly demanded, as well as the mechanism to make decisions on time (T. Yan et al. 2023; Ni et al. 2023).

In the recent past, technological advancements have made it possible to use real‐time monitoring systems that can collect more detailed information on various water quality parameters (WQPs). They can be applied in providing an overview of the chemical, biological, and physical parameters of water bodies by enabling continuous water collection of the waters (Liu et al. 2023; Georgescu et al. 2023). Patterns and trends can also be observed that would not be identified using the traditional methods because of the availability of high‐frequency data. This type of real‐time information is vital in the initial detection of pollutants, deviations, and timely intervention to prevent potential environmental and health risks.

In addition to being relevant to human water consumption, water quality monitoring is crucial for agriculture, fisheries, aquaculture, and ecosystem management. Water is preserved in the highest quality that facilitates the irrigation of crops, fish farming, and the conservation of aquatic life. With the help of implementing modern monitoring systems into water management practices, it is possible to create improved resource allocation, enhanced environmental planning, and proactive conservation strategies. These systems enable interested parties, policymakers, and environmental managers to make sound decisions by ensuring that the information about the state of the water is provided in an accurate and continuous manner, hence protecting the water resources and ensuring that the water ecosystems are managed sustainably to benefit future generations (M. Yan and Wang 2024; Leggesse et al. 2023; X. Wang, Li, et al. 2023). Current IoT‐driven monitoring of water quality systems tends to have great computational demands, reliance on extensive data sets, the inability to process noisy sensor data, and real‐time implementation and scaling issues. This paper will examine these difficulties in this context.

1.1. Novelty and Contribution

The innovations and contributions of this study are listed as follows:

• An IoT‐driven DL pipeline is introduced by the proposed structure that integrates advanced preprocessing, optimized feature selection, and hybrid neural architectures to achieve highly reliable water quality prediction through water quality index (WQI) and water quality class (WQC) estimation.

• The preprocessing stage using quality‐aware fuzzy min–max neural network (QAFMMN) generates noise‐resilient, quality‐aware feature boundaries, enabling stable learning even under irregular sensor variations and dynamic environmental disturbances present in real‐world aquatic monitoring.

• In the feature selection mechanism, addax optimization algorithm is combined with the Nadam optimizer (AOA‐NO), where addax‐based search behavior is used with Nadam‐based gradient refinement to ensure the optimal attribute reduction without any significant determinant being dropped when assessing water quality.

• The Swin‐Transformer–depthwise separable convolutional neural network (ST‐DSCNN) with hybrid deep learning architecture boosts multiscale feature gathering as it combines transformer‐based attention with lightweight depthwise convolution, which has enhanced the representation of nonlinear temporal–spatial reliance across WQP.

• The CSBO‐driven bias optimization of ST‐DSCNN offers a circulatory flow‐based adaptive search optimization that enhances convergence consistency and boosts generalization in different sensor sets and also in different aquatic environments.

• The Swin‐Transformer–depthwise separable convolutional neural network with circulatory system‐based optimization (ST‐DSCNN‐CSBO) pipeline takes raw streams of the IoT sensors and applies robust preprocessing, selective dimensionality reduction, deep learning of features, and optimized prediction to robustly track and evaluate the water's quality in real time.

The manuscript is divided into five main portions. An overview and background of the study are given in Section 1. Current approaches and related work are presented in Section 2. The aspects of the recommended strategy are described in Section 3. The experimental results and their interpretation are given in Section 4. Lastly, a summary of the results and a study conclusion are presented in Section 5.

2. Literature Survey

Several research studies have been directed toward using the IoT data to predict and track the water quality. In this section, several of them are discussed.

Z. Li et al. (2024) developed gated graph neural network (GGNN) in the publication Real‐Time Quality of Water Prediction, which combined hydraulic flow directions and water quality data for simulation. The learned model was considered part of the sites being monitored and could make reasonable predictions of the entire distribution network. This approach was sufficient but had limitations based on the fact that it depends on the optimal positioning of sensors, which was the most significant limitation.

According to Arepalli and Naik (2024), there was an enhanced DSTCNN algorithm using data collected by the Internet of Things regarding water quality, classified according to WQI standards into fish‐friendly water or hazardous water. The model was effective in modeling complex spatial and temporal correlations through a hybrid activation function and dilated convolutions and was capable of making a better analysis of dynamic water patterns. This approach had improved predictive performance and had a high requirement for computational resources, which is a disadvantage of the suggested approach.

Qiao et al. (2024) introduced a spatiotemporal prediction framework, which implied the fusion spatiotemporal graph convolution neural network (FSGCN) that deals with nonlinear water quality sequence processing with temporal attention, spatial correlations in river networks via graph convolution, and temporal residual learning to enhance long‐term forecasting. It was found to have more predictive potential based on real‐world data, but the density and uniformity of sensor dispersion remained a limitation to the method and an issue with the suggested method.

Wiryasaputra et al. (2024) introduced a decision tree–based IoT‐based water quality monitoring and prediction system. The temperature, acidity, and contamination levels of the water are monitored using the sensors, which are connected to each other. The readings are processed and sent to the cloud server for visualization purposes, along with the Arduino controller. The method improved monitoring by efficiency but was still restricted by dependency on sensors and networks.

Tian et al. (2024) presented the deep temporal convolutional network‐gated recurrent unit (DeepTCN‐GRU) convolutional network and trained it to predict the most prominent water quality indicators, relying on the data of the water source area of the Kaifeng Yellow River. As the experiment revealed, preprocessing worked well on the dataset improvement, and the hybrid architecture worked well on predictive stability during different conditions and water sources. The description focused on the performance consistency, but the model was still constrained by the reliance on the optimal choice of lag values.

Bi et al. (2024) introduced a hybrid prediction model named variational mode decomposition–bidirectional attention encoder–decoder (VBAED) for complex water quality sequences, bidirectional input attention, temporal representations obtained by means of a BiLSTM encoder, and improved predictions by a BiLSTM decoder with temporal attention. The method was useful in modeling nonlinear dynamics in real‐time water quality prediction; however, it was constrained by its high computational intensity.

A data preprocessing unit, a feature improvement unit, and a data prediction unit comprise Xu et al.'s (2024) DL model for forecasting water quality. Wavelet transformation technology was employed to decompose complicated water quality data and environmental data, whereas the informer encoder helped enhance the time series feature encoding process. For water quality prediction, SBiLSTM proved to be useful. One of the weaknesses of the method was the high computational complexity because of its complex structure.

Z. Wang, Wang, and Wu (2023) presented a thorough weight‐based feature selection technique that used the Pearson correlation coefficient with entropy weighting to identify significant features for water quality prediction. The method that measured the significance of features in recognition considered the information content and correlation value, which enhanced the quality and strength of prediction. Different machine learning models were also experimented with, and LSTM proved to be outstanding at working with time series data. Nevertheless, the approach was limited due to the quality of the data and the preprocessing. Table 1 compares the existing approaches.

TABLE 1.

An examination of the current methods.

Authors' names	Methods	Advantage	Disadvantage
Z. Li et al. (2024)	GGNN	Accurately predicts water quality at unmonitored sites	Performance depends on optimal sensor placement accuracy
Arepalli and Naik (2024)	DSTCNN	Improved spatiotemporal feature extraction and robust classification	Higher model complexity increases computational resource demand
Qiao et al. (2024)	FSGCN	Improved long‐term accuracy with strong spatiotemporal learning	Model performance depends heavily on sensor distribution
Wiryasaputra et al. (2024)	Decision tree (DT)	Provides continuous monitoring with efficient drinkability prediction	Dependent on network stability and sensor consistency
Tian et al. (2024)	DeepTCN‐GRU	Enhanced accuracy through combined temporal feature learning	Performance depends heavily on lag value selection
Bi et al. (2024)	VBAED	Improved stability and accuracy for nonlinear water data	High computational cost for large‐scale forecasting
Xu et al. (2024)	SBiLSTM	Accurate nonlinear prediction with enhanced feature extraction	Complex structure increases computational requirements
Z. Wang, Wang, and Wu (2023)	LSTM	Improves accuracy by selecting relevant correlated features	Performance depends on data quality and preprocessing
Geetha et al. (2024)	CNN–BiLSTM	Integrates sensor and image data for real‐time river monitoring	Limited to a single river dataset
Kumar and Vellaichamy (2026)	Attention‐quantized GRU	High accuracy with reduced computation	Focused only on aquaculture ponds
Huan et al. (2025)	WD‐LSTM‐Transformer	Multisource data improves prediction accuracy	Complex model with high computational cost

A mixed CNN–BiLSTM network was used by Geetha et al. (2024) to assess the Kaveri River's water quality. Images and sensor readings were captured on‐site, and their data were collected in real‐time in the cloud environment. The promise of DL for sustainable river water resource management is demonstrated by the suggested model's increased accuracy of 4.62% when identifying impurities.

Kumar and Vellaichamy (2026) suggested an attention‐quantized GRU (AQGRU) to classify water quality in aquaculture in real time. The strategy combines the attention mechanisms to emphasize important temporal patterns and the quantization to minimize the cost of computing. The experimental results were 99.2% accurate, which is better than AODEGRU, DSTCNN, and LSTM models, and thus, easy monitoring and management of the aquaculture environment is realized.

Huan et al. (2025) developed a novel WD‐LSTM‐Transformer architecture for forecasting the total phosphorus and nitrogen concentrations in the Beijing‐Hangzhou Canal. The use of wavelet denoising improved the quality of the data; LSTM and Transformer architectures enabled learning the time dependencies and nonlinear relationships. SHAP analysis ensured better interpretability of the findings, and the performance was excellent, with an R ² above 0.9 achieved, representing a huge improvement in prediction accuracy.

