Artificial Intelligence for Water Quality Improvement in Bangladesh: A Review of Methods, Applications, and an Integrated Roadmap for Arsenic, Salinity, and River Pollution

1. Introduction

Access to safe water is a foundational determinant of public health and economic development, yet for Bangladesh it remains precarious. The country sits on the world’s largest delta, formed by the Ganges, Brahmaputra (Jamuna), and Meghna rivers, and is simultaneously water-abundant and water-insecure. Three problems dominate. First, the shift from microbially unsafe surface water to groundwater tube wells in the late twentieth century unintentionally exposed the population to naturally occurring arsenic, in what has been described as one of the largest mass poisonings in history. Second, sea-level rise, reduced dry-season river flow, and the expansion of brackish-water shrimp aquaculture have driven salinity intrusion deep into the coastal southwest, contaminating both surface and groundwater. Third, rapid, poorly regulated industrialisation has turned the rivers around Dhaka — the Buriganga, Turag, Balu, and Tongi Khal — into heavily polluted channels with dissolved oxygen near zero in the dry season [1]. Figure 1 locates these three challenge zones.

Managing these problems demands dense, timely, and spatially comprehensive information that conventional sampling cannot provide. Laboratory analysis of arsenic or salinity across millions of wells is infeasible to repeat frequently, and the physico-chemical drivers are highly non-linear, seasonal, and spatially heterogeneous. Artificial intelligence offers a route to extend limited measurements: by learning the relationship between cheaply observed predictors and expensive or future targets, ML models can map contamination risk, forecast seasonal deterioration, and support real-time decisions [1,3]. A growing body of Bangladesh-specific work — spanning arsenic prediction in the northwest [3], coastal-aquifer modelling [4], and WQI prediction for Dhaka’s rivers [1] — demonstrates that these methods are both accurate and locally relevant.

This paper has three objectives: (i) to characterise the three water-quality challenges of Bangladesh and the data landscape available for modelling them; (ii) to review, with their governing equations, the AI methods now being applied and the performance they achieve; and (iii) to propose an integrated, interpretable, and deployable AI framework, together with a research and policy roadmap, tailored to Bangladesh’s institutions and aligned with Sustainable Development Goal 6 (clean water and sanitation). Unlike global reviews, the emphasis throughout is on the specific hydroclimatic and governance context of Bangladesh.

2. Water-Quality Challenges in Bangladesh

2.1. Groundwater Arsenic

Arsenic enters Bangladesh’s shallow aquifers through the reductive dissolution of arsenic-bearing iron oxyhydroxides in Holocene alluvial sediments, a process favoured by reducing, organic-rich conditions [3]. Concentrations are spatially erratic, so that adjacent wells can differ by an order of magnitude, which makes blanket testing impractical and statistical prediction attractive. The World Health Organization guideline value is 10 µg L⁻¹, whereas the Bangladesh national standard is 50 µg L⁻¹ [18,19]; large populations are exposed above one or both thresholds, with elevated arsenic frequently co-occurring with iron, manganese, phosphate, and bicarbonate [3,4].

2.2. Coastal Salinity Intrusion

In the southwest coastal belt — Khulna, Satkhira, and Bagerhat — freshwater availability is increasingly limited by salinity. Drivers include sea-level rise, diminished upstream dry-season flow, tidal propagation up estuaries, cyclonic storm surge, and the conversion of paddy land to brackish shrimp ponds. Rising salinity degrades drinking-water sources and agricultural soils and is linked to adverse health outcomes. Because salinity varies strongly with season and tide, it is well suited to time-series forecasting once sufficient electrical-conductivity records are assembled [4,5].

2.3. River and Surface-Water Pollution

The rivers encircling Dhaka receive untreated effluent from thousands of tanneries, textile and dyeing units, and dense domestic discharge. Measurements show extreme organic loading, high ammonia, and dissolved oxygen approaching zero in the dry season [1]. Recent ML studies of the Buriganga, Balu, Tongi Khal, and Turag report that most physico-chemical parameters exceed permissible limits, placing average water quality in the lowest category and rendering the water unfit for domestic use [1,7]. Table 1 lists the parameters most relevant to these three problems together with the applicable Bangladesh and WHO thresholds.

