Machine Learning for Surface Water Quality Management in Malaysia: Integrating Physicochemical Parameters, Nutrient Loads, Land Use, Hydrological Conditions, and Anthropogenic Pressure

 

1.Introduction
Surface water quality has been substantially affected during recent decades by the discharge of domestic wastewater [1], industrial effluents, agricultural pollutants, nutrients, suspended materials, organic contaminants, and other wastes [2]. Deteriorating water quality creates risks for human health, aquatic biodiversity, ecosystem functioning, water supply, irrigation, fisheries, and other beneficial uses [3,4,5]. The problem becomes more complex when climate variability modifies rainfall patterns, river discharge, water temperature, dilution capacity, and pollutant transport [6,7].
Sustainable surface-water management therefore requires systematic identification of the principal pollution sources and environmental pressures affecting rivers and other freshwater systems [8,9]. Traditional water-quality management relies heavily on periodic field sampling and laboratory analysis [10]. Although these procedures provide reliable measurements, their temporal and spatial coverage may be insufficient for rapidly changing pollution conditions [11]. Real-time sensors can improve temporal resolution, but their deployment can involve substantial acquisition, maintenance, calibration, and data-quality requirements [12].
Machine learning (ML) provides an alternative or complementary approach because it can identify nonlinear relationships among multiple environmental variables and transform large monitoring datasets into predictions, classifications, and early-warning information [13]. Previous research has demonstrated ML applications in water-quality classification, water-quality prediction, anomaly detection, pollution-source identification, and hydrological forecasting [14,15].
Malaysia provides a particularly relevant setting for this development. The Department of Environment (DOE) operates a national framework for river-water-quality standards and the River Water Quality Index [16]. Malaysian river waters are classified into Classes I, IIA, IIB, III, IV, and V according to designated uses and water-quality conditions [17]. Class I includes conservation and very limited-treatment uses, while Classes II–V correspond to progressively different treatment, fishery, recreational, livestock, irrigation, and other uses [18].
The Malaysian WQI framework also provides a direct basis for machine-learning classification. According to DOE documentation, WQI values of 81–100 are classified as clean, 60–80 as slightly polluted, and 0–59 as polluted [19]. The Malaysian classification framework incorporates parameters including ammoniacal nitrogen, biochemical oxygen demand, chemical oxygen demand, dissolved oxygen, pH, and total suspended solids [20].
Previous Malaysian research has already demonstrated the potential of ML for this purpose. Shamsuddin et al. developed ANN, Decision Tree, and Support Vector Machine models for multiclass water-quality classification in the Langat River Basin and reported that SVM provided strong classification performance [21]. Earlier research on the Langat River also combined chemometric techniques and ANN using extensive physicochemical measurements, identifying dissolved oxygen, BOD, pH, ammoniacal nitrogen, chlorine, E. coli, and coliform as important variables in explaining spatial water-quality variation [22].
However, a surface-water management system based only on conventional WQI parameters may not fully represent the processes controlling water quality. River condition is influenced not only by measured physicochemical properties but also by rainfall, discharge, catchment land use, agricultural activity, industrialization, urban expansion, wastewater inputs, and population pressure. Therefore, a more comprehensive ML framework should integrate these variables [23].
The present study develops such a framework for Malaysia. Rather than treating water quality as a function of water chemistry alone, the proposed framework conceptualizes surface-water quality as the product of interactions among five major dimensions: physicochemical conditions, nutrient loading, land use, hydrological conditions, and anthropogenic pressure.
The objectives are therefore to:
    review the major ML algorithms applicable to surface-water-quality management;
    establish a Malaysia-oriented set of physicochemical, nutrient, hydrological, land-use, and anthropogenic variables;
    integrate Malaysian WQI and national water-quality classes into an ML decision framework;
    evaluate the potential roles of DT, RF, SVM/SVR, ANN, LSTM, and ensemble models;
    establish a framework for WQI classification, continuous water-quality prediction, and anomaly detection;
    introduce explainable ML for identifying the environmental variables that most strongly influence model predictions; and
    identify priorities for real-time, multi-source, and sustainable surface-water management in Malaysia.