2.1. Problem Statement

The current approaches to water quality prediction frequently fail to manage noisy sensor data, redundant features, poor generalization, and poor ability to model complicated temporal–spatial complexities. Numerous models rely on traditional learning methods that are not able to keep the necessary accuracy in changing environmental conditions and massive IoT streams of data. The proposed ST‐DSCNN‐CSBO framework provides the resolution to these restrictions by combining quality‐conscious preprocessing, intelligent feature removal, transformer‐based DL, and circulatory‐inspired bias optimization. This aggregate yields increased strength to resistance to noise, strength of feature relevance, enhancement of multiple‐scale representation, and stability of convergence, which makes highly reliable and flexible WQI and WQC prediction in real‐world IoT‐based monitoring settings.

3. Proposed Methodology

The suggested ST‐DSCNN‐CSBO framework starts with the IoT‐based data collection, where temperature, pH, turbidity, and TDS values are obtained on a continuous basis through distributed sensors. The resulting data is sent to the cloud and is preprocessed using QAFMMN to clean up any discrepancies and noisy data and normalize the value of the data to maintain the integrity of the data. The resulting processed data is then put through the feature selection step, where AOA‐NO identifies attributes that produce the greatest effect by balancing predictive relevance and dimensional efficiency. The DL phase is aimed at these features and is motivated by the ST‐DSCNN that detects hierarchical spatial–temporal patterns and predicts the WQI and the WQC. To further guarantee the stability of the prediction, ST‐DSCNN bias parameters are optimized using the circulatory system‐based optimization (CSBO) that simulates the circulatory dynamics to provide the model with the optimal bias configuration. The outcome is high‐quality data preparation, dimensional reduction, deep hierarchical feature learning, and an adaptive bias tuning workflow to enable accurate and reliable water quality monitoring and prediction by IoT means. The proposed approach is explained schematically in Figure 1.

FIGURE 1.

Illustration of the suggested methodology.

3.1. Input Acquisition

Input acquisition involved collecting real‐time turbidity, temperature, pH, and TDS data through IoT sensors, with wireless transmission to the base station for storage and processing.

The sensing module and coordinator module are explained as follows.

• •Sensing module
  The sensing module in the proposed study is responsible for collecting real‐time WQPs, which serve as the input for subsequent prediction and analysis. It is made up of sensors placed in water bodies to measure turbidity, temperature, pH, and TDS.

  • ○
    Temperature sensor: This gauges the water's temperature to examine thermal characteristics that affect chemical processes and aquatic life.
  • ○
    pH sensor: It determines the degree of chemical equilibrium and water suitability by measuring the acidity or alkalinity of the water.
  • ○
    Turbidity sensor: It gauges the water's turbidity, or the quantity of suspended particles that could be contaminants.
  • ○
    TDS sensor: It allows one to evaluate the water's quality and usability by measuring the total amount of all dissolved solids, such as salt and mineral content.

• •
  Coordinator module

  • ○
    It controls the process of data acquisition and transfer from the sensing module to the processing module by using an Arduino controller along with LoRa wireless technology.
  • ○
    Arduino controller: It captures data from the sensors, analyzes it, and sends it to the cloud.
  • ○
    LoRa: In order to transmit real‐time data from far locations to the module in charge of data processing, it offers wireless communication capabilities with a greater range and reduced power consumption.
    This module makes sure that data continues flowing continuously and reliably. The second step that follows the collection of data is preprocessing, where raw sensor data will be cleansed, normalized, and made ready to use in a feature selection and prediction process.

3.2. Preprocessing Using QAFMMN

Preprocessing contribution to the water quality detection based on IoT is the raw sensor data of turbidity, temperature, pH, and TDS sensors. These measurements are converted into patterned forms by the preprocessing phase that can be applied in the QAFMMN (Hou et al. 2024).

The edge‐to‐edge (E2E) convolution derives the connection between WQPs based on the adjacency matrix of dynamic correlation. It is defined as Equation (1):

$$E_{p,q}^{\left(a\right)}\left(d\right)=\text{Re}\mathit{LU}\left({\sum }_{u=1}^{\left|\Omega \right|}\left(h_u^m{X}_{p,u}^s\left(d\right)+g_u^m{X}_{u,q}^s\left(d\right)\right)+y^a\right)$$ (1)

where $E_{p,q}^{\left(a\right)}$ is the feature of the ath filter between sensors $p$ and $q$ at time window $d$, $X^s$ is the adjacency matrix representing correlations between sensors, $h_u^m$ and $g_u^m$ are learning weights, $y^a$ is the bias, and $\left|\mathrm{\Omega }\right|=4$ is the number of WQPs. Equation (1) formulates weighted interactions between all sensor pairs.

The edge‐to‐node (E2N) convolution is an edge‐wise aggregation of edge features into node features in each parameter. It is defined as Equation (2):

$$x_p^{\left(a\right)}\left(d\right)=\text{Re}\mathit{LU}\left({\sum }_{n=1}^N{\sum }_{u=1}^{\left|\Omega \right|}{h}_u^{n,a}{E}_{p,u}^n\left(d\right)+y^a\right)$$ (2)

where $x_p^{\left(a\right)}\left(d\right)$ stands for the aggregate attribute of sensor $p$ at time $d$, $E_{p,u}^n$ stands for the edge attribute computed by Equation (1), $h_u^{n,a}$ is the weight factor used in training nth inputs and ath outputs, and $N$ indicates the number of edge attributes.

At last, fuzzy hyperboxes are created to represent the uncertainty and fuzziness of the preprocessed IoT data. It is defined as Equation (3):

$$E{Y}_q^d=\left\{F_m^d,{\mathit{YD}}_q^d,{\mathit{YW}}_q^d,\mu \left(F_m^d,E{Y}_q^d\right)\right\}$$ (3)

where $E{Y}_q^d$ is the qth hyperbox for view $d$, $F_m^d$ is the input pattern for sample $m$ (temperature, pH, turbidity, and TDS) at view $d$, ${\mathit{YD}}_q^d$ and ${\mathit{YW}}_q^d$ are the minimum and maximum bounds of the hyperbox, and $\mu \left(F_m^d,E{Y}_q^d\right)$ is the membership function measuring the fit of the sample to the hyperbox. Equation (3) reduces ambiguity in the IoT measurements and prepares reliable patterns for feature selection.

Therefore, the preprocessed IoT‐based water quality data, such as edge features, node features, and fuzzy hyperboxes, are submitted to the module of feature selection as input.

3.3. Water Quality Assessment Using WQI and WQC

• WQI

  Using a mixture of various water quality measures, WQI provides a single value to describe the overall water quality. This study uses four parameters, which include turbidity, temperature, pH, and TDS. The importance of a parameter establishes the weight of the parameter. To calculate the WQI, each parameter (scaled to lie between 0 and 100) is multiplied by its weight, the multiplication is summed, and then the result is divided by the sum of the weights as shown in Equation (4):

  $$\mathit{WQI}=\frac{\left(\sum P_{\mathit{\text{VALUE}}}+W_{\mathit{\text{FACTOR}}}\right)}{\sum W_{\mathit{\text{FACTOR}}}}$$ (4)

  where $P_{\mathit{\text{VALUE}}}$ denotes the normalized score of a particular WQP, ranging from 0 to 100, and $W_{\mathit{\text{FACTOR}}}$ indicates how important it is to the overall quality of the water. Equation (4) makes sure that the overall effect of all the parameters is captured in a balanced measure of water quality WQI.

• WQCs

  WQCs provide a system through which the appropriateness of water to numerous applications can be measured and applied, for example, drinking, recreation, or the well‐being of an ecosystem. Measurements of the most important water quality measures and other general indicators, including the WQI, are used to identify these classes. The definition of WQCs can vary depending on the regulatory norms or parameters, which are established by the various bodies. These parameters are typically determined as being in a specific range or thresholds, and water is classified into a category depending on how its measurements fall within the range. The quality of water can be defined according to the WQI as follows:

  • Excellent—Water that met all legal requirements and was safe to drink.
  • Fair—Water that met some standards but required additional treatment or closer monitoring.
  • Good—Water that complied with most regulations but had minor variances with regard to certain parameters.
  • Marginal—Water that had major deviations from many regulations, hence requiring considerable processing.
  • Poor—Water that could not be used for consumption or irrigation purposes because of contamination.

3.4. Feature Selection Using AOA‐NO

The feature selection using AOA‐NO attempts to fine‐tune the preprocessed data by establishing the informative and minimal subset of attributes. With the help of iterative optimization, the method balances the relevance of features with the dimensionality reduction, leaving behind only the strongest determinants. The strategy successfully uses the addax optimization algorithm (AOA) in the world exploration approach, coupled with refinement of the adaptive parameter of Nadam to select the high‐value features that can increase the overall effectiveness and reliability of the following water quality prediction step.

3.4.1. AOA (Hamadneh et al. 2024)

The AOA algorithm is inspired by the natural intelligence of searching of the addax antelope, taking into account its efficient behavior when foraging for food and its usual method of digging. The addax antelope uses intelligent searching movements to locate vegetation and alters its searching behavior depending on the presence of available resources, which could be considered as exploration in the algorithm. The digging behavior of the animal represents exploitation. AOA converts these behaviors into mathematical operators that give candidate solutions a direction to more optimal regions and avoid less promising ones, guaranteeing an effective trade‐off between local refining and global search.