3. Data Sources and AI Methods

3.1. Data Landscape in Bangladesh

Modelling depends on data from several national and external sources: the Department of Public Health Engineering (DPHE) and the British Geological Survey legacy surveys for groundwater arsenic; the Bangladesh Water Development Board (BWDB) for river stage, flow, and salinity; the Department of Environment (DoE) for surface-water quality; agricultural agencies for coastal salinity; and freely available satellite imagery (Sentinel-2/3, Landsat) for optical retrieval. Increasingly, low-cost IoT sensors measuring pH, electrical conductivity, turbidity, and total dissolved solids supply high-frequency streams [6,11]. These sources are heterogeneous, unevenly distributed, and gap-prone, so pre-processing — cleaning, imputation, feature engineering, and normalisation — is a critical and often decisive step. A common min–max normalisation rescales each predictor,

$$x\text{'}={\displaystyle \frac{x-x_{min}}{x_{max}-x_{min}}}$$

(1)

so that variables of differing magnitude contribute comparably to the learner.

3.2. The Water-Quality Index

For rivers, the weighted-arithmetic water-quality index (WQI) is the most common aggregate target,

$$WQI={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{W}_i{Q}_i}{{\displaystyle {\sum }_{i=1}^n}{W}_i}}$$

(2)

where the quality rating and weight of the i-th parameter are

$$Q_i=100\times {\displaystyle \frac{C_i-C_0}{S_i-C_0}},W_i={\displaystyle \frac{k}{S_i}}\left(3\right)$$

(3)

with C_i the observed concentration, S_i the permissible standard from Table 1, C₀ the ideal value, and k a constant. Studies then pose either WQI regression or water-quality-class classification [1].

3.3. Learning Algorithms

Tree ensembles dominate tabular water-quality prediction in Bangladesh. The random forest averages an ensemble of decorrelated trees,

$${ŷ}_{RF}={\displaystyle \frac{1}{B}}{\displaystyle \sum _{b=1}^B}{T}_b\left(x\right)\left(4\right)$$

(4)

where T_b is the b-th of B trees [13]; gradient-boosted variants (XGBoost, CatBoost) build trees sequentially to correct residual error [14]. For temporal problems such as salinity, the long short-term memory (LSTM) network captures long-range dependencies through gated recurrence [15], and has been combined with process knowledge in physics-informed neural networks (PINNs) that add the governing advection–dispersion–reaction residual to the loss,

$$\mathcal{L}={\mathcal{L}}_{data}+\lambda {\mathcal{L}}_{phys},{\mathcal{L}}_{phys}={\displaystyle \frac{1}{M}}\sum _j{\Vert {\displaystyle \frac{\partial Ĉ}{\partial t}}+u{\displaystyle \frac{\partial Ĉ}{\partial x}}-D{\displaystyle \frac{\partial Ĉ}{\partial x^2}}+kR\Vert }^2\left(5\right)$$

(5)

where Ĉ is the predicted concentration, u velocity, D dispersion, kR a reaction term, and λ balances data and physics. PINNs improve generalisation where labelled data are scarce — a recurring constraint in Bangladesh — and were central to the first integrated explainable framework for the monsoon-driven Padma River Basin [2,17].

3.4. Evaluation and Interpretability

Regression skill is reported with the root-mean-square error, mean absolute error,

$$\mathcal{L}={\mathcal{L}}_{data}+\lambda {\mathcal{L}}_{phys},{\mathcal{L}}_{phys}={\displaystyle \frac{1}{M}}\sum _j{\Vert {\displaystyle \frac{\partial Ĉ}{\partial t}}+u{\displaystyle \frac{\partial Ĉ}{\partial x}}-D{\displaystyle \frac{\partial Ĉ}{\partial x^2}}+kR\Vert }^2\left(6\right)$$

(6)

and the coefficient of determination,

$$R^2=1-{\displaystyle \frac{\sum _i{\left(y_i-{ŷ}_i\right)}^2}{\sum _i{\left(y_i-ȳ\right)}^2}}\left(7\right)$$

(7)

with the Nash–Sutcliffe efficiency (NSE) sharing the same form. Because predictions inform public-health and regulatory decisions, interpretability is essential. SHAP values, grounded in cooperative game theory, decompose each prediction additively into feature contributions,

$$ŷ\left(x\right)={\phi }_0+{\displaystyle \sum _{j=1}^M}{\phi }_j,{\phi }_j=\sum _{S\subseteq F\setminus \left\{j\right\}}{\displaystyle \frac{\left|S\right|!\left(\right|F|-|S|-1)!}{\left|F\right|!}}\left[f(S\cup \{j\left\}\right)-f\left(S\right)\right]\left(8\right)$$

(8)

where φ_j is the contribution of feature j and φ₀ the baseline [16]. Such attribution lets analysts confirm that an arsenic model relies on geochemically plausible drivers rather than spurious correlations. The complete pipeline, tailored to Bangladesh, is shown in Figure 2.