2. Conceptual Framework for Surface Water Quality Management
2.1 Water Quality as a Multidimensional Environmental System
Water quality is a multidimensional concept involving physical, chemical, biological, and microbiological characteristics [24]. The conventional monitoring approach generally emphasizes direct measurements in water samples [25]. However, the observed value of a water-quality parameter can be influenced by hydrological conditions, upstream land use, pollution sources, meteorological conditions, and seasonal processes [26].
Consequently, the proposed ML framework represents surface-water quality through five interacting groups [27,28,29,30]:
Group 1: Physicochemical parameters
    pH
    dissolved oxygen (DO)
    water temperature (WT)
    electrical conductivity (EC)
    turbidity
    total suspended solids (TSS)
    total dissolved solids (TDS)
    biochemical oxygen demand (BOD)
    chemical oxygen demand (COD)
Group 2: Nutrient parameters
    ammoniacal nitrogen (NH₃-N)
    nitrate (NO₃⁻)
    total nitrogen (TN)
    phosphate
    total phosphorus (TP)
Group 3: Hydrological and meteorological parameters
    rainfall intensity
    cumulative rainfall
    antecedent rainfall
    streamflow/discharge
    water level
    seasonal conditions
    antecedent dry period
Group 4: Land-use parameters
    urban/built-up land
    agricultural land
    industrial land
    forest land
    impervious-surface proportion
    road density
Group 5: Anthropogenic pressure
    population density
    wastewater discharge
    industrial intensity
    agricultural intensity
    urban expansion
    proximity to pollution sources
This structure allows ML models to move beyond a purely chemical description of water quality and toward a catchment-scale representation of pollution processes.
2.2 Water Quality Index and Classification
The Water Quality Index (WQI) [31] provides a numerical representation of overall water condition by integrating multiple parameters into a single index.
A general weighted formulation can be expressed as:
WQI=(∑_(i=1)^n▒q_i  w_i)/(∑_(i=1)^n▒w_i )
where:
    q_iis the sub-index value of parameter i;
    w_iis the corresponding parameter weight; and
    nis the number of parameters.
The conventional WQI approach is useful because it transforms multiple measurements into a common management indicator. However, the calculation requires measurements for multiple parameters and may become computationally demanding when monitoring frequency and spatial coverage increase.
ML can complement this process by learning the relationship between available input parameters and WQI or WQI class. The model can subsequently estimate water-quality condition from a reduced or expanded set of observations.
Table 1. Malaysian Water Quality Index Classification
WQI range    DOE classification    General interpretation
81–100    Clean    Relatively good water quality
60–80    Slightly polluted    Moderate deterioration
0–59    Polluted    Significant pollution
Source: Malaysian Department of Environment water-quality classification.
Malaysia additionally uses Classes I, IIA, IIB, III, IV, and V to link water-quality conditions with designated uses[32].
3. Machine Learning Algorithms for Surface Water Quality Management
3.1 Decision Tree
Decision Tree (DT) is a supervised-learning algorithm used for both classification and regression. It recursively divides observations according to selected predictor variables until terminal nodes representing predicted classes or numerical values are generated [33].
DT is particularly attractive for environmental applications because its decision structure can be interpreted comparatively easily. For example, a tree may identify combinations of low DO, high BOD, elevated ammoniacal nitrogen, and high TSS as conditions associated with degraded water-quality classes [34].
However, an individual tree can become excessively complex and overfit the training dataset. Pruning, cross-validation, and ensemble methods can therefore be used to improve generalization [35].
3.2 Random Forest
Random Forest (RF) combines multiple decision trees constructed using different bootstrap samples and subsets of predictor variables [36].
For water-quality management, RF has several advantages [37]:
    ability to model nonlinear relationships;
    capacity to accommodate many predictors;
    relatively strong resistance to overfitting compared with individual trees;
    ability to estimate variable importance;
    suitability for classification and regression; and
    applicability to mixed environmental datasets.