3.4.2. Nadam Optimizer (NO) (L. Li et al. 2020)

The NO is inspired by the fusion of two practical optimization ideas, which are momentum and adaptive learning. The physical properties of objects accelerating with time inspired momentum to aid optimization to get out of shallow minima. Adaptive learning came about as a result of the necessity to set the learning rates of each parameter depending on the previous gradients. Nadam combines Nesterov‐accelerated momentum with future gradient expectations, with Adam adaptive updates, to form a more responsive and stable optimizer that converges more effectively in deep learning applications.

• Initialization

  The hybrid AOA‐NO is initiated with the generation of every candidate solution within specified ranges with the help of a random value, whereas the first and second moment vectors of Nadam are set to zero. This is used to set the initial search population and optimizer state as given in Equation (5):

  $$P_a^{\left(0\right)}=\mathit{lB}+k\cdot \left(\mathit{uB}-\mathit{lB}\right),x_0=0,y_0=0$$ (5)

  In the AOA‐NO initialization, each candidate solution $P_a^{\left(0\right)}$ is obtained by using a random value $k\in \left[0,1\right]$ within the lower and upper bounds $\left(\mathit{lB},\mathit{uB}\right)$, according to the AOA position‐based initialization. Simultaneously, Nadam's internal adaptive parameters are initialized with the first‐moment vector $x_0=0$ and second‐moment vector $y_0=0$, in order to prepare the optimizer during feature‐selection updates and refinement throughout the exploitation process. Unified initialization enables the population to start its global exploration and enables Nadam to perform gradient‐based fine adjustments later.

• Fitness function

  The quality of each candidate feature subset of the AOA‐NO fitness function is determined by the relevance of the chosen features. The function focuses on subsets that capture highly informative features to effectively reduce dimensionality and enhance the quality of input to a model.

  $${\mathit{\text{Fitness}}}_a=\frac{1}{{\sum }_{b\in D_a}{L}_b}$$ (6)

  The fitness function described in Equation (6) assesses each feature subset based on its overall relevance. Here, every subset $D_a$ encompasses selected features, whereas each feature, $L_b$, contributes a relevance score calculated with the aid of correlation, entropy weight, or mutual information. Thus, the higher the relevance, the smaller the fitness value, which maintains the subset as superior. In contrast, low relevance corresponds to the most significant value of fitness, indicating poor feature quality.

• Exploration phase

  In AOA, the initial stage simulates the broad positional changes of addaxes in foraging based on their capacity to follow rainfall and search large expanses of terrain on their own in arid environments. These actions are used to direct large‐scale search in the space of problems. Every addax picks a random but appropriate vegetation zone, and the new location is calculated based on Equation (7). When this position enhances the objective function, it takes the place of the previous position as indicated by Equation (8):

  $$p_{a,b}^{X1}=p_{a,b}+k_{a,b}\cdot \left({\mathit{Ch}}_{a,b}-A_{a,b}\cdot p_{a,b}\right)$$ (7)

  $$P_a=\left\{\begin{array}{cc}{P}_a^{X1},& L_a^{X1}\le L_a\\ P_a,& \mathit{\text{Else}}\end{array}\right.$$ (8)

  In the above formulation, ${\mathit{Ch}}_a$ is the selected specific foraging zone for the ath addax, and ${\mathit{Ch}}_{a,b}$ determines its value in the bth dimension. $P_a^{X1}$ denotes the generated updated position for the ath addax during the foraging stage of AOA, whereas the bth‐dimensional component is represented as $p_{a,b}^{X1}$, and $L_a^{X1}$ is the objective value of the corresponding addax in this phase. The terms $k_{a,b}$ are randomly generated values within [0, 1], whereas $A_{a,b}$ are randomly chosen with values of either 1 or 2.

• Exploitation phase

  The exploitation stage in Nadam is when the optimizer is used to perfect the parameter updates with Nesterov‐accelerated adaptive gradients, so that the model can move in an optimal region. The mathematical expression of the exploitation phase is expressed in Equation (9):

  $$e_f=e_{f-1}-{\alpha }_f\frac{{\widehat{n}}_f}{\sqrt{{\widehat{m}}_f}+\epsilon }$$ (9)

  where $e_f$ denotes the updated parameter at step $f$ and ${\alpha }_f$ denotes the learning rate. During exploitation, Nadam makes use of the predictive Nesterov momentum ${\widehat{n}}_f$, which includes both past gradients and an anticipatory step toward the next gradient direction. Simultaneously, the adaptive second‐moment estimate ${\widehat{m}}_f$ controls the update size across every dimension.

• Termination

  In feature selection, the AOA‐NO algorithmically assesses and updates candidate subsets until the stopping criterion, say, maximum iterations or convergence to fitness, is met. The approach selectively narrows down each subset by striking a balance between the relevance of a feature and dimensionality reduction and finally giving rise to a compact and very informative feature set. After getting this optimal subset, the next step is water quality prediction, where the selected features are used as the predictive input.

3.5. ST‐DSCNN for IoT‐Based Water Quality Prediction Using WQI and WQC

The ST‐DSCNN combines the advantages of the convolutional Swin‐Transformer network (CSTN) (C. Wang et al. 2024) and depthwise separable convolutional neural network (DSCNN) (Hu et al. 2021) in improving the forecast of water quality using IoT. The method employs the selected features of water quality to serve as the predictive input in the process and accurately determines the numerical values of WQI and the label of WQC categories. The hybrid approach increases feature extraction, reinforces contextual learning, and promotes high‐quality assessment of the general water quality conditions.

This study is new in that it combines the use of Swin‐Transformer and depthwise separable CNN in a water quality prediction framework that operates on IoT. The proposed model uses the Swin‐Transformer and depthwise separable CNN‐based techniques to extract lightweight local features with minimal computational complexity and to capture the long‐range dependencies and contextual relationships between the WQPs. This mixed integration facilitates effective multiscale learning of features on sensor data. Moreover, AOA‐NO–based feature selection and CSBO‐based bias optimization are added to the architecture, which improve convergence stability, redundancy reduction, and prediction reliability. Therefore, the value is in the synergistic architecture design and optimization approach specific to the real‐time application of the IoT‐based water quality prediction and not the independent application of the existing architectures. Figure 2 illustrates the detailed architecture of the proposed ST‐DSCNN model, where Swin‐Transformer blocks capture long‐range contextual dependencies and depthwise separable CNN layers extract lightweight local spatial features for accurate WQI prediction and WQ classification.

FIGURE 2.

Architecture of the proposed ST‐DSCNN for WQI prediction and WQ classification.

3.5.1. CSTN

CSTN is a deep learning model that models the short‐ and long‐range dependence between the parameters of the IoT water‐quality models to predict numerical WQI values and classify WQC as good, fair, marginal, poor, and excellent. The CSTN employs convolutional projections, window self‐attention, and depthwise convolution in order to obtain hierarchical feature representations. The process of feature extraction and prediction may be expressed in the following equations:

$$\begin{array}{l}{\widehat{u}}^k=\mathit{WC}-\mathit{MSA}\left(\mathit{\text{conv\hspace{0.17em}proj}}\left(\mathit{Ln}\left(u^{k-1}\right)\right)\right)+u^{k-1}\hfill \\ u^k=\mathit{DFN}\left({\widehat{u}}^k\right)+{\widehat{u}}^k\hfill \end{array}$$ (10)

Equation (10) illustrates that input embeddings $u^{k-1}$ are normalized $\left(\mathit{Ln}\left(\cdot \right)\right)$, projected by convolution $\left(\mathit{\text{conv\hspace{0.17em}proj}}\left(\cdot \right)\right)$, and fed into window‐based convolutional multihead self‐attention $\left(\mathit{WC}-\mathit{MSA}\left(\cdot \right)\right)$. Furthermore, the depth‐convolution forward network, $\mathit{DFN}\left(\cdot \right)$, enhances cross‐window information. Residual connections $\left(+u^{k-1},+{\widehat{u}}^k\right)$ preserve gradient stability. ${\widehat{u}}^k$ and $u^k$ denote the intermediate and final output of the kth block, respectively.

$$\begin{array}{l}\left[Z,F,S\right]=\mathit{\text{conv\hspace{0.17em}proj}}\left(u^{k-1}\right)\hfill \\ =\mathit{\text{flatten}}\left(\mathit{\text{depth\hspace{0.17em}conv}}\left(\left(W\mathit{\text{in\hspace{0.17em}partition}}\left(\mathit{\text{reshape}}\left(u^{k-1}\right)\right)\right)\right)\right)\hfill \end{array}$$ (11)

Equation (11) defines how, given the kth input embedding $u^{k-1}$, the query, key, and value matrices $\left(Z,F,S\right)$ are generated, respectively, through reshaping, flattening, and depthwise convolution $\mathit{\text{depth\hspace{0.17em}conv}}\left(\cdot \right)$. These matrices encode local spatial relationships among the IoT WQPs: pH, temperature, turbidity, and TDS.

$$\mathit{MSA}\left(Z,F,S\right)=\text{soft max}\left(Z{F}^{{\rm T}}/\sqrt{d}+Y\right)S$$ (12)

Equation (12) performs multihead self‐attention, where $Z,F,S\in R^{N^2\times d}$ are convolutional embeddings; $d$ is the dimensionality of query/key vectors; $Y\in R^{N^2\times N^2}$ is the learnable relative position bias; and $N^2$ stands for the number of patches. This operation computes attention‐enhanced features, which capture dependencies among WQPs.

$$\begin{array}{c}b=\mathit{Ln}\left(\mathit{\text{depth}}-\mathit{\text{conv}}\left({\widehat{u}}^{k-1}\right)\right)\hfill \\ \widehat{b}=\mathit{\text{point}}-\mathit{\text{conv}}\left(\mathit{\text{GELU}}\left(\mathit{\text{point}}-\mathit{\text{conv}}\left(b\right)\right)\right)\hfill \\ u^k={\widehat{u}}^{k-1}+\widehat{b}\hfill \end{array}$$ (13)

Equation (13) shows the DFN operation, where ${\widehat{u}}^{k-1}$ denotes the input from the attention module, $\mathit{\text{depth}}-\mathit{\text{conv}}\left(\cdot \right)$ denotes the depthwise convolution, $\mathit{\text{point}}-\mathit{\text{conv}}\left(\cdot \right)$ means the pointwise convolution, $\mathit{\text{GELU}}\left(\cdot \right)$ is the activation, and $u^k$ is the final output of the kth CSTB. This layer ensures the local spatial context is preserved while supporting WQI prediction and WQC classification.