4. AI Applications for Bangladesh

This section reviews five application clusters, summarised with reported performance in Table 2.

4.1. Arsenic Risk Prediction and Mapping

Predicting whether a well exceeds an arsenic threshold from cheaper geochemical and spatial predictors is among the most valuable applications. In Chapainawabganj, a comparison of naïve Bayes, random forest, support-vector machine, decision tree, and logistic regression identified pH, electrical conductivity, sulphate, nitrate, iron, manganese, sodium, and bicarbonate as informative features and produced spatial arsenic-occurrence maps to guide mitigation [3]. In the coastal region, multiple ML algorithms were trained on comprehensive water-quality parameters to predict arsenic that ranged widely from 0.01 to 0.72 mg L⁻¹ [4]. Explainable analysis of such models consistently surfaces reductive-dissolution drivers — iron, redox state, depth, and phosphate — as dominant, as illustrated in Figure 4. These risk maps let limited testing budgets be directed to the highest-risk areas.

4.2. Coastal Salinity Forecasting

Salinity is strongly seasonal, peaking in the dry pre-monsoon months as freshwater flow recedes and tidal influence advances. Recurrent and physics-informed models can learn this cycle and project short-horizon salinity from meteorological and hydrological drivers, supporting planning of safe-water provision, rainwater harvesting, and managed aquifer recharge [4,5]. Figure 5 shows an illustrative forecast in which an LSTM anticipates the seasonal rise and crossing of an irrigation threshold, with an uncertainty band — an output type that is far more actionable than a single point estimate.

4.3. River Water-Quality-Index Prediction

For Dhaka’s rivers, a recent study compared fourteen ML models to predict the WQI of the Buriganga, Balu, Tongi Khal, and Turag. An artificial neural network gave the best skill, with a root-mean-square error of 2.34, mean absolute error of 1.24, Nash–Sutcliffe efficiency of 0.97, and R² of 0.97, with random forest also performing strongly; the authors emphasise the cost-effectiveness gained by coupling feature selection with ML [1]. In the monsoon-driven Padma Basin, an integrated framework combining ML, deep learning, and PINNs with explainable analysis and probabilistic uncertainty quantification was used to predict, interpret, and spatially regionalise seasonal WQI dynamics [2]. Figure 3 summarises the indicative performance envelope across such Bangladesh studies.

4.4. Satellite and IoT Monitoring

Remote sensing extends observation to whole river reaches and ungauged waters. For the Buriganga, optical indices for coloured dissolved organic matter, chlorophyll, and floating algae were derived from imagery and forecast with random-forest and LSTM models, keeping errors within an acceptable range and confirming heavy organic pollution [7]. Comparable hybrid pipelines — physically based retrieval feeding gradient-boosting and time-series models — have mapped chlorophyll-a in the shared Ganges system with high accuracy [8]. At ground level, low-cost IoT systems built on inexpensive microcontrollers and pH, turbidity, total-dissolved-solids, and temperature sensors stream data to ML classifiers for real-time safety screening, including for drinking water in tourist areas [6,11]. Together these technologies move monitoring from sparse snapshots toward continuous, spatially complete coverage.

Figure 4. Explainable-AI (SHAP-style) attribution of groundwater-arsenic prediction to its geochemical drivers. The ranking is illustrative but consistent with the reductive-dissolution mechanism that controls arsenic mobilisation in Bangladesh aquifers.

Figure 4. Explainable-AI (SHAP-style) attribution of groundwater-arsenic prediction to its geochemical drivers. The ranking is illustrative but consistent with the reductive-dissolution mechanism that controls arsenic mobilisation in Bangladesh aquifers.