RF can therefore be used to predict WQI, classify water-quality classes, estimate individual pollutant concentrations, or identify the relative contribution of hydrological, land-use, and anthropogenic variables [38].
3.3 Support Vector Machine and Support Vector Regression
Support Vector Machine (SVM) is a supervised-learning method based on identifying an optimal decision boundary between classes. Kernel functions allow SVM to represent nonlinear relationships [39].
For continuous water-quality prediction, the corresponding Support Vector Regression (SVR) framework can be applied [40].
SVM/SVR is especially useful when [41]:
    datasets are relatively small;
    predictors are highly nonlinear;
    the number of variables is relatively high;
    generalization is important; and
    complex boundaries exist between water-quality classes.
A Malaysian study of the Langat River Basin demonstrated the applicability of SVM, ANN, and DT to multiclass WQI classification [42].
3.4 Artificial Neural Network
Artificial Neural Networks (ANNs) consist of interconnected computational nodes organized into input, hidden, and output layers [43].
For surface-water-quality management, ANN can learn nonlinear relationships between environmental variables and WQI or individual water-quality parameters [44].
For example:
WQI=f(Dⓜ,Oⓜ,Bⓜ,Oⓜ,Dⓜ,Cⓜ,Oⓜ,Dⓜ,pⓜ,HNH_3-NTSSRainfallFlowLandUseAnthropogenicPressure)
The major advantage of ANN is its capacity to approximate highly nonlinear relationships [45]. However, ANN generally requires adequate training data, appropriate architecture selection, parameter tuning, and computational resources. Its internal decision process may also be difficult for environmental managers to interpret [46].
3.5 Long Short-Term Memory Networks
Water quality is inherently temporal. Measurements observed today may depend on conditions during previous hours, days, or weeks [47].
Long Short-Term Memory (LSTM) networks are a specialized form of recurrent neural network designed to learn temporal dependencies [48].
An LSTM-based framework can be represented conceptually as:
WQI_t=f(X_tⓜ,X_(t-1)ⓜ,.ⓜ,..X_(t-k) )
where X_trepresents the environmental variables observed at time t, and krepresents the historical time window.
Potential LSTM inputs include [49,50]:
    previous DO values;
    previous BOD/COD;
    rainfall;
    cumulative rainfall;
    streamflow;
    water level;
    nutrient concentrations;
    temperature; and
    previous WQI.
This makes LSTM particularly relevant for early-warning systems.
4. Proposed Malaysian Surface-Water ML Architecture
The proposed framework contains six sequential stages.
Stage 1: Multi-source data acquisition
Data should be collected from:
    DOE monitoring stations;
    automated water-quality sensors;
    meteorological stations;
    hydrological stations;
    GIS databases;
    satellite-derived land-use products;
    population databases;
    industrial databases; and
    wastewater-discharge records.
Stage 2: Data preprocessing
The dataset should undergo:
    quality control;
    missing-value treatment;
    outlier identification;
    temporal synchronization;
    spatial matching;
    normalization where required;
    feature selection; and
    class-balance assessment.
Stage 3: Feature integration
The five predictor groups should be merged into a unified feature matrix.
Stage 4: Model development
At least four model families should be compared:
    DT;
    RF;
    SVM/SVR;
    ANN/LSTM.
Ensemble models may subsequently combine the strongest models.
Stage 5: Model interpretation
SHAP and LIME should be used to identify the contribution of each predictor.
Stage 6: Management application
The final system should produce:
    WQI;
    water-quality class;
    pollutant predictions;
    anomaly alerts;
    pollution-pressure indicators; and
    management recommendations.