3.5.2. DSCNN

DSCNN represents a kind of deep learning network that effectively discovers spatial and cross‐channel attributes in the input of the IoT sensor (temperature, pH, turbidity, and TDS). The network forecasts WQI and WQC using depthwise convolution and batch normalization, followed by ReLU and pointwise convolution functions.

$$S_k^{e_{k-1}^g,s_k^e,s_k^W}={\sum }_{h^{\mathit{se}}=1}^{H^{\mathit{se}}}{\sum }_{h^{\mathit{sW}}=1}^{H^{\mathit{sW}}}{E}_k^{e_{k-1}^g,v_k\cdot s_k^e+h^{\mathit{se}},v_k\cdot s_k^e+h^{\mathit{sW}}}{H}_k^{e_{k-1}^g,h^{\mathit{se}},h^{\mathit{sW}}}+x_k^{e_{k-1}^g}$$ (14)

Equation (14) describes the depthwise convolution in the kth DSCNN block, where $E^{k-1}$ denotes the input feature map from the previous layer; $H_k^{\mathit{se}}$ and $H_k^{\mathit{sW}}$ are the kernel height and width, $v_k$ is the stride, $s_k^e$ and $s_k^W$ are spatial indices, $H^k$ is the depthwise convolution kernel, and $x^k$ is the bias. This operation extracts spatial information independently per input channel.

$$S_k^{\prime }=\text{Re}\mathit{LU}\left(\mathit{Bn}\left(S_k\right)\right)$$ (15)

$\mathit{Bn}\left(\cdot \right)$ is a batch normalization followed by a ReLU activation function applied on depthwise output $S_k$, given by Equation (15). $S_k^{\prime }$ represents the normalized and activated feature map. This also stabilizes learning and brings nonlinearity to record intricate connections between WQPs.

$$V_k^{e_k^g,s_k^e,s_k^W}={\sum }_{e_{k-1}^g=1}^{E_{k-1}^g}{S^{\prime }}_k^{e_{k-1}^g,s_k^e,s_k^W}{U}_k^{e_k^g,e_{k-1}^g}+y_k^{e_k^g}$$ (16)

Equation (16) describes the pointwise or 1 × 1 convolution, where $S_k^{\prime }$ is the depthwise feature map and $U_k$, $y_k$, and $V_k$ are the pointwise convolution kernel, bias, and output, respectively. This layer aggregates cross‐channel information and prepares the features for the final flattening and classification into WQI values and WQC classes.

ST‐DSCNN methodology works with the integration of CSTN window‐based attention and DSCNN lightweight convolution to derive the rich spatial–temporal patterns of the chosen water quality features serving as prediction inputs. These processed representations allow proper estimation of WQI values, as well as correct classification of WQC. In order to further improve predictive reliability, the CSBO algorithm is used to optimally tune the bias parameters of the ST‐DSCNN model.

3.6. Bias Tuning of ST‐DSCNN via CSBO

The dynamics of the human circulatory system, which continuously circulates blood throughout the body via the systemic and pulmonary channels to support life, served as the model for CSBO (Ghasemi et al. 2022). This is a natural process that expresses a natural process of refinement, renewal, and selective improvement. In CSBO, the blood droplets are all possible solutions, and their motion in the circulatory cycle is analogous to the transitions between weaker deoxygenated states and stronger oxygenated states. Through this biological behavior, the algorithm reinforces high‐quality solutions, eliminates poorer ones, and, through a series of steps, guides the population to good areas in the search space. The process of CSBO is presented in Figure 3. Yet despite being a relatively new metaheuristic, CSBO was chosen in the study due to its balanced exploration–exploitation nature and derailing convergence properties during neural network parameter optimization. In comparison to the gradient‐only updates or population methods, which can converge early, CSBO uses adaptive circulation‐based movements to keep the population diverse but refines its promising proposals further. It is especially helpful in the optimization of deep networks with a very nonlinear optimization landscape to tune the bias parameters.

FIGURE 3.

Step‐by‐step procedure of CSBO.

• Step 1:

  Initialization

  The optimization process begins after the CSBO method's parameters are chosen. The bias parameter can be effectively and iteratively optimized using ST‐DSCNN once the optimization environment is set up.

• Step 2:

  Random generation

  At the beginning of the search, potential options are chosen at random. Variability can be managed to keep solutions from converging during the random generating step. The solution contains a single potential ST‐DSCNN bias parameter. To identify the best random selections, CSBO first iteratively refines them.

• Step 3:

  Fitness function

  The fitness function uses the current bias parameter to assess the performance quality of the ST‐DSCNN candidate solution. Equation (17) assesses the quality of the solution by changing the bias value to compute the fitness function and provides performance feedback to direct correction.

  $$\mathit{\text{Fitness}}\mathit{\text{function}}=\mathit{Min}\left(y\right)$$ (17)

  where $y$ denotes the bias parameter of ST‐DSCNN.
• Step 4:

  Vein movement phase

  This stage transports every blood mass via the veins to more promising solutions through the influences of pressure‐driven fitness.

  $${\mathit{bM}}_p^{\mathit{New}}={\mathit{bM}}_p+L_{p1}\times x_p\times \left({\mathit{bM}}_p-{\mathit{bM}}_1\right)+L_{23}\times x_p\times \left({\mathit{bM}}_3-{\mathit{bM}}_2\right)$$ (18)

  Equation (18) updates the blood mass ${\mathit{bM}}_p^{\mathit{New}}$ with a change in the current position based on pressure factors $L_{p1}$, $L_{23}$, and the displacement controller $x_p$. The terms $\left({\mathit{bM}}_p-{\mathit{bM}}_1\right)$ and $\left({\mathit{bM}}_3-{\mathit{bM}}_2\right)$ direct the movement toward the space between these two reference masses according to the differences between them to allow directional shifting toward the region of better fitness.

• Step 5:

  Pulmonary circulation phase

  This step renews the weaker masses of blood by oxygenation‐inspired random perturbations, enhancing exploration.

  $${\mathit{bM}}_p^{\mathit{New}}={\mathit{bM}}_p+\left(\frac{{\mathit{\text{Rand}}}_m}{\mathit{pf}}\right)\times {\mathit{\text{Rand}}}_g\left(1,V\right)$$ (19)

  As shown in Equation (19), the new position, ${\mathit{bM}}_p^{\mathit{New}}$, was produced by the addition of a perturbation, which consisted of a Gaussian random value, ${\mathit{\text{Rand}}}_m$, scaled by the iteration index $\mathit{pf}$, and a Cauchy‐distributed vector ${\mathit{\text{Rand}}}_g\left(1,V\right)$. This mix of random distributions helps to escape from local optima by reintroducing diversity.

• Step 6:

  Systematic circulation phase

  This step increases exploitation by moving strong blood masses to the best regions of the search space.

  $${\mathit{bM}}_{p,q}^{\mathit{New}}={\mathit{bM}}_{1,q}+x_p*\left({\mathit{bM}}_{3,q}-{\mathit{bM}}_{2,q}\right)$$ (20)

  From Equation (20), the new value for the dimension ${\mathit{bM}}_{p,q}^{\mathit{New}}$ is found by moving the mass of blood toward the optimal candidate. The optimal solution is ${\mathit{bM}}_{1,q}$, whereas $x_p*\left({\mathit{bM}}_{3,q}-{\mathit{bM}}_{2,q}\right)$ adds a fitness‐based correction factor. This targeted approach improves exploitation through concentrated efforts within good regions.

• Step 7:

  Termination

  After the completion of the termination stage and when all the termination criteria have been met, the process of optimization is achieved. The proposed CSBO algorithm manages to reduce the bias value of the ST‐DSCNN at the termination stage. Further reduction will not add much to the performance of the system, even if there is convergence.

The ST‐DSCNN‐CSBO system can be used to monitor the water quality of an IoT sensor through the collection of turbidity, temperature, pH, and TDS data in real time, provided that the distribution of sensors is comprehensive. The data is transferred to the cloud, where it is first processed with the QAFMMN to remove inconsistencies and normalize values. The AOA‐NO is used to conduct feature selection, indicating the main attributes in the most appropriate way. The ST‐DSCNN works on the selected features and predicts the WQC and WQI. CSBO is used to optimize bias parameters, which improves stability in prediction. On the whole, this framework will help to carry out correct assessments of the water quality based on IoT due to multiple data management and deep learning methods.

The computational complexity and convergence behavior of the CSBO optimization process were used to determine its efficiency. The complexity of the overall time of CSBO is largely determined by the population size ($N$), number of iterations ($T$), and dimensionality of the search space ($D$). The optimum solutions of each iteration are obtained by revising the candidate solutions during the venous movement, pulmonary circulation, and systemic circulation. Thus, the computational complexity of the algorithm can be estimated as $O\left(N\times T\times D\right)$, which is similar to population‐based metaheuristics that are widely used. With the lightweight arithmetic operations CSBO uses, but no complex gradient computations, the optimization process is still computationally efficient in neural network parameter tuning.