4.5. Treatment and Early Warning

Beyond assessment, AI supports intervention. ML models guide the design and operation of arsenic- and iron-removal units and predict removal efficiency from operating conditions, while anomaly-detection and forecasting systems provide early warning of pollution episodes and effluent non-compliance [9,10]. Reviews of AI-based detection and control for river pollution highlight the shift from reactive measurement toward predictive, automated management [10]. Interpretable deep reinforcement learning has improved the operation of urban drainage systems, cutting untreated discharges while keeping the control logic auditable [12] — directly relevant to Dhaka’s overburdened drainage and frequent overflows. For Bangladesh, the highest near-term value lies in targeting scarce treatment and testing resources where models indicate the greatest risk.

Figure 5. Illustrative seasonal salinity-intrusion forecast for coastal Bangladesh. An LSTM-type model anticipates the dry-season rise and the crossing of an irrigation-salinity threshold, with an uncertainty band supporting risk-based planning. Synthetic data shown for illustration.

Figure 5. Illustrative seasonal salinity-intrusion forecast for coastal Bangladesh. An LSTM-type model anticipates the dry-season rise and the crossing of an irrigation-salinity threshold, with an uncertainty band supporting risk-based planning. Synthetic data shown for illustration.

5. An Integrated AI Roadmap for Bangladesh

The reviewed studies are individually strong but fragmented: each addresses a single problem, region, or model. We argue that the next advance for Bangladesh is integration — connecting the three challenges to shared data infrastructure, interpretable models, and decision channels (Figure 6). Table 3 maps each challenge to suitable AI methods, the data they require, and the management outcome they enable.

6. Challenges and Limitations

Data scarcity and quality. Long, consistent, openly accessible records are rare in Bangladesh. Monitoring is unevenly distributed, instruments and laboratories vary, and many historical arsenic surveys are now dated. Models trained on small or biased samples risk overfitting and poor real-world skill [3,5].

Monsoon seasonality and non-stationarity. Hydroclimatic forcing produces strong wet–dry contrasts and inter-annual variability; a model fitted to one season or year may fail in another. Explicitly seasonal and physics-informed designs help, but demonstrating skill across the monsoon cycle remains essential [2].

Transferability. Arsenic geochemistry and pollution regimes differ between basins, so a model calibrated in the northwest will not automatically transfer to the coast. Validation on independent districts — not merely held-out samples — is the meaningful test [3,4].

Interpretability and trust. For regulatory and public-health use, opaque predictions are insufficient. Explainable AI and knowledge-informed models are narrowing this gap and should be built in from the start [2,16].

Deployment and capacity. Sustained operation requires reliable sensors, connectivity, maintenance, and trained personnel — resources that are unevenly available. Few studies report long-term field performance, and IoT deployments must contend with calibration drift and harsh conditions [6,11].

7. Future Directions and Policy Recommendations

Realising the framework of Section 5 depends as much on institutions as on algorithms. Table 4 sets out prioritised recommendations spanning data, methods, deployment, and governance.

Methodologically, the most promising directions are physics-informed and explainable models that generalise from few labels [2,17]; data-assimilation and edge-AI architectures that fuse IoT streams with forecasts for real-time warning; and foundation or transfer-learning approaches that exploit data from data-rich basins to bootstrap models in data-poor districts. Each directly targets the data-scarcity and transferability barriers that most constrain Bangladesh.

8. Conclusion

Bangladesh’s water-quality crises — arsenic, salinity, and river pollution — are too large, too dynamic, and too spatially heterogeneous for conventional monitoring alone. The recent literature reviewed here shows that artificial intelligence already delivers accurate, locally validated tools: machine-learning models predict the water-quality index of Dhaka’s rivers with R² near 0.97 [1]; explainable and physics-informed frameworks interpret seasonal dynamics in the Padma Basin [2]; and ML risk maps and forecasts target arsenic and salinity where testing budgets are scarce [3,4]. The decisive next step is integration — linking the country’s monitoring agencies, remote sensing, and low-cost sensors to interpretable models and concrete decisions, as set out in the framework and roadmap proposed here. With investment in open data, calibrated sensing, transparent methods, and local capacity, AI can become not merely a means of observing Bangladesh’s water but a practical lever for improving it, in direct service of Sustainable Development Goal 6.