5. Proposed Variables and Their Environmental Roles
Table 2. Integrated predictor framework
Dimension    Variables    Expected management relevance
Physicochemical    pH, DO, WT, EC, turbidity, TSS, TDS    Direct representation of water condition
Organic pollution    BOD, COD    Organic and chemical pollution
Nutrients    NH₃-N, NO₃⁻, TN, phosphate, TP    Nutrient enrichment and eutrophication
Hydrology    rainfall, discharge, water level    Pollutant dilution and transport
Climate    temperature, rainfall intensity, antecedent rainfall    Seasonal and event-driven changes
Land use    urban, agriculture, forest, industrial    Spatial pollution pressure
Urbanization    built-up area, impervious surface    Runoff and pollutant loading
Anthropogenic    population density, wastewater, industrial intensity    Human-generated pressure
Transport    road density, drainage connectivity    Pollutant mobilization
Management    WQI, WQC class    Final management indicators
The major methodological innovation of the proposed framework is therefore not simply replacing statistical models with ML. It is the integration of water chemistry with catchment-scale environmental drivers.
________________________________________
6. Machine Learning for Water-Quality Classification
Water-quality classification is one of the most established ML applications in surface-water management.
A classification model can estimate:
Y∈{Cⓜ,lⓜ,eanSlightly" " PollutedPolluted}
or, at the Malaysian regulatory level:
Y∈{Iⓜ,Iⓜ,Iⓜ,Aⓜ,Iⓜ,IBIIIIVV}
The input vector can contain both conventional water-quality parameters and contextual environmental variables.
Table 3. Candidate classification models
Model    Main application    Major advantage    Main limitation
DT    WQI classification    Easy interpretation    Overfitting
RF    Multiclass WQI    Robust nonlinear modeling    Lower transparency
SVM    WQI classification    Strong nonlinear classification    Parameter tuning
ANN    Complex classification    Nonlinear learning    Requires sufficient data
KNN    Similarity classification    Simple implementation    Sensitive to scale/outliers
XGBoost    Ensemble classification    High predictive capacity    Hyperparameter complexity
LSTM    Temporal classification    Learns temporal dependence    Data and computational demand
Existing Malaysian evidence supports the feasibility of this approach. Research in the Langat River Basin used ANN, DT, and SVM for multiclass water-quality classification, while other Malaysian work has investigated ANN-based WQI prediction.
________________________________________
7. Machine Learning for Continuous Water-Quality Prediction
Classification alone does not provide sufficient information for proactive management. Environmental agencies also require estimates of future pollutant concentrations and WQI.
A continuous prediction model can be formulated as:
Y ̂_(t+h)=f(X_tⓜ,X_(t-1)ⓜ,.ⓜ,..X_(t-k) )
where:
    Y ̂_(t+h)is the predicted water-quality variable;
    his the prediction horizon;
    X_tis the current environmental state; and
    krepresents historical observations.
Potential prediction targets include:
    DO;
    BOD;
    COD;
    NH₃-N;
    NO₃⁻;
    TP;
    TSS;
    turbidity;
    WQI; and
    water-quality class.
The inclusion of hydrological variables is particularly important because rainfall and discharge can alter pollutant concentration through dilution, mobilization, and transport.
8. Nutrient-Load Modeling
Nutrient pollution represents an important extension of conventional WQI-based modeling[51].
Nitrogen and phosphorus can originate from[52]:
    agricultural fertilizers;
    livestock operations;
    domestic wastewater;
    industrial wastewater;
    urban runoff;
    soil erosion; and
    atmospheric deposition.
The proposed ML framework therefore includes:
N_load=f(Nⓜ,H_3ⓜ,-ⓜ,Nⓜ,Nⓜ,O_3^- TNRainfallFlowLandUse)
and
P_load=f(Pⓜ,O_4^(3-)ⓜ,Tⓜ,Pⓜ,RainfallFlowLandUse)
These relationships can be used to identify periods and locations where nutrient loading is likely to increase.
Importantly, concentration and load should not be treated as identical variables. Concentration describes pollutant abundance per unit water volume, whereas load incorporates flow and therefore provides information about the total pollutant mass transported through a river system[53].
9. Land Use and Spatial Water-Quality Modeling
Land use determines the spatial distribution of many pollution pressures[54].
Urban areas may contribute[55]:
    road runoff;
    suspended solids;
    metals;
    hydrocarbons;
    nutrients; and
    wastewater-related pressures.