ALGORITHM 1. CSBO‐based bias optimization for ST‐DSCNN.

Input: Initial population of blood masses ${\mathit{bM}}_p$, bias parameter $y$, population size $P$, maximum iterations $\mathit{pf}$Output: Optimal bias parameter $y^{*}$InitializationGenerate initial blood masses ${\mathit{bM}}_p$ randomly within the search space.Set iteration counter $f=1$.Fitness EvaluationFor each blood mass ${\mathit{bM}}_p$, evaluate the fitness using   $\mathit{\text{Fitness}}\mathit{\text{function}}=\mathit{Min}\left(y\right)$   where $y$ denotes the ST-DSCNN bias parameter.While $f<\mathit{pf}$ doVein Movement Phase (Exploration)Update each blood mass position using   ${\mathit{bM}}_p^{\mathit{New}}={\mathit{bM}}_p+L_{p1}\times x_p\times \left({\mathit{bM}}_p-{\mathit{bM}}_1\right)+L_{23}\times x_p\times \left({\mathit{bM}}_3-{\mathit{bM}}_2\right)$Pulmonary Circulation Phase (Diversity Enhancement)Apply oxygenation-inspired perturbation:   ${\mathit{bM}}_p^{\mathit{New}}={\mathit{bM}}_p+\left(\frac{{\mathit{\text{Rand}}}_m}{\mathit{pf}}\right)\times {\mathit{\text{Rand}}}_g\left(1,V\right)$Systematic Circulation Phase (Exploitation)Move blood masses toward the best candidate solution:   ${\mathit{bM}}_{p,q}^{\mathit{New}}={\mathit{bM}}_{1,q}+x_p*\left({\mathit{bM}}_{3,q}-{\mathit{bM}}_{2,q}\right)$Update PopulationReplace the previous ${\mathit{bM}}_p$ with ${\mathit{bM}}_{p,q}^{\mathit{New}}$ if the new solution improves the fitness.Update Best SolutionIdentify the best blood mass ${\mathit{bM}}_{1,q}$ corresponding to the minimum fitness value.Increment Iteration   $f=f+1$End WhileReturn Optimal Bias Parameter$y^{*}={\mathit{bM}}_{1,q}$

4. Results and Discussion

The proposed ST‐DSCNN‐CSBO was developed in Python with the help of TensorFlow to ensure computational efficiency, stability, and simple scaling of the water quality data to large scales. The model has been critically evaluated against several performance metrics—precision, specificity, recall, accuracy, F1‐score, MAE, R ², RMSE, and MSE—to comprehensively analyze the model performance in predicting behavior in WQI and WQC. This thorough research established the robustness, reliability, and predictive quality of the proposed framework in handling complex water quality scenarios that are inspired by IoT. Table 2 provides a detailed experimental configuration and processing workflow.

TABLE 2.

Implementation requirements for the proposed method.

Component	Hyperparameter	Value
System	Operating system	Windows 11 (64‐bit)
	Platform	Python 3.10
	Deep learning framework	TensorFlow 2.12
	Additional library	NumPy
	Hardware requirement	NVIDIA RTX 3060 GPU
Data preprocessing (QAFMMN)	Normalization method	Min–max scaling
	Noise filtering window size	5
AOA‐NO feature selection	Population size	30
	Maximum iterations	50
	Exploration coefficient	0.5
	Feature selection threshold	0.6
Swin‐Transformer	Patch size	4
	Embedding dimension	96
	Number of attention heads	4
	Window size	7
	Transformer layers	2
Separable CNN (ST‐DSCNN)	Convolution filter size	3 × 3
	Number of filters	64
	Activation function	ReLU
	Batch size	32
	Learning rate	0.001
	Optimizer	Adam
	Epochs	100
CSBO optimization	Population size	25
	Maximum iterations	50
	Search dimension	Bias parameters
	Convergence tolerance	10−5
Model evaluation	Train–test split	80:20
	Cross‐validation	5‐fold

4.1. Data Collection Framework

• Study area

  The study area was the Rohri Canal, which is situated in Nawabshah City, Shaheed Benazirabad (SBA) District, and Sindh Province, Pakistan. This canal draws water from the Indus River near Sukkur and has a gravity‐fed system that supplies irrigation water and drinking water to the surrounding communities. It develops an extensive network of distributors and minor canals that support the agricultural activities within the SBA District. This canal needs continuous monitoring due to its importance as a regional source of water. The conventional monitoring relied on manual sampling at various sites, which motivated the application of an IoT‐enabled system capable of efficient, continuous, and automated measurement of WQPs.

• Dataset generation

  This dataset was generated using an IoT‐based water quality monitoring framework deployed at the Rohri Canal. Four sensors were located in the water body for real‐time variations in WQP: turbidity, temperature, pH, and TDS. These sensors used LoRa connectivity to transmit their readings wirelessly back to a central base station as these parameters continued to be measured. This incoming data was saved in the database by the base station for later processing. In total, 6000 observations for each of the parameters were collected over the span of 4 months during the autumn season; nevertheless, because of the security risk involved, it was possible to collect data only during the daytime hours. In order to test the efficiency of different machine learning algorithms, the entire dataset was split into Dataset 1 having 600 observations per parameter and Dataset 2 containing 6000 observations per parameter. The data was divided, with 80% being used as a training model and 20% as a test model. The reliable and representative design of this method ensured that the analysis and prediction tasks had a reliable and representative design (Rahu et al. 2023). The utilized data can be collected from https://www.pcrwr.gov.pk/water‐quality/?utm_source=copilot.com.

  The acceptable limits presented in Table 3 form the ground on which one interprets whether the read values are within the acceptable safety ranges. These limits are subsequently applied in the WQI calculation and in determining the general WQC. The datasets used in this study were obtained from publicly available water quality monitoring repositories (https://www.kaggle.com/datasets/adityakadiwal/water‐potability) and previously reported environmental datasets commonly used for water quality prediction research. These parameters are temperature, pH, turbidity, and TDS, which are standard parameters measured using calibrated IoT‐based environmental sensors. Sensor measurements are common with the rest of the practices of environmental monitoring, in which devices are regularly calibrated as specified by the manufacturer to ensure validity. Because the study utilizes publicly available environmental data without involving human or animal subjects, formal ethical clearance was not required.

TABLE 3.

Acceptable limits of water quality parameters for drinking and irrigation purposes.

Parameter	Unit	Acceptable limit for drinking water	Acceptable limit for irrigation water
pH	n/a	6.5–8.5	6.0–7.5
TDS	mg/L	600	< 2000
Temperature	°C	< 25	< 25
Turbidity	NTU	1.0–5.0	n/a

4.2. Evaluation Metrics

Precision, recall, specificity, accuracy, F1‐score, MAE, R ², RMSE, and MSE were used to determine the reliability, robustness, and computational efficiency of the proposed ST‐DSCNN‐CSBO model. Its integrated feature selection, feature extraction, and bias optimization with CSBO allow the correct, scalable, and reliable prediction of the IoT‐based water quality, which in its turn provides the whole picture in effective water quality monitoring and decision‐making.

The equation and performance measures are displayed in Table 4, where $T^P$ denotes the true positive, $F^P$ represents the false positive, $T^N$ denotes the true negative, $F^N$ represents the false negative, $y_i$ is the observed or actual value, $\widehat{y}$ is the predicted value extracted from the model, and $n$ is the number of data samples employed for model evaluation.

TABLE 4.

Metrics of evaluation and equation.

Metrics	Formula
Accuracy	$\mathit{Acc}=\frac{T^P+T^N}{T^P+T^N+F^P+F^N}$
Precision	$\mathrm{Pre}=\frac{T^P}{T^P+F^P}$
Recall	$\mathrm{Rec}=\frac{T^P}{T^P+F^N}$
F1‐score	$F1=\frac{2*\mathrm{Pre}*\mathrm{Rec}}{\mathrm{Pre}+\mathrm{Rec}}$
Specificity	$\mathit{\text{Spec}}=\frac{T^N}{T^N+F^P}$
Mean absolute error (MAE)	$\mathit{MAE}=\frac{1}{n}\sum _{i=1}^n\left|y_i-{\widehat{y}}_i\right|$
Mean squared error (MSE)	$\mathit{MSE}=\frac{1}{n}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2$
Root mean square error (RMSE)	$\mathit{\text{RMSE}}=\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$
R 2 error	$R^2=1-\frac{\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$

4.3. Performance Analysis

The performance evaluation of the ST‐DSCNN‐CSBO is examined here.

Figure 4 shows the convergence pattern of the proposed ST‐DSCNN‐CSBO model when training and testing it on Dataset 1 using 100 epochs. The high loss rates during the initial periods of time enhance the sharp decline in the loss points on the effective learning of parameters and optimization. The performance of the two methods of training and testing approaches both the high values and the stabilization in a slow way and with minimal variation, which reflects good generalization. The strong correlation between the curves shows that the model is not over‐fitted and shows consistent predictive quality on unknown samples of water quality.

FIGURE 4.

Training and testing loss–accuracy curves for Dataset 1.

Figure 5 illustrates the training dynamics of the model using Dataset 2, which is larger. The loss curves exhibit a sharp decrease in early iterations and then are smooth, meaning that model parameters are optimized. At the same time, the accuracy curves continuously increase and approach large values in both the training and testing sets. The fact that the distance between the curves is minimal indicates high generalization ability, and this validates that the proposed architecture is efficient in learning complex patterns when using larger amounts of data in IoT‐based water quality.

FIGURE 5.

Training and testing loss–accuracy curves for Dataset 2.

Figure 6 demonstrates the correlation relations between temperature, pH, turbidity, TDS, and WQI in Dataset 1. The heatmap shows that there exists a moderate positive relation between WQI and individual parameters, which proves their applicability in the evaluation of water quality. In the meantime, the existence of weak correlations among parameters means that there is no significant redundancy among features. This independence justifies the usefulness of the feature selection phase and makes sure that each parameter is a source of unique information to be used in the predictive modeling process.

FIGURE 6.

Correlation heatmap—Dataset 1.

Figure 7 illustrates the correlation framework of the bigger Dataset 2. Just like in the case of Dataset 1, the heatmap of Dataset 2 shows that there are consistently moderate correlations among WQI and sensor parameters, which establish their predictive significance. The weak negative relationships between the input parameters suggest that there is no significant multicollinearity; hence, the stable learning is facilitated in the deep learning system. The patterns of correlations observed in different datasets also confirm the accuracy and strength of the obtained measurements of IoT sensors.

FIGURE 7.

Correlation heatmap—Dataset 2.

The pairwise visualization of a correlation among the WQPs and WQI is presented in Figure 8. The individual distribution of each parameter is shown in the diagonal plots, whereas the interaction patterns between pairs of variables are shown in the scatter plots. TDS and WQI have a strong association, which suggests that they have a greater impact on the water quality metric than other factors. Such visualization assists in uncovering missing dependencies and enables the multiparameter learning of the suggested model.

FIGURE 8.

Pair plot of water quality parameters.

Violin plots are used as shown in Figure 9 to determine the distribution and variability of turbidity, temperature, pH, and TDS. Densities and central tendencies of the parameters can be viewed in the form of the shapes of the plots. The distribution of temperature, pH, and turbidity is relatively narrow, which means that there are stable environmental conditions at the time of data collection. TDS, conversely, is less predictable and variably shows the changes in dissolved mineral levels. This analysis gives information on the dispersion of parameters that affect the prediction of water quality.

FIGURE 9.

Violin plot of water quality parameters.

Figure 10 shows the correlation between the temperature and WQI in the form of a scatter distribution. The cloudy pattern shows a comparatively low linear relationship between the two variables, which shows that temperature in itself is not a very strong factor that determines the overall water quality. Nevertheless, small groups around moderate temperatures imply indirect contact with other parameters. This point confirms the necessity of multifeature analysis in the prediction of water quality instead of the use of single‐parameter relationships.

FIGURE 10.

Scatter plot of temperature versus WQI.

The empirical cumulative distribution function of WQI values is shown in Figure 11. The increasing curve of the ECDF shows that the observations of water quality are evenly distributed throughout the dataset and do not show any sudden spikes in concentration. The steeper parts are those ranges in which the data points are thicker. This analysis assists in determining percentile thresholds and patterns of the distribution of WQI that are applied in the predictive model to aid in a better understanding of the boundaries of water quality classification as applied.

FIGURE 11.

ECDF of WQI.

The frequency distribution of the temperature measurements is presented in Figure 12 with a density curve. The majority of the observations are in the 24°C to 28°C, which points to the relative stability of the thermal conditions during the period of monitoring. A bell‐shaped pattern implies a more or less normal distribution of the value of the temperature. The congruency of the bars with the density curve in the histograms signifies that measurements of the environment are constant and, thus, give acceptable input data to the predictive modeling process.

FIGURE 12.

Temperature distribution histogram with density curve.

In Figure 13, the real WQI values are compared to the predicted results of the proposed model of Dataset 1. The value of most data points is near the ideal diagonal reference line, implying that both predicted values and the true values are in close agreement. This small error spread about the line indicates very little error in prediction and successful learning of the patterns of underlying water quality. This outcome revealed the ability of the model to represent nonlinear associations between parameters, even when using a relatively small dataset.

FIGURE 13.

Actual versus predicted WQI—Dataset 1.

The accuracy of the prediction of the model on the bigger Dataset 2 is presented in Figure 14. The ample concentration of the points around the diagonal line is evidence of very high predictive consistency in actual and estimated WQI values. The distribution is less spread as compared to Dataset 1, which is a sign of improved learning in the presence of more training data. This proves the scalability of the proposed framework and its capability to ensure high prediction accuracy in the case of the availability of larger IoT datasets.

FIGURE 14.

Actual versus predicted WQI—Dataset 2.

Figure 15 shows the prediction performance of the model under various water quality categories, with the classes being excellent, good, fair, marginal, and poor. Close attachment of each color‐coded cluster to the reference diagonal shows that all the categories possess the same level of predictable accuracy. The overall distribution is well balanced, thus indicating that the model is effective in differentiating the various water quality conditions without giving preference to any particular class, hence supporting the fact that the model can be utilized in various practical monitoring and classification tasks.

FIGURE 15.

Actual versus predicted WQI across quality categories.

The average values of WQI of each category of water quality are displayed in Figure 16. The upward trend in the values of WQI by the classes indicates the anticipated trend of the progression of lower to higher pollution levels. The categorical distinction is clear enough to prove that the dataset offers clear‐cut points of classification. The effectiveness of the model in acquiring different patterns with different water quality conditions is also supported by this distribution.

FIGURE 16.

WQI distribution across quality classes.

Figure 17 displays the changes in temperature and turbidity with time in Dataset 1. The temperature sequence indicates that the temperature varies moderately around the same mean, indicating the seasonal changes in the natural environment over the course of the monitoring. On the other hand, turbidity reflects more variations in the dynamics of sedimentary movement and water flow. The trends help to prove the dynamism of the water quality indicators and the need to track them in real time with the assistance of the IoT sensors.

FIGURE 17.

Time‐series plots—Dataset 1: (a) temperature and (b) turbidity.

The temporal variations of TDS and pH of Dataset 1 are shown in Figure 18. TDS changes over time, which means that the concentrations of the dissolved minerals change with the environmental and hydrological changes. At the same time, pH levels are relatively constant with minor variations toward the neutral levels. The observations show the inherent chemical balance of the water body and give practical temporal trends that can be acquired by the predictive model.

FIGURE 18.

Time‐series plots—Dataset 1: (a) TDS and (b) pH.

Figure 19 represents the time dependence of temperature and turbidity of the larger Dataset 2. Temperature estimates exhibit slow changes that indicate how temperature indeed varies in the environment, whereas turbidity exhibits sporadic peaks that indicate that there is a disturbance in sediment or a change in flow. The stable patterns of measurements show that the sensor can be used reliably throughout the time of monitoring. These temporal variations underscore the relevance of temporal modeling in the correct prediction of water quality.

FIGURE 19.

Time‐series plots—Dataset 2: (a) temperature and (b) turbidity.

Figure 20 shows the time‐series patterns of TDS and pH of Dataset 2. The TDS values show periodic change due to changes in mineral content and water flow conditions. Conversely, pH values are comparatively constant with minor variations from neutral conditions. The trends that are similar in the two parameters imply that the sensor data is reliable and that the temporal data is meaningful to the deep learning model to learn water quality dynamics.

FIGURE 20.

Time‐series plots—Dataset 2: (a) TDS and (b) pH.

4.4. Comparative Analysis

The suggested ST‐DSCNN‐CSBO was compared to various current approaches, such as GGNN (Z. Li et al. 2024), DSTCNN (Arepalli and Naik 2024), FSGCN (Qiao et al. 2024), DeepTCN‐GRU (Tian et al. 2024), VBAED (Bi et al. 2024), SBiLSTM (Xu et al. 2024), LSTM (Z. Wang, Wang, and Wu 2023), hybrid CNN–BiLSTM (Geetha et al. 2024), AQGRU (Kumar and Vellaichamy 2026), and WD‐LSTM‐Transformer (Huan et al. 2025). The experimental results show that ST‐DSCNN‐CSBO provides superior results in terms of several parameters, such as F1‐score, recall, accuracy, specificity, precision, MAE, RMSE, and R ². The model is very precise and fast in predicting water quality; it has outperformed other available ML and DL models because it integrates feature selection, deep feature extraction, and bias optimization.

Table 5 shows the performance comparison between the current methods and the proposed ST‐DSCNN‐CSBO in estimating water quality using IoT. In existing frameworks such as GGNN, DSTCNN, FSGCN, DeepTCN‐GRU, VBAED, SBiLSTM, and LSTM, an F1‐score, recall, accuracy, specificity, and precision of 89.65%–92.34% are achieved. In contrast, the proposed ST‐DSCNN‐CSBO framework demonstrates improved predictive performance, achieving accuracy, precision, recall, F1‐score, and specificity of 99.96%, 99.94%, 99.95%, 99.93%, and 99.92%, respectively.

TABLE 5.

Comparative performance of existing methods and proposed ST‐DSCNN‐CSBO.