Agricultural areas may contribute[56]:
    nitrogen;
    phosphorus;
    pesticides;
    sediments; and
    organic matter.
Industrial areas may contribute[57]:
    chemical pollutants;
    metals;
    organic contaminants; and
    elevated thermal or conductivity signals.
Forested areas generally represent a different hydrological and pollution-pressure regime[58].
Accordingly, GIS-derived land-use variables should be incorporated into ML models rather than treating every monitoring station as environmentally independent.
Potential spatial features include:
UrbanRatio=A_urban/A_catchment 
AgriculturalRatio=A_agriculture/A_catchment 
IndustrialRatio=A_industrial/A_catchment 
and:
ImperviousRatio=A_impervious/A_catchment 
These variables can be calculated for different buffer distances around monitoring stations or upstream subcatchments.
10. Hydrological Conditions and Water Quality
Hydrology provides an essential link between pollution sources and observed water quality[59].
Rainfall can increase surface runoff and transport pollutants from urban and agricultural surfaces into rivers. Conversely, prolonged low-flow conditions can reduce dilution capacity and increase the concentration of some pollutants[60].
The proposed framework therefore incorporates:
    current rainfall;
    cumulative rainfall;
    24-h rainfall;
    72-h rainfall;
    antecedent precipitation;
    discharge;
    water level;
    flow velocity where available; and
    dry-period duration.
An antecedent rainfall index may be represented conceptually as:
API_t=P_t+kP_(t-1)+k^2 P_(t-2)+...+k^n P_(t-n)
where P_tis rainfall and kis a recession coefficient.
The inclusion of these variables allows the ML model to distinguish between pollution generated under normal flow and pollution mobilized during rainfall events.
11. Anthropogenic Pressure Index
A further extension of the framework is the construction of an Anthropogenic Pressure Index (API*) based on normalized human-pressure variables[61].
A general formulation can be expressed as:
API^*=∑_(j=1)^m▒w_j  Z_j
where Z_jrepresents a normalized anthropogenic-pressure variable and w_jrepresents its assigned or data-derived weight.
Potential components include:
    population density;
    industrial density;
    wastewater discharge;
    agricultural intensity;
    built-up proportion;
    road density; and
    proximity to wastewater outlets.
The purpose of this index is not to replace direct pollution measurements. Rather, it provides a contextual representation of human pressure that can improve spatial prediction and source interpretation.
12. Explainable Machine Learning
High predictive accuracy alone is insufficient for environmental decision-making. Managers need to know why a model predicts deterioration[62].
SHAP (SHapley Additive exPlanations) can quantify the contribution of individual variables to predictions[63].
For example, a SHAP analysis may reveal that[64,65]:
    rainfall has a strong positive contribution to TSS;
    high BOD contributes negatively to WQI;
    elevated DO contributes positively to WQI;
    urban land proportion increases pollution probability;
    ammoniacal nitrogen strongly influences WQI class.
LIME can provide local explanations for individual predictions.
The proposed framework therefore combines:
Prediction+Explanation+Environmental" " Interpretation
rather than prediction alone.
13. Water-Quality Anomaly Detection
Pollution incidents do not necessarily correspond to a gradual decline in WQI. Sudden changes can indicate[66,67]:
    accidental discharge;
    wastewater-system failure;
    intense storm runoff;
    illegal discharge;
    sensor malfunction; or
    unusual hydrological conditions.
Unsupervised ML can detect unusual observations without requiring a complete labeled database.
Potential methods include[68,69]:
    Isolation Forest;
    One-Class SVM;
    clustering;
    autoencoders;
    Bayesian approaches; and
    LSTM-based temporal anomaly detection.
An anomaly score can be used to generate early warnings:
A_t=f(X_t-X ̂_t )
where X_trepresents observed conditions and X ̂_trepresents expected conditions.
14. Performance Evaluation
The proposed models should be evaluated using metrics appropriate to the prediction task.