Methods	Accuracy (%)	Precision (%)	Recall (%)	F1‐score (%)	Specificity (%)
GGNN (Z. Li et al. 2024)	90.12	89.95	90.25	90.10	89.98
DSTCNN (Arepalli and Naik 2024)	91.05	90.88	91.12	91.00	90.95
FSGCN (Qiao et al. 2024)	89.87	89.65	89.90	89.80	89.70
DeepTCN‐GRU (Tian et al. 2024)	91.56	91.40	91.50	91.45	91.38
VBAED (Bi et al. 2024)	92.12	92.00	92.05	92.03	91.95
SBiLSTM (Xu et al. 2024)	90.78	90.65	90.80	90.72	90.68
LSTM (Z. Wang, Wang, and Wu 2023)	91.34	91.20	91.30	91.25	91.18
Hybrid CNN–BiLSTM (Geetha et al. 2024)	93.85	93.6	93.75	93.67	93.5
AQGRU (Kumar and Vellaichamy 2026)	96.2	96.05	96.1	96.07	96
WD‐LSTM‐Transformer (Huan et al. 2025)	95.48	95.3	95.42	95.36	95.25
ST‐DSCNN‐CSBO (proposed)	99.96	99.94	99.95	99.93	99.92

Table 6 compares the performance of different models, such as GGNN, DSTCNN, FSGCN, DeepTCN‐GRU, VBAED, SBiLSTM, LSTM, and the proposed ST‐DSCNN‐CSBO on the basis of MAE, MSE, RMSE, R ², and computing time. Current approaches have MAE values of between 0.115 and 0.124 and R ² of between 0.889 and 0.899, with a computational time of 0.38–0.45 s. ST‐DSCNN‐CSBO has the best errors (MAE 0.008, MSE 0.006, RMSE 0.024), the highest R ² of 0.999, and the lowest computational time of 0.02 s, indicating that it is more predictive accurate, robust, and efficient than the current methods. The proposed model achieves not only superior prediction accuracy but also significantly reduced edge inference time (34 ms) and memory footprint (96 MB) compared to existing methods. This efficiency is mainly due to the use of depthwise separable convolutions and optimized bias tuning, making the model suitable for real‐time IoT deployment. The results confirm that the framework balances high predictive performance with practical deployability in resource‐constrained environments.

TABLE 6.

Comparison of regression performance and computational efficiency of existing methods and proposed ST‐DSCNN‐CSBO.

Methods	MAE	MSE	RMSE	R 2	Computational time (s)	Edge inference time (ms)
GGNN (Z. Li et al. 2024)	0.124	0.038	0.195	0.891	0.45	128
DSTCNN (Arepalli and Naik 2024)	0.118	0.035	0.187	0.894	0.41	112
FSGCN (Qiao et al. 2024)	0.121	0.037	0.192	0.889	0.43	119
DeepTCN‐GRU (Tian et al. 2024)	0.116	0.034	0.184	0.898	0.39	105
VBAED (Bi et al. 2024)	0.119	0.036	0.19	0.892	0.42	116
SBiLSTM (Xu et al. 2024)	0.117	0.035	0.187	0.895	0.4	109
LSTM (Z. Wang, Wang, and Wu 2023)	0.115	0.033	0.181	0.899	0.38	101
Hybrid CNN–BiLSTM (Geetha et al. 2024)	0.102	0.029	0.17	0.915	0.34	92
AQGRU (Kumar and Vellaichamy 2026)	0.089	0.024	0.155	0.938	0.31	85
WD‐LSTM‐Transformer (Huan et al. 2025)	0.093	0.026	0.161	0.931	0.33	88
ST‐DSCNN‐CSBO (proposed)	0.008	0.006	0.024	0.999	0.02	34

Table 7 shows the result of fivefold cross‐validation, where the prediction accuracy of the newly developed model, ST‐DSCNN‐CSBO, is relatively consistent for all data partitioning schemes. There is not much difference in the standard deviation ranging from 99.72% to 99.83%, whereas the average R ² value is 0.998. The fact that there is only a slight difference with the folds indicates that the model is reliable in generalization and does not require any particular training set. These findings also justify the strength and stability of the suggested framework in terms of predicting water quality with the use of the IoT.

TABLE 7.

Fivefold cross‐validation performance of the proposed ST‐DSCNN‐CSBO model.

Fold	Accuracy (%)	Precision (%)	Recall (%)	F1‐score (%)	R 2
Fold 1	99.72	99.68	99.70	99.69	0.997
Fold 2	99.81	99.77	99.79	99.78	0.998
Fold 3	99.76	99.71	99.74	99.72	0.998
Fold 4	99.83	99.80	99.81	99.80	0.998
Fold 5	99.79	99.75	99.77	99.76	0.998
Average	99.78	99.74	99.76	99.75	0.998

4.5. Theoretical Justification and Comparative Positioning of CSBO

Even though the CSBO algorithm improves the optimization phase of the suggested approach, its innovation is still considered a step‐by‐step development of already established optimization techniques. For more emphasis on its value, the algorithm is provided with an additional theoretical perspective and a more detailed analysis, among other latest approaches used in optimization. From a theoretical point of view, the CSBO algorithm provides a unique two‐stage circulation strategy that includes both exploration and exploitation phases and operates as a biologically inspired circulation. In contrast to traditional algorithms, where the ratio between exploration and exploitation processes is fixed and predetermined, in CSBO, diversification occurs due to pulmonary circulation, and intensification occurs because of systemic circulation.

In order to prove the efficiency of the introduced CSBO approach, a comparative analysis was made by applying the algorithm to solve the same problem as with five other recently suggested algorithms, namely, the Harris Hawks optimization (HHO) approach, slime mold algorithm (SMA), marine predators algorithm (MPA), arithmetic optimization algorithm (AOA), and whale optimization algorithm (WOA).

Table 8 illustrates that CSBO requires the least amount of processing time while achieving the best performance across all evaluation metrics. The improvement, though incremental, highlights its efficient balance between exploration and exploitation with lower complexity compared to recent algorithms. This supports its suitability for lightweight, real‐time IoT‐based optimization while maintaining competitive accuracy.

TABLE 8.

Comparison of CSBO with recent metaheuristic algorithms.

Method	Accuracy (%)	MAE	RMSE	R 2	Computation time (s)
WOA	98.62	0.017	0.066	0.994	0.05
AOA	98.74	0.015	0.061	0.996	0.05
HHO	98.88	0.013	0.054	0.997	0.06
SMA	99.02	0.012	0.049	0.997	0.06
MPA	99.15	0.011	0.046	0.998	0.06
CSBO (proposed)	99.96	0.008	0.024	0.999	0.02

4.6. Ablation Study

An ablation experimentally eliminates or changes the elements of a model to determine the contribution of each component. It assists in interpreting the importance of preprocessing, feature selection process, model architecture, and optimization procedure and determining which modules have the most significant effect on predictive performance and which is sufficient to guarantee the strength of the resulting integrated model. Table 9 shows that ablation study of ST‐DSCNN‐CSBO for water quality prediction.

TABLE 9.

Ablation study of ST‐DSCNN‐CSBO for water quality prediction.

Model variant	Accuracy (%)	Precision (%)	Recall (%)	F1‐score (%)	Specificity (%)
Without preprocessing	95.87	95.74	95.69	95.71	95.62
Without feature selection	96.42	96.31	96.28	96.29	96.19
Without CSTN	96.05	95.92	95.88	95.90	95.81
Without DSCNN	96.57	96.44	96.39	96.41	96.31
Without CSBO optimization	98.23	98.12	98.07	98.09	97.98
Proposed model (ST‐DSCNN‐CSBO)	99.96	99.94	99.95	99.93	99.92

The removal of AOA‐NO makes the accuracy drop to 96.42%, whereas recall goes down to 96.28%. The deletion of Swin‐Transformer and depthwise separable CNN makes the accuracy drop to 96.05% and 96.57%, respectively. Accuracy in the absence of CSBO optimization is 98.23%. The master model has the highest accuracy (99.96%), precision (99.94%), recall (99.95%), F1‐score (99.93%), and specificity (99.92%), which confirms the significance of all the modules.

4.7. Statistical Analysis

A statistical analysis is a methodical process of collecting, summarizing, interpreting, and evaluating the data with the purpose of listing patterns, relations, and tendencies. It assists in determining the significance of the differences available, to be able to estimate the reliability and to be able to make conclusions based on the information. Hypothesis testing, correlation, regression, and variance analysis are the standard methods that will give an objective assessment of the performance and strength, as well as predictive reliability of the model.

The results for statistical test analysis from the proposed ST‐DSCNN‐CSBO method compared to the existing methods are listed in Table 10. The p‐values for paired t‐test are from 0.0031 to 0.0041, whereas for the chi‐square test, the values are from 0.0027 to 0.0035. These results indicate statistically differences between the proposed ST‐DSCNN‐CSBO model and the GGNN, and DST methods. The comparative analysis supports the reliability and effectiveness of the proposed deep learning‐driven IoT‐based water quality forecasting model, demonstrating improved predictive performance compared with the existing techniques.

TABLE 10.

Chi‐square and paired t‐test comparison of existing methods with ST‐DSCNN‐CSBO.

Methods compared with ST‐DSCNN‐CSBO	Paired t‐test (p)	Chi‐square test (p)
GGNN (Z. Li et al. 2024)	0.0032	0.0029
DSTCNN (Arepalli and Naik 2024)	0.0036	0.0031
FSGCN (Qiao et al. 2024)	0.0031	0.0027
DeepTCN‐GRU (Tian et al. 2024)	0.0040	0.0034
VBAED (Bi et al. 2024)	0.0035	0.0030
SBiLSTM (Xu et al. 2024)	0.0033	0.0028
LSTM (Z. Wang, Wang, and Wu 2023)	0.0041	0.0035

4.8. IoT System Reliability and Environmental Robustness Analysis

Even though the given framework is mainly oriented toward predictive modeling, one more assessment was carried out to examine the feasibility of the practical reliability of the IoT sensing infrastructure. Snores, communication latency, and environmental perturbations tend to affect real‐world IoT systems. Thus, the strength of the proposed framework was evaluated based on sensor stability, data transmission latency, and predictions in the case of noisy sensor inputs. Gaussian noise and simulated transmission delays were added to model conditions of the real‐world environment and network conditions.

The reliability results in Table 11 indicate that the IoT‐based monitoring framework maintains stable predictive performance under noisy sensing and communication conditions. Though network delays and environmental noise contribute slightly to the latency and the ratio of prediction error, the accuracy is over 98.8%, indicating that it is highly robust. These findings suggest that the proposed preprocessing and feature extraction solutions are effective in alleviating the effect of sensor noise and transmission variability and support the possibility of the framework in the real‐world IoT‐based water quality monitoring.

TABLE 11.

Reliability evaluation of IoT sensor network and prediction stability.

Condition	Sensor data loss (%)	Average latency (ms)	Prediction accuracy (%)	MAE
Normal sensor operation	0.8	42	99.96	0.008
Moderate environmental noise	2.1	48	99.42	0.011
High environmental noise	3.7	56	98.87	0.015
Network delay simulation	1.5	73	99.21	0.012

Despite the excellent accuracy and R ² obtained from the model, there is an element of doubt that arises from the constrained feature space used (with just four parameters) and the controlled manner in which data is collected. This may make the learning process much easier than it would be in the real world. Further validation of the proposed model on external data and different water sources should therefore be undertaken to verify its ability to generalize to unseen data. To test the generalization ability of the proposed ST‐DSCNN‐CSBO algorithm, an independent water quality dataset named Kaggle Water Potability dataset (publicly available on Kaggle site) was tested as an external validation approach. The dataset is significantly different from the original IoT data in terms of geographical location, sampling, and parameter variation.

The external validation dataset comprises typical measures of water quality, including pH, TDS, and turbidity, among others. To allow easier comparison with the primary data utilized for developing the models, only those typical parameters were included (pH, TDS, turbidity, and temperature scaling).

Although it is not as good as it was for the original set, Table 12 shows that the model's performance is still very strong for the external set, indicating reduced data reliance bias. The decrease in accuracy is due to the increasing data complexity, which proves that the almost perfect performance obtained previously is at least partially due to specific features of the used datasets.

TABLE 12.

External validation performance on independent dataset.

Metric	Value
Accuracy (%)	97.84
Recall (%)	97.71
Precision (%)	97.65
F1‐score (%)	97.68
R 2	0.972
MAE	0.021
RMSE	0.058

4.9. Real‐World Deployment Considerations and Case Study Validation

As illustrated in Table 12, although the accuracy of the proposed model for the external dataset is slightly lower than the previous set, the accuracy rate still shows an excellent result, indicating low data dependency bias. The reason behind the drop in accuracy is that the complexity of data increases, confirming that the outstanding accuracy obtained previously was partly because of the characteristics of the datasets.

As illustrated in Table 13, the proposed model exhibits excellent prediction accuracy even with the introduction of practical noise and delays, with only minor loss in performance. Although these results clearly suggest a high degree of robustness and practical implementation capability, an actual deployment for a longer period of time within multiple geographical areas is required.

TABLE 13.

Case study–based IoT deployment evaluation.

Condition	Data loss (%)	Latency (ms)	Accuracy (%)	MAE
Controlled environment	0.8	42	99.96	0.008
Field‐like noise conditions	2.3	55	99.21	0.012
Communication delay scenario	1.7	71	98.95	0.014

4.10. Feature Importance and Explainability Analysis

In order to improve the interpretability of the suggested model in making decisions about the environment, an assessment of the importance of features was carried out based on SHAP attributions. The method allows us to determine how important each parameter is (pH, TDS, turbidity, and temperature) in determining the WQI and WQC values.

As can be seen from Table 14, the factors that play the most important roles in the prediction procedure are the turbidity level and total dissolved solids, which means they play the biggest roles in changing water quality. The second factor with a fairly significant impact on the prediction process is pH due to its regulative function in chemical interactions, whereas temperature plays a relatively smaller role.

TABLE 14.

Feature importance scores based on SHAP analysis.

Feature	Importance score
TDS	0.34
Turbidity	0.29
pH	0.21
Temperature	0.16

4.11. Scalability Analysis for Large‐Scale IoT Deployment

By simulating an increasing number of IoT sensor nodes and analyzing system latency, throughput, and prediction performance, a quantitative analysis was carried out to assess the scalability of the suggested architecture.

As seen from Table 15, when the number of nodes is increased to 500, there is a more significant increase in the delay time and decrease in the throughput caused by the communication overhead and the process of data aggregation. There is also a minor drop in the accuracy of the results owing to higher data variability, which means that the scalability of the proposed model is good.

TABLE 15.

Scalability evaluation with increasing IoT nodes.

Number of sensors	Avg latency (ms)	Throughput (samples/s)	Accuracy (%)	MAE
10	38	520	99.96	0.008
50	46	505	99.82	0.009
100	58	488	99.67	0.010
200	72	462	99.41	0.012
500	109	421	98.96	0.016

4.12. Discussion

The proposed ST‐DSCNN‐CSBO model showed excellent results in predicting water quality, with highly accurate and reliable predictive results for WQI and WQC. The dimensionality reduction applied during preprocessing by QAFMMN and feature selection by AOA‐NO was effective in terms of retaining vital information and reducing dimensionality. The hybrid ST‐DSCNN also managed to recognize the intricate temporal and spatial associations in sensor data, and this aspect also led to the enhancement of prediction accuracy. In addition, the network parameters were also optimized using CSBO so that the bias could be reduced and network learning could be robust and stable. The proposed solution comes in handy for processing large amounts of data with IoT sensors when it comes to giving consistent and reliable predictions. It has the advantages of a higher feature representation, reduced computational complexity, superior convergence, and the capability to react to various patterns of water quality. Overall, ST‐DSCNN‐CSBO is an innovative, scalable, and effective solution for real‐time monitoring and prediction of water quality when using IoT.

Apart from the predictive performance of the algorithm, the obtained results present valuable information regarding the environment by highlighting the significance of the analyzed parameters for the process. According to the results, TDS and turbidity demonstrate greater effects on WQI variation, which means that dissolved substances and particulate matter play a critical role in the deterioration of the studied waters' quality. At the same time, there were no notable changes in the temperature and pH values of the samples, indicating that these physicochemical parameters remained relatively stable throughout the study period. As far as environmental issues are concerned, increased TDS values are usually associated with either higher water salinity or excessive amounts of minerals; therefore, such water samples cannot be recommended for irrigation purposes. Increased turbidity may also indicate that the studied bodies of water have been contaminated. As a result, reduced light exposure affects the functioning of aquatic organisms and interferes with their ability to survive. Lastly, the consistency of pollution classification across all of the WQC classes implies that the developed model allows capturing pollution transitions smoothly.

Because the suggested ST‐DSCNN‐CSBO model performs well, overfitting is not indicated. The training and testing curves have nearly identical convergence dynamics without significant leaps, indicating that the model is training with stable generalization. The dataset was divided using the 80:20 train–test split. QAFMMN preprocessing eliminates sensor noise, and the AOA‐NO feature selection eliminates duplicate attributes and constrains the model complexity. Moreover, depthwise separable convolution makes fewer parameters trainable, and CSBO optimization stabilizes the convergence. All these mechanisms eliminate memorization and facilitate effective generalization in both datasets.

The results of the experiments prove that the suggested framework provides very precise and consistent predictions for the considered datasets, which proves the success of the combined Swin‐Transformer, separable CNN framework, and optimization based on the use of CSBO. The model has robust generalization and can effectively manage the extracted feature representations, hence its relevance to the application of the system to IoT‐assisted monitoring settings. Moreover, the architecture can be scaled computationally because the transformer‐based extraction of features and the lightweight separable convolution are computationally efficient, and both can be used to efficiently process larger datasets and distributed streams of IoT data. Despite the likely presence of sensor variability, environmental, and implementation‐specific factors in real‐world applications, the general framework will offer a flexible and adaptable framework that can be further expanded by adding new data sources and calibration processes. These observations reveal the practical possibilities of the proposed system to support reliable, large‐scale monitoring and intelligent decision‐support applications.

The superior accuracy achieved by the proposed ST‐DSCNN‐CSBO framework is mainly attributed to the integration of quality‐aware preprocessing, optimized feature selection, hybrid transformer–CNN feature learning, and adaptive bias optimization. The QAFMMN stage effectively reduces sensor noise and improves data consistency, whereas AOA‐NO selects the most informative attributes and removes redundant features. Furthermore, the combination of Swin‐Transformer and DSCNN enables simultaneous learning of long‐range contextual dependencies and local spatial patterns from IoT sensor data. The inclusion of CSBO‐based bias optimization further improves convergence stability and generalization capability, leading to enhanced predictive performance compared with existing approaches.

5. Conclusions

The proposed ST‐DSCNN‐CSBO model is a powerful and robust model in monitoring and predicting water quality on the basis of IoT. The method ensures the use of the most informative and relevant features, which are provided by preprocessing according to QAFMMN and feature selection with the assistance of AOA‐NO, and reduces the dimensionality of data without losing the key information. The ST‐DSCNN also proves to be effective in capturing the temporal and spatial relationships in the data of the water quality sensor, which makes it more predictable. Bias optimization in the network parameters is also performed with the use of CSBO, such that convergence is strong and generalization is increased. The approach possesses several significant strengths, including the fact that it correctly predicts both WQI and WQC, it can predict large quantities of IoT data, and it is resilient to noisy or missing data. Its weakness, however, is that it is less efficient in real‐time deployments as it demands more processing capacity compared to single integrated modules and optimization steps. Nevertheless, ST‐DSCNN‐CSBO proves to be more capable than the traditional and existing deep learning models. Further research is aimed at extending the framework to contamination prediction, which will allow monitoring hazardous pollutants proactively and improve environmental and population health.

Author Contributions

Navala Rani Gunti: conceptualization, funding acquisition, writing – original draft, validation, methodology, formal analysis, resources, data curation, software. Kodukula Subrahmanyam: supervision, resources, project administration, visualization, writing – review and editing, writing – original draft, investigation, data curation.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The data that support the findings of this study can be provided upon request from the author.