For regression:
MAE=1/n ∑_(i=1)^n▒∣ y_i-y ̂_i∣
RMSE=√(1/n ∑_(i=1)^n▒( ├ y_i-y ̂_i ├ )┤^2 ┤ )
R^2=1-(∑▒( ├ y_i-y ̂_i ├ )┤^2 ┤)/(∑▒( ├ y_i-y ̅├ )┤^2 ┤ )
For classification:
    accuracy;
    precision;
    recall;
    F1-score;
    balanced accuracy;
    Cohen's kappa; and
    confusion matrix.
For imbalanced water-quality classes, accuracy alone should not be considered sufficient. Macro-F1 and class-specific recall should also be reported.
15. Validation Strategy
Random train-test splitting can produce overly optimistic results when environmental observations are temporally or spatially dependent.
Therefore, the Malaysian framework should use:
Temporal validation
Training on earlier observations and testing on later observations.
Spatial validation
Training on some monitoring stations and testing on geographically separate stations.
Catchment-level validation
Training on selected catchments and testing on an independent catchment.
Extreme-event validation
Testing specifically during high-rainfall and unusual pollution episodes.
These procedures provide stronger evidence of whether an ML model can generalize beyond the data used for model development.
16. Proposed Integrated Decision-Support Framework
Figure 1. Conceptual architecture of the proposed Malaysian ML system
Data sources
DOE monitoring + IoT sensors + rainfall + hydrology + GIS + land use + population + industrial/wastewater data
↓
Data preprocessing
Quality control → missing-value treatment → anomaly screening → normalization → temporal/spatial alignment
↓
Feature integration
Physicochemical + nutrient + hydrological + land-use + anthropogenic variables
↓
Machine-learning layer
DT | RF | SVM/SVR | ANN | LSTM | XGBoost | Hybrid models
↓
Outputs
WQI prediction + WQC classification + pollutant prediction + anomaly detection
↓
Explainability
SHAP + LIME + variable importance
↓
Management
Early warning + pollution-source prioritization + monitoring optimization + water-quality protection
17. Comparative Interpretation of Machine Learning Approaches
Table 4. Recommended role of ML models in the Malaysian framework
Model    Classification    Regression    Time series    Explainability    Recommended role
DT    High    High    Low    High    Transparent baseline
RF    High    High    Moderate    Moderate–High    Main ensemble model
SVM    High    High    Moderate    Moderate    Small/medium datasets
ANN    High    High    Moderate    Low    Nonlinear relationships
LSTM    Moderate    High    Very high    Low    Temporal forecasting
XGBoost    High    High    Moderate    Moderate–High    High-performance tabular data
KNN    Moderate    Moderate    Low    High    Benchmark/similarity
Hybrid models    High    High    Very high    Variable    Advanced prediction

18. Evidence from Malaysian Applications
The proposed framework is consistent with the direction of existing Malaysian research.
The Langat River Basin has been used extensively as a Malaysian case for data-driven water-quality analysis. Earlier research analyzed 23 physicochemical parameters from seven monitoring stations and used multivariate techniques and ANN to distinguish pollution regions[70,71,72,73,74,75,76,77,78,79].
More recent research developed ANN, DT, and SVM models for multiclass WQI classification in the Langat River Basin. The study showed that ML can reduce the computational complexity associated with conventional WQI classification and demonstrated the practical relevance of SVM-based classification.
Other Malaysian research has developed ANN models for WQI estimation in the Langat River, while ML studies have also investigated SVM-based water-quality prediction for catchments where direct observations may be limited.
These studies provide an empirical basis for applying ML to Malaysian river management. Nevertheless, the proposed framework extends the scope by integrating additional predictors representing land use, hydrology, nutrient loading, and anthropogenic pressure.
19. Advantages of the Proposed Framework
The integrated approach offers several potential advantages.
First, it can reduce dependence on a limited set of conventional water-quality variables by incorporating contextual environmental information.
Second, it can provide predictions between conventional sampling events.
Third, it can combine spatial and temporal information.
Fourth, it can identify variables associated with deterioration through feature-importance and explainability methods.
Fifth, it can support early-warning systems for sudden pollution events.
Sixth, it can improve monitoring efficiency by identifying locations and periods where additional sampling is most valuable.
Finally, it can provide a common analytical architecture that can potentially be transferred among Malaysian catchments after appropriate local calibration.
20. Limitations and Research Challenges
Despite its potential, several limitations must be addressed.
20.1 Data availability
ML performance depends heavily on the quantity, quality, temporal resolution, and representativeness of training data.
20.2 Missing observations
Environmental datasets frequently contain missing observations resulting from sensor failure, sampling interruptions, laboratory limitations, or changes in monitoring programs.
20.3 Spatial heterogeneity
Different Malaysian catchments have different hydrological, geological, climatic, land-use, and anthropogenic characteristics. A model trained in one basin should therefore not automatically be assumed to generalize to another.
20.4 Concept drift
Water-quality relationships can change because of urban expansion, climate variability, wastewater-treatment improvements, industrial changes, or land-use conversion.
20.5 Black-box models
Deep-learning models can provide strong predictive performance while remaining difficult to interpret.
20.6 Data leakage
Variables used to calculate WQI should be carefully distinguished from independent predictors when evaluating ML models intended to predict WQI. Otherwise, performance may be artificially inflated.
20.7 Regulatory compatibility
ML outputs should complement rather than replace official Malaysian regulatory procedures unless formally validated and adopted by the relevant authorities.
21. Future Research Directions
Future Malaysian studies should focus on seven priorities.
21.1 Multi-source data fusion
The integration of DOE observations, IoT sensors, remote sensing, GIS, rainfall, hydrology, and socioeconomic data can produce a more comprehensive representation of catchment conditions.
21.2 Hybrid models
Hybrid architectures combining RF/XGBoost with LSTM or CNN-LSTM may be particularly useful for datasets containing both spatial and temporal patterns.
21.3 Explainable AI
SHAP, LIME, partial-dependence analysis, and feature-interaction analysis should become standard components of environmental ML studies.
21.4 Transfer learning
Models trained in data-rich Malaysian basins could potentially be adapted to data-limited basins through transfer-learning approaches, subject to rigorous validation.
21.5 Real-time monitoring
The integration of IoT sensors with ML could allow continuous water-quality assessment and automated warning generation.
21.6 Uncertainty quantification
Future models should provide prediction intervals rather than only point predictions.
21.7 Climate-change integration
Rainfall extremes, temperature changes, drought periods, and changing flow regimes should be incorporated into long-term water-quality prediction frameworks.
22. Management Implications for Malaysia
The proposed framework can support environmental management at several levels.
At the monitoring level, ML can identify which variables and stations require more frequent measurements.
At the operational level, real-time models can provide early warnings.
At the catchment level, land-use and anthropogenic variables can identify areas where pollution pressure is concentrated.
At the regulatory level, predicted WQI and water-quality classes can provide supplementary information for prioritizing investigation.
At the strategic level, long-term models can evaluate relationships between urbanization, climate variability, nutrient loading, and river-water quality.
Malaysia's existing DOE standards provide an appropriate regulatory reference because they link water-quality measurements to water-use classes and WQI categories.
23. Conclusions
Machine learning has substantial potential to strengthen surface-water-quality management in Malaysia by transforming conventional monitoring data into classification, prediction, anomaly-detection, and early-warning systems.
The principal contribution of the proposed framework is the integration of five dimensions of environmental information: physicochemical parameters, nutrient loads, land use, hydrological conditions, and anthropogenic pressure. This multidimensional approach recognizes that observed river-water quality is not determined exclusively by the chemical properties of a water sample but emerges from interactions among pollution sources, catchment characteristics, hydrological processes, climate conditions, and human activities.
DT, RF, SVM/SVR, ANN, LSTM, and ensemble algorithms each have different strengths. Tree-based methods provide strong performance on heterogeneous tabular datasets and can offer useful feature-importance information. SVM can be effective for nonlinear classification and relatively limited datasets. ANN can represent complex nonlinear relationships, while LSTM is particularly suitable for temporal forecasting. Hybrid models may provide additional advantages when spatial, temporal, and nonlinear processes interact.
The Malaysian regulatory context provides a strong foundation for ML development. The DOE framework uses WQI and water-use classes to characterize river-water quality, while existing Malaysian research has already demonstrated the feasibility of ANN, DT, SVM, and related approaches in the Langat River Basin.
The next generation of Malaysian water-quality ML systems should therefore move beyond models based solely on conventional WQI parameters. Integrating rainfall, streamflow, antecedent hydrological conditions, land-use composition, urbanization, agricultural activity, industrial pressure, population density, and wastewater information can provide a more complete representation of the mechanisms controlling surface-water quality.
At the same time, predictive accuracy should not be the only criterion for model selection. Explainability, uncertainty, spatial transferability, temporal robustness, data quality, regulatory compatibility, and operational feasibility are equally important. SHAP and LIME can help transform complex ML models into more interpretable decision-support tools by showing how individual environmental variables contribute to predictions.
A mature Malaysian surface-water ML system should ultimately connect monitoring → prediction → explanation → early warning → management response. Such an integrated architecture can complement conventional environmental monitoring, improve the efficiency of water-quality assessment, and provide a scientifically grounded foundation for proactive surface-water management.
The proposed framework is therefore suitable as a methodological basis for future empirical studies using Malaysian DOE monitoring records, IoT observations, hydrological datasets, remote-sensing products, and GIS-derived catchment characteristics. Its final empirical validation should be performed using independently collected Malaysian observations and should include spatial, temporal, and event-based validation to establish the reliability and generalizability of the developed models.


Ethical Considerations
Not applicable. This study did not require ethical approval because it does not include human or animal subjects and does not involve any personal or sensitive data.

List of Abbrevations:
(WQI): The Water Quality Index ; (ML): Machine learning;(DOE): Department of Environment;  (DO): dissolved oxygen; (WT): water temperature ; (EC): electrical conductivity; (TSS): total suspended solids ; (TDS): total dissolved solids; (BOD): biochemical oxygen demand; (COD): chemical oxygen demand; (NH₃-N): ammoniacal nitrogen; (NO₃⁻): nitrate; (TN): total nitrogen; (TP): total phosphorus;  (DT) : Decision Tree; (RF) : Random Forest;(SVM): Support Vector Machine ; (SVR) : Support Vector Regression;  (ANNs): Artificial Neural Networks; (LSTM): Long Short-Term Memory.

Acknowledgment: 
The authors would like to express their sincere gratitude to the Faculty of Engineering and Technology, Multimedia University, Ayer Keroh 75450, Melaka, Malaysia, for providing partial research funding under Grant No. Malaysia.grant.2026.PTGF.26YH, which supported the conduct of this study. The authors also gratefully acknowledge The International Journal of Engineering Sciences – Noor Al-Ilm for Publishing and Distribution for its generous support in providing a full waiver of the publication fees and facilitating the publication of this manuscript free of charge. The authors highly appreciate the institution's and journal's commitment to supporting scientific research and advancing opportunities for researchers.

Author Contribution: 
All authors contributed equally to the main contributor to this paper. All authors read and approved the final paper.


Declaration of generative AI and AI-assisted technologies in the writing process
The authors hereby declare that no generative artificial intelligence or AI-assisted technologies were used at any stage during the preparation of this manuscript, including language editing, proofreading, or content development. The authors take full responsibility for the originality and integrity of the work presented in this publication.

Funding: 
This research received partial financial support from the Faculty of Engineering and Technology, Multimedia University, Ayer Keroh 75450, Melaka, Malaysia, under Grant No. Malaysia.grant.2026.PTGF.26YH, to support the conduct of the present study. The authors also acknowledge The International Journal of Engineering Sciences, Noor Al-Ilm for Publishing and Distribution, for providing a full waiver of the publication fees. The publication fee waiver was provided as editorial and publication support and did not involve any financial contribution to, or influence on, the research design, data collection, analysis, interpretation of results, or reporting of the research.


Conflicts of Interest: 
“The authors declare no conflict of interest